Search

To optimize proteins for particular traits holds great promise for industrial and pharmaceutical purposes. Machine Learning is increasingly applied in this field to predict properties of proteins, thereby guiding the experimental optimization process. A natural question is: How much progress are we making with such predictions, and how important is the choice of regressor and representation? In this paper, we demonstrate that different assessment criteria for regressor performance can lead to dramatically different conclusions, depending on the choice of metric, and how one defines generalization. We highlight the fundamental issues of sample bias in typical regression scenarios and how this can lead to misleading conclusions about regressor performance. Finally, we make the case for the importance of calibrated uncertainty in this domain. Author summary: Supervised machine learning is increasingly used to predict the function and properties of proteins. The performance obtained with these methods relies on a multitude of factors including how data is represented, how observations are distributed, how training is conducted, and how performance is measured. In this paper, we systematically assess the importance of these different components in a protein regression pipeline. We discuss the benefits of using representations extracted from protein language models, the impact of the choice of regression algorithm, and the role of uncertainty. Finally, to avoid misleading performance claims, we stress the need for carefully aligning the train/test setup to reflect the setting in which the prediction algorithm will ultimately be applied. [ABSTRACT FROM AUTHOR]

The heterogeneity of Hepatocellular Carcinoma (HCC) poses a barrier to effective treatment. Stratifying highly heterogeneous HCC into molecular subtypes with similar features is crucial for personalized anti-tumor therapies. Although driver genes play pivotal roles in cancer progression, their potential in HCC subtyping has been largely overlooked. This study aims to utilize driver genes to construct HCC subtype models and unravel their molecular mechanisms. Utilizing a novel computational framework, we expanded the initially identified 96 driver genes to 1192 based on mutational aspects and an additional 233 considering driver dysregulation. These genes were subsequently employed as stratification markers for further analyses. A novel multi-omics subtype classification algorithm was developed, leveraging mutation and expression data of the identified stratification genes. This algorithm successfully categorized HCC into two distinct subtypes, CLASS A and CLASS B, demonstrating significant differences in survival outcomes. Integrating multi-omics and single-cell data unveiled substantial distinctions between these subtypes regarding transcriptomics, mutations, copy number variations, and epigenomics. Moreover, our prognostic model exhibited excellent predictive performance in training and external validation cohorts. Finally, a 10-gene classification model for these subtypes identified TTK as a promising therapeutic target with robust classification capabilities. This comprehensive study provides a novel perspective on HCC stratification, offering crucial insights for a deeper understanding of its pathogenesis and the development of promising treatment strategies. Author summary: Dividing highly heterogeneous HCC into molecular subtypes with similar characteristics is crucial for personalized anti-tumor therapies. Although driver genes play pivotal roles in cancer progression, their potential in HCC subtyping has been largely overlooked. In this work, we developed a multi-omics network-based stratification algorithm that utilizes patient mutation data and requires smaller computational resources for subtype assignment. Through this algorithm, we categorized HCC into two subtypes, CLASS A and CLASS B. Using multi-omics and single-cell data, we identified differences between these subtypes in gene expression, methylation, immune infiltration, and other aspects. Beyond subtype characterization, our study established a robust clinical prediction model (https://mike-wang-bjut.shinyapps.io/DynNomapp%5fHCC%5fSutypes/) incorporating subtype information and typical clinical features, enabling precise survival predictions. Finally, we developed a high-performing machine learning classifier for our subtype. Analyzing this classification model and reviewing previous experimental papers, we identified TTK as a potential diagnostic marker and therapeutic target specific to our subtypes. In conclusion, our research offers a novel perspective on HCC stratification, which is crucial for a deeper understanding of its pathogenesis and developing promising treatment strategies. [ABSTRACT FROM AUTHOR]

The modelling of many real-world problems relies on computationally heavy simulations of randomly interacting individuals or agents. However, the values of the parameters that underlie the interactions between agents are typically poorly known, and hence they need to be inferred from macroscopic observations of the system. Since statistical inference rests on repeated simulations to sample the parameter space, the high computational expense of these simulations can become a stumbling block. In this paper, we compare two ways to mitigate this issue in a Bayesian setting through the use of machine learning methods: One approach is to construct lightweight surrogate models to substitute the simulations used in inference. Alternatively, one might altogether circumvent the need for Bayesian sampling schemes and directly estimate the posterior distribution. We focus on stochastic simulations that track autonomous agents and present two case studies: tumour growths and the spread of infectious diseases. We demonstrate that good accuracy in inference can be achieved with a relatively small number of simulations, making our machine learning approaches orders of magnitude faster than classical simulation-based methods that rely on sampling the parameter space. However, we find that while some methods generally produce more robust results than others, no algorithm offers a one-size-fits-all solution when attempting to infer model parameters from observations. Instead, one must choose the inference technique with the specific real-world application in mind. The stochastic nature of the considered real-world phenomena poses an additional challenge that can become insurmountable for some approaches. Overall, we find machine learning approaches that create direct inference machines to be promising for real-world applications. We present our findings as general guidelines for modelling practitioners. Author summary: Computer simulations play a vital role in modern science as they are commonly used to compare theory with observations. One can infer the properties of a system by comparing the data to the predicted behaviour in different scenarios. Each scenario corresponds to a simulation with slightly different settings. However, since real-world problems are highly complex, the simulations often require extensive computational resources, making direct comparisons with data challenging, if not insurmountable. It is, therefore, necessary to resort to inference methods that mitigate this issue, but it is not clear-cut what path to choose for any specific research problem. In this paper, we provide general guidelines for how to make this choice. We do so by studying examples from oncology and epidemiology and by taking advantage of machine learning. More specifically, we focus on simulations that track the behaviour of autonomous agents, such as single cells or individuals. We show that the best way forward is problem-dependent and highlight the methods that yield the most robust results across the different case studies. Rather than relying on a single inference technique, we recommend employing several methods and selecting the most reliable based on predetermined criteria. [ABSTRACT FROM AUTHOR]

When grown on agar surfaces, microbes can produce distinct multicellular spatial structures called colonies, which contain characteristic sizes, shapes, edges, textures, and degrees of opacity and color. For over one hundred years, researchers have used these morphology cues to classify bacteria and guide more targeted treatment of pathogens. Advances in genome sequencing technology have revolutionized our ability to classify bacterial isolates and while genomic methods are in the ascendancy, morphological characterization of bacterial species has made a resurgence due to increased computing capacities and widespread application of machine learning tools. In this paper, we revisit the topic of colony morphotype on the within-species scale and apply concepts from image processing, computer vision, and deep learning to a dataset of 69 environmental and clinical Pseudomonas aeruginosa strains. We find that colony morphology and complexity under common laboratory conditions is a robust, repeatable phenotype on the level of individual strains, and therefore forms a potential basis for strain classification. We then use a deep convolutional neural network approach with a combination of data augmentation and transfer learning to overcome the typical data starvation problem in biological applications of deep learning. Using a train/validation/test split, our results achieve an average validation accuracy of 92.9% and an average test accuracy of 90.7% for the classification of individual strains. These results indicate that bacterial strains have characteristic visual 'fingerprints' that can serve as the basis of classification on a sub-species level. Our work illustrates the potential of image-based classification of bacterial pathogens and highlights the potential to use similar approaches to predict medically relevant strain characteristics like antibiotic resistance and virulence from colony data., Competing Interests: The authors declare no competing interests., (Copyright: © 2023 Rattray et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.)

Recent studies have established that the circadian clock influences onset, progression and therapeutic outcomes in a number of diseases including cancer and heart diseases. Therefore, there is a need for tools to measure the functional state of the molecular circadian clock and its downstream targets in patients. Moreover, the clock is a multi-dimensional stochastic oscillator and there are few tools for analysing it as a noisy multigene dynamical system. In this paper we consider the methodology behind TimeTeller, a machine learning tool that analyses the clock as a noisy multigene dynamical system and aims to estimate circadian clock function from a single transcriptome by modelling the multi-dimensional state of the clock. We demonstrate its potential for clock systems assessment by applying it to mouse, baboon and human microarray and RNA-seq data and show how to visualise and quantify the global structure of the clock, quantitatively stratify individual transcriptomic samples by clock dysfunction and globally compare clocks across individuals, conditions and tissues thus highlighting its potential relevance for advancing circadian medicine. Author summary: The cellular circadian clock consists of an interacting set of genes that through their interactions oscillate throughout the day. This oscillator also responds to external cues so that the genes oscillate in phase with external environmental rhythms. A cell therefore uses its circadian clock to provide its genes with information about the external time. In this way it can coordinate many of the processes taking place in the cell and allocate some of these processes to specific times of the day. It is becoming increasingly clear that the quality of this timing information influences onset progression and outcome in a number of chronic diseases such as cancer. Our aim is therefore to develop a machine-learning tool that can assess how well the clock is working. We want to use this with patients and therefore, for clinical utility, it needs to work with only a single clinical sample and to produce reproducible results that can be clearly interpreted and easily compared. [ABSTRACT FROM AUTHOR]

Mechanistic models have been used for centuries to describe complex interconnected processes, including biological ones. As the scope of these models has widened, so have their computational demands. This complexity can limit their suitability when running many simulations or when real-time results are required. Surrogate machine learning (ML) models can be used to approximate the behaviour of complex mechanistic models, and once built, their computational demands are several orders of magnitude lower. This paper provides an overview of the relevant literature, both from an applicability and a theoretical perspective. For the latter, the paper focuses on the design and training of the underlying ML models. Application-wise, we show how ML surrogates have been used to approximate different mechanistic models. We present a perspective on how these approaches can be applied to models representing biological processes with potential industrial applications (e.g., metabolism and whole-cell modelling) and show why surrogate ML models may hold the key to making the simulation of complex biological systems possible using a typical desktop computer. [ABSTRACT FROM AUTHOR]

We propose the "runtime learning" hypothesis which states that people quickly learn to perform unfamiliar tasks as the tasks arise by using task-relevant instances of concepts stored in memory during mental training. To make learning rapid, the hypothesis claims that only a few class instances are used, but these instances are especially valuable for training. The paper motivates the hypothesis by describing related ideas from the cognitive science and machine learning literatures. Using computer simulation, we show that deep neural networks (DNNs) can learn effectively from small, curated training sets, and that valuable training items tend to lie toward the centers of data item clusters in an abstract feature space. In a series of three behavioral experiments, we show that people can also learn effectively from small, curated training sets. Critically, we find that participant reaction times and fitted drift rates are best accounted for by the confidences of DNNs trained on small datasets of highly valuable items. We conclude that the runtime learning hypothesis is a novel conjecture about the relationship between learning and memory with the potential for explaining a wide variety of cognitive phenomena., Competing Interests: No competing interests., (Copyright: © 2023 German et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.)

The ubiquitous availability of genome sequencing data explains the popularity of machine learning-based methods for the prediction of protein properties from their amino acid sequences. Over the years, while revising our own work, reading submitted manuscripts as well as published papers, we have noticed several recurring issues, which make some reported findings hard to understand and replicate. We suspect this may be due to biologists being unfamiliar with machine learning methodology, or conversely, machine learning experts may miss some of the knowledge needed to correctly apply their methods to proteins. Here, we aim to bridge this gap for developers of such methods. The most striking issues are linked to a lack of clarity: how were annotations of interest obtained; which benchmark metrics were used; how are positives and negatives defined. Others relate to a lack of rigor: If you sneak in structural information, your method is not sequence-based; if you compare your own model to "state-of-the-art," take the best methods; if you want to conclude that some method is better than another, obtain a significance estimate to support this claim. These, and other issues, we will cover in detail. These points may have seemed obvious to the authors during writing; however, they are not always clear-cut to the readers. We also expect many of these tips to hold for other machine learning-based applications in biology. Therefore, many computational biologists who develop methods in this particular subject will benefit from a concise overview of what to avoid and what to do instead. [ABSTRACT FROM AUTHOR]

Proteolysis-targeting chimeras (PROTACs) are hetero-bifunctional molecules that induce the degradation of target proteins by recruiting an E3 ligase. PROTACs have the potential to inactivate disease-related genes that are considered undruggable by small molecules, making them a promising therapy for the treatment of incurable diseases. However, only a few hundred proteins have been experimentally tested for their amenability to PROTACs, and it remains unclear which other proteins in the entire human genome can be targeted by PROTACs. In this study, we have developed PrePROTAC, an interpretable machine learning model based on a transformer-based protein sequence descriptor and random forest classification. PrePROTAC predicts genome-wide targets that can be degraded by CRBN, one of the E3 ligases. In the benchmark studies, PrePROTAC achieved a ROC-AUC of 0.81, an average precision of 0.84, and over 40% sensitivity at a false positive rate of 0.05. When evaluated by an external test set which comprised proteins from different structural folds than those in the training set, the performance of PrePROTAC did not drop significantly, indicating its generalizability. Furthermore, we developed an embedding SHapley Additive exPlanations (eSHAP) method, which extends conventional SHAP analysis for original features to an embedding space through in silico mutagenesis. This method allowed us to identify key residues in the protein structure that play critical roles in PROTAC activity. The identified key residues were consistent with existing knowledge. Using PrePROTAC, we identified over 600 novel understudied proteins that are potentially degradable by CRBN and proposed PROTAC compounds for three novel drug targets associated with Alzheimer's disease. Author summary: Many human diseases remain incurable because disease-causing genes cannot by selectively and effectively targeted by small molecules. Proteolysis-targeting chimera (PROTAC), an organic compound that binds to both a target and a degradation-mediating E3 ligase, has emerged as a promising approach to selectively target disease-driving genes that are not druggable by small molecules. However, not all of proteins can be accommodated by E3 ligases, and be effectively degraded. Knowledge about the degradability of a protein will be crucial for PROTAC design. However, only hundreds of proteins have been experimentally tested if they are amenable to the PROTACs. This leaves us uncertain about which other proteins in the entire human genome can be targeted by PROTACs. In this paper, we propose an intepretable machine learning model, PrePROTAC, which takes advantage of powerful protein language modeling. PrePROTAC achieves high accuracy when evaluated with an external dataset which comes from different gene families from the proteins in the training data, suggesting the generalizability of this model. We apply PrePROTAC to the human genome, and identify more than 600 understudied proteins that are potentially responsive to PROTACs. Furthermore, we design PROTAC compounds for three novel drug targets associated with Alzheimer's disease. [ABSTRACT FROM AUTHOR]

Quantitative understanding of microbial growth is an essential prerequisite for successful control of pathogens as well as various biotechnology applications. Even though the growth of cell populations has been extensively studied, microbial growth remains poorly characterised at the spatial level. Indeed, even isogenic populations growing at different locations on solid growth medium typically show significant location-dependent variability in growth. Here we show that this variability can be attributed to the initial physiological states of the populations, the interplay between populations interacting with their local environment and the diffusion of nutrients and energy sources coupling the environments. We further show how the causes of this variability change throughout the growth of a population. We use a dual approach, first applying machine learning regression models to discover that location dominates growth variability at specific times, and, in parallel, developing explicit population growth models to describe this spatial effect. In particular, treating nutrient and energy source concentration as a latent variable allows us to develop a mechanistic resource consumer model that captures growth variability across the shared environment. As a consequence, we are able to determine intrinsic growth parameters for each local population, removing confounders common to location-dependent variability in growth. Importantly, our explicit low-parametric model for the environment paves the way for massively parallel experimentation with configurable spatial niches for testing specific eco-evolutionary hypotheses. Author summary: Image-based platforms allow obtaining population size estimates for massively parallel growth experiments on substrate plates at a relatively low cost. However, such population size data has been shown to display a high degree of spatial variability, which occurs even with isogenic populations. Here we first quantified the importance of spatial location on growth variation using a machine learning approach, and then developed spatially aware population growth models to explain the spatial structure of the growth data. Ultimately, we show that a spatial resource consumer model with local microhabitats connected via diffusion, and a parameter capturing the initial physiological state of a population, can fully explain the observed spatial variation in growth while allowing the inference of intrinsic growth parameters of specific populations. This result provides a method for systematic extraction of spatial growth models and paves the way for massively parallel eco-evolutionary experimentation. [ABSTRACT FROM AUTHOR]

Empirical work has shown that human cultural transmission can be heavily influenced by population age structure. We aim to explore the role of such age structure in shaping the cultural composition of a population when cultural transmission occurs in an unbiased way. In particular, we are interested in understanding the effect induced by the interplay between age structure and the cultural transmission process by allowing cultural transmission from individuals within a limited age range only. To this end we develop an age-structured cultural transmission model and find that age-structured and non age-structured populations evolving through unbiased transmission possess very similar cultural compositions (at a single point in time) at the population and sample level if the copy pool consists of a sufficiently large fraction of the population. If, however, an age constraint—a structural constraint restricting the pool of potential role models to individuals of a limited age range— exists, the cultural compositions of age-structured and non age-structured population show stark differences. This may have drastic consequences for our ability to correctly analyse cultural data sets. Rejections of tests of neutrality, blind to age structure and, importantly, the interaction between age structure and cultural transmission, are only indicative of biased transmission if it is known a priori that there are no or only weak age constraints acting on the pool of role models. As this knowledge is rarely available for specific empirical case studies we develop a generative inference approach based on our age-structured cultural transmission model and machine learning techniques. We show that in some circumstances it is possible to simultaneously infer the characteristics of the age structure, the nature of the transmission process, and the interplay between them from observed samples of cultural variants. Our results also point to hard limits on inference from population-level data at a single point in time, regardless of the approach used. Author summary: Age structure is undoubtedly a feature of human populations and empirical cultural evolutionary research has shown that the specifics of age structure can heavily influence how humans learn and transmit different cultural traits. Nevertheless, such demographic properties are rarely included in theoretical models of human cultural evolution. This may present a problem, especially for model-based inferential analyses. In this paper, I show when the population-level signatures of age-structured and non age-structured populations—evolving through the process of unbiased transmission—show detectable differences and therefore when demographic properties must be included in models of cultural evolution. I demonstrate that a failure to do so may lead to erroneous conclusions about evolutionary processes, in particular processes of cultural transmission, underlying observed patterns of cultural change. To mitigate this risk I develop a generative inference framework combining cultural evolutionary and machine learning techniques. This framework also provides valuable information about hard limits on inference from sparse population-level data. [ABSTRACT FROM AUTHOR]

Machine learning tools have proven useful across biological disciplines, allowing researchers to draw conclusions from large datasets, and opening up new opportunities for interpreting complex and heterogeneous biological data. Alongside the rapid growth of machine learning, there have also been growing pains: some models that appear to perform well have later been revealed to rely on features of the data that are artifactual or biased; this feeds into the general criticism that machine learning models are designed to optimize model performance over the creation of new biological insights. A natural question arises: how do we develop machine learning models that are inherently interpretable or explainable? In this manuscript, we describe the SWIF(r) reliability score (SRS), a method building on the SWIF(r) generative framework that reflects the trustworthiness of the classification of a specific instance. The concept of the reliability score has the potential to generalize to other machine learning methods. We demonstrate the utility of the SRS when faced with common challenges in machine learning including: 1) an unknown class present in testing data that was not present in training data, 2) systemic mismatch between training and testing data, and 3) instances of testing data that have missing values for some attributes. We explore these applications of the SRS using a range of biological datasets, from agricultural data on seed morphology, to 22 quantitative traits in the UK Biobank, and population genetic simulations and 1000 Genomes Project data. With each of these examples, we demonstrate how the SRS can allow researchers to interrogate their data and training approach thoroughly, and to pair their domain-specific knowledge with powerful machine-learning frameworks. We also compare the SRS to related tools for outlier and novelty detection, and find that it has comparable performance, with the advantage of being able to operate when some data are missing. The SRS, and the broader discussion of interpretable scientific machine learning, will aid researchers in the biological machine learning space as they seek to harness the power of machine learning without sacrificing rigor and biological insight. Author summary: Machine learning methods are incredibly powerful at performing tasks such as classification and clustering, but they also pose unique problems that can limit new insights. Complex machine learning models may reach conclusions that are difficult or impossible for researchers to understand after-the-fact, sometimes producing biased or meaningless results. It is therefore essential that researchers have tools that allow them to understand how machine learning tools reach their conclusions, so that they can effectively design models. This paper builds on the machine learning method SWIF(r), originally designed to detect regions in the genome targeted by natural selection. Our new method, the SWIF(r) Reliability Score (SRS), can help researchers evaluate how trustworthy the prediction of a SWIF(r) model is when classifying a specific instance of data. We also show how SWIF(r) and the SRS can be used for biological problems outside the original scope of SWIF(r). We show that the SRS is helpful in situations where the data used to train the machine learning model fails to represent the testing data in some way. The SRS can be used across many different disciplines, and has unique properties for scientific machine learning research. [ABSTRACT FROM AUTHOR]

Measles is one the best-documented and most-mechanistically-studied non-linear infectious disease dynamical systems. However, systematic investigation into the comparative performance of traditional mechanistic models and machine learning approaches in forecasting the transmission dynamics of this pathogen are still rare. Here, we compare one of the most widely used semi-mechanistic models for measles (TSIR) with a commonly used machine learning approach (LASSO), comparing performance and limits in predicting short to long term outbreak trajectories and seasonality for both regular and less regular measles outbreaks in England and Wales (E&W) and the United States. First, our results indicate that the proposed LASSO model can efficiently use data from multiple major cities and achieve similar short-to-medium term forecasting performance to semi-mechanistic models for E&W epidemics. Second, interestingly, the LASSO model also captures annual to biennial bifurcation of measles epidemics in E&W caused by susceptible response to the late 1940s baby boom. LASSO may also outperform TSIR for predicting less-regular dynamics such as those observed in major cities in US between 1932–45. Although both approaches capture short-term forecasts, accuracy suffers for both methods as we attempt longer-term predictions in highly irregular, post-vaccination outbreaks in E&W. Finally, we illustrate that the LASSO model can both qualitatively and quantitatively reconstruct mechanistic assumptions, notably susceptible dynamics, in the TSIR model. Our results characterize the limits of predictability of infectious disease dynamics for strongly immunizing pathogens with both mechanistic and machine learning models, and identify connections between these two approaches. Author summary: Machine learning techniques in infectious disease modeling have grown in popularity in recent years. However, systematic investigation into the comparative performance of these approaches with traditional mechanistic models are still rare. In this paper, we compare one of the most widely used semi-mechanistic models for measles (TSIR) with a commonly used machine learning approach (LASSO), comparing performance and limits in predicting short to long term outbreaks of measles, one of the best-documented and most-mechanistically-studied non-linear infectious disease dynamical systems. Our results show that in general the LASSO outperform TSIR for predicting less-regular dynamics, and it can achieve similar performance in other scenarios when compared to the TSIR. The LASSO also has the advantages of not requiring explicit demographic data in model training. Finally, we identify connections between these two approaches and show that the LASSO model can both qualitatively and quantitatively reconstruct mechanistic assumptions, notably susceptible dynamics, in the TSIR model. [ABSTRACT FROM AUTHOR]

The origin of life involved complicated evolutionary processes. Computer modeling is a promising way to reveal relevant mechanisms. However, due to the limitation of our knowledge on prebiotic chemistry, it is usually difficult to justify parameter-setting for the modeling. Thus, typically, the studies were conducted in a reverse way: the parameter-space was explored to find those parameter values "supporting" a hypothetical scene (that is, leaving the parameter-justification a later job when sufficient knowledge is available). Exploring the parameter-space manually is an arduous job (especially when the modeling becomes complicated) and additionally, difficult to characterize as regular "Methods" in a paper. Here we show that a machine-learning-like approach may be adopted, automatically optimizing the parameters. With this efficient parameter-exploring approach, the evolutionary modeling on the origin of life would become much more powerful. In particular, based on this, it is expected that more near-reality (complex) models could be introduced, and thereby theoretical research would be more tightly associated with experimental investigation in this field-hopefully leading to significant steps forward in respect to our understanding on the origin of life., Competing Interests: The authors have declared that no competing interests exist.

Generative models of protein sequence families are an important tool in the repertoire of protein scientists and engineers alike. However, state-of-the-art generative approaches face inference, accuracy, and overfitting- related obstacles when modeling moderately sized to large proteins and/or protein families with low sequence coverage. Here, we present a simple to learn, tunable, and accurate generative model, GENERALIST: GENERAtive nonLInear tenSor-factorizaTion for protein sequences. GENERALIST accurately captures several high order summary statistics of amino acid covariation. GENERALIST also predicts conservative local optimal sequences which are likely to fold in stable 3D structure. Importantly, unlike current methods, the density of sequences in GENERALIST-modeled sequence ensembles closely resembles the corresponding natural ensembles. Finally, GENERALIST embeds protein sequences in an informative latent space. GENERALIST will be an important tool to study protein sequence variability. Author summary: Protein sequence families show tremendous sequence variation. Yet, it is thought that a large portion of the functional sequence space remains unexplored. Generative models are machine learning methods that allow us to learn what makes proteins functional using sequences of naturally occurring proteins. Here, we present a new type of generative model GENERALIST: GENERAtive nonLInear tenSor-factorizaTion for protein sequences that is accurate, easy to implement, and works with very small datasets. We believe that GENERALIST will be an important tool in the repertoire of protein scientists and engineers alike. [ABSTRACT FROM AUTHOR]

Segmenting three-dimensional (3D) microscopy images is essential for understanding phenomena like morphogenesis, cell division, cellular growth, and genetic expression patterns. Recently, deep learning (DL) pipelines have been developed, which claim to provide high accuracy segmentation of cellular images and are increasingly considered as the state of the art for image segmentation problems. However, it remains difficult to define their relative performances as the concurrent diversity and lack of uniform evaluation strategies makes it difficult to know how their results compare. In this paper, we first made an inventory of the available DL methods for 3D cell segmentation. We next implemented and quantitatively compared a number of representative DL pipelines, alongside a highly efficient non-DL method named MARS. The DL methods were trained on a common dataset of 3D cellular confocal microscopy images. Their segmentation accuracies were also tested in the presence of different image artifacts. A specific method for segmentation quality evaluation was adopted, which isolates segmentation errors due to under- or oversegmentation. This is complemented with a 3D visualization strategy for interactive exploration of segmentation quality. Our analysis shows that the DL pipelines have different levels of accuracy. Two of them, which are end-to-end 3D and were originally designed for cell boundary detection, show high performance and offer clear advantages in terms of adaptability to new data. Author summary: In recent years, a number of deep learning (DL) algorithms based on computational neural networks have been developed, which claim to achieve high accuracy and automatic segmentation of three-dimensional (3D) microscopy images. Although these algorithms have received considerable attention in the literature, it is difficult to evaluate their relative performances, while it remains unclear whether they really perform better than other, more classical segmentation methods. To clarify these issues, we performed a detailed, quantitative analysis of a number of representative DL pipelines for cell instance segmentation from 3D confocal microscopy image datasets. We developed a protocol for benchmarking the performances of such DL-based segmentation pipelines using common training and test datasets, evaluation metrics, and visualizations. Using this protocol, we evaluated and compared 4 different DL pipelines to identify their strengths and limitations. A high performance non-DL method was also included in the evaluation. We show that DL pipelines may show significant differences in their performances depending on their model architecture and pipeline components but overall show excellent adaptability to unseen data. We also show that our benchmarking protocol could be extended to a variety of segmentation pipelines and datasets. [ABSTRACT FROM AUTHOR]

Machine learning algorithms, including recent advances in deep learning, are promising for tools for detection and classification of broadband high frequency signals in passive acoustic recordings. However, these methods are generally data-hungry and progress has been limited by challenges related to the lack of labeled datasets adequate for training and testing. Large quantities of known and as yet unidentified broadband signal types mingle in marine recordings, with variability introduced by acoustic propagation, source depths and orientations, and interacting signals. Manual classification of these datasets is unmanageable without an in-depth knowledge of the acoustic context of each recording location. A signal classification pipeline is presented which combines unsupervised and supervised learning phases with opportunities for expert oversight to label signals of interest. The method is illustrated with a case study using unsupervised clustering to identify five toothed whale echolocation click types and two anthropogenic signal categories. These categories are used to train a deep network to classify detected signals in either averaged time bins or as individual detections, in two independent datasets. Bin-level classification achieved higher overall precision (>99%) than click-level classification. However, click-level classification had the advantage of providing a label for every signal, and achieved higher overall recall, with overall precision from 92 to 94%. The results suggest that unsupervised learning is a viable solution for efficiently generating the large, representative training sets needed for applications of deep learning in passive acoustics. Author summary: Machine learning algorithms have proven to be effective for tools for detection and classification tasks in many fields, however, these processes are generally data-hungry and their use in marine acoustics has been limited by a lack of large labeled datasets for algorithms to learn from. In underwater acoustic recordings, many signals generated by animals, human activities and physical processes mingle together, and their sounds can change depending on ocean temperatures, locations, and behavior. Manual classification of these datasets is unmanageable without an in-depth knowledge of the acoustic context of each recording location. This paper presents a process which combines unsupervised and supervised learning phases and expert oversight to generate and use large datasets for acoustic classification of marine mammal and human-generated signals. Unsupervised learning is used to automatically generate the large training datasets needed to teach a supervised learning algorithm to correctly classify seven different signal types commonly recorded in the Southern California Bight. Using this process, researchers with large unlabeled acoustic datasets can begin to take advantage of widespread advances in machine learning. [ABSTRACT FROM AUTHOR]

The brain learns representations of sensory information from experience, but the algorithms by which it does so remain unknown. One popular theory formalizes representations as inferred factors in a generative model of sensory stimuli, meaning that learning must improve this generative model and inference procedure. This framework underlies many classic computational theories of sensory learning, such as Boltzmann machines, the Wake/Sleep algorithm, and a more recent proposal that the brain learns with an adversarial algorithm that compares waking and dreaming activity. However, in order for such theories to provide insights into the cellular mechanisms of sensory learning, they must be first linked to the cell types in the brain that mediate them. In this study, we examine whether a subtype of cortical interneurons might mediate sensory learning by serving as discriminators, a crucial component in an adversarial algorithm for representation learning. We describe how such interneurons would be characterized by a plasticity rule that switches from Hebbian plasticity during waking states to anti-Hebbian plasticity in dreaming states. Evaluating the computational advantages and disadvantages of this algorithm, we find that it excels at learning representations in networks with recurrent connections but scales poorly with network size. This limitation can be partially addressed if the network also oscillates between evoked activity and generative samples on faster timescales. Consequently, we propose that an adversarial algorithm with interneurons as discriminators is a plausible and testable strategy for sensory learning in biological systems. Author summary: After raw sensory data is received at the periphery, it is transformed by various neural pathways and delivered to the sensory cortex. There, neural activity forms an internal model of the state of the external world, which is updated appropriately by new information. A goal of learning, then, is to learn how to transform information into the appropriate representational form within the brain's internal model. Here, we look to artificial intelligence for new possible theories of how the brain might learn representations that resolve issues with previously proposed theories. We describe how one particular algorithm—adversarial learning—resolves a major issue with previous hypotheses relating to recurrence. Furthermore, this algorithm resembles broad features of the organization and learning dynamics of the brain, such as wake and sleep cycles. Considering seriously how this algorithm would appear if implemented by the brain, we map its features to known physiology and make testable predictions for how neural circuits learn new representations of information. [ABSTRACT FROM AUTHOR]

Aedes mosquitoes, known as vectors of mosquito-borne diseases, pose significant risks to public health and safety. Modeling the population dynamics of Aedes mosquitoes requires comprehensive approaches due to the complex interplay between biological mechanisms and environmental factors. This study developed a model that couples differential equations with a neural network to simulate the dynamics of mosquito population, and explore the relationships between oviposition rate, temperature, and precipitation. Data from nine cities in Guangdong Province spanning four years were used for model training and parameter estimation, while data from the remaining three cities were reserved for model validation. The trained model successfully simulated the mosquito population dynamics across all twelve cities using the same set of parameters. Correlation coefficients between simulated results and observed data exceeded 0.7 across all cities, with some cities surpassing 0.85, demonstrating high model performance. The coupled neural network in the model effectively revealed the relationships among oviposition rate, temperature, and precipitation, aligning with biological patterns. Furthermore, symbolic regression was used to identify the optimal functional expression for these relationships. By integrating the traditional dynamic model with machine learning, our model can adhere to specific biological mechanisms while extracting patterns from data, thus enhancing its interpretability in biology. Our approach provides both accurate modeling and an avenue for uncovering potential unknown biological mechanisms. Our conclusions can provide valuable insights into designing strategies for controlling mosquito-borne diseases and developing related prediction and early warning systems. Author summary: When simulating population dynamics, scholars often rely on predefined functions to describe processes such as growth, development, and mortality. However, these functions may not be able to accurately capture the underlying biological mechanisms, posing challenges for related research. In this study, we adopted a data-driven approach integrated with mechanistic principles to simulate mosquito dynamics. Given the unknown functional relationships between oviposition rate, temperature, and precipitation, we employed a neural network to model these relationships, resulting in a model that combines the neural network with traditional mechanistic approaches. This method can successfully simulate the changing trends of mosquitoes, infer unknown parameters, train the neural network, and capture the relationships between oviposition rate, temperature, and precipitation. Data from nine cities in Guangdong Province were used for model training, while data from the remaining three cities were used for model validation. Our model successfully fitted data from all cities using the same set of parameters, demonstrating satisfactory performance. On this basis, we successfully identified the functional form between oviposition rate, temperature, and precipitation through the symbolic regression algorithm, which was consistent with the results obtained from the neural network. These findings show that our method can accurately simulate population dynamics while exploring unknown biological mechanisms. [ABSTRACT FROM AUTHOR]

A fundamental feature of the human brain is its capacity to learn novel motor skills. This capacity requires the formation of vastly different visuomotor mappings. Using a grid navigation task, we investigated whether training variability would enhance the flexible use of a visuomotor mapping (key-to-direction rule), leading to better generalization performance. Experiments 1 and 2 show that participants trained to move between multiple start-target pairs exhibited greater generalization to both distal and proximal targets compared to participants trained to move between a single pair. This finding suggests that limited variability can impair decisions even in simple tasks without planning. In addition, during the training phase, participants exposed to higher variability were more inclined to choose options that, counterintuitively, moved the cursor away from the target while minimizing its actual distance under the constrained mapping, suggesting a greater engagement in model-based computations. In Experiments 3 and 4, we showed that the limited generalization performance in participants trained with a single pair can be enhanced by a short period of variability introduced early in learning or by incorporating stochasticity into the visuomotor mapping. Our computational modeling analyses revealed that a hybrid model between model-free and model-based computations with different mixing weights for the training and generalization phases, best described participants' data. Importantly, the differences in the model-based weights between our experimental groups, paralleled the behavioral findings during training and generalization. Taken together, our results suggest that training variability enables the flexible use of the visuomotor mapping, potentially by preventing the consolidation of habits due to the continuous demand to change responses. Author summary: The development of new motor skills often requires the learning of novel associations between actions and outcomes. These novel mappings can be flexible and generalize to new situations, or more local with narrow generalization, similar to stimulus-action associations. In a series of experiments using a navigation task, we showed that generalizable mappings are favored under a training variability regime, while local mappings with narrow generalization are developed in the absence of variability. Training variability was generated in our experiments either with multiple goals or with stochasticity in the action-outcome mapping, with both regimes leading to successful generalization. In addition, we showed that the benefits in generalization from training variability can be observed even when participants are subsequently exposed to no variability for a prolonged period of time. These results were best described by a mixture of model-free and model-based reinforcement learning algorithms, with different mixture weights for the training and generalization phases. [ABSTRACT FROM AUTHOR]

Viruses of microbes are ubiquitous biological entities that reprogram their hosts' metabolisms during infection in order to produce viral progeny, impacting the ecology and evolution of microbiomes with broad implications for human and environmental health. Advances in genome sequencing have led to the discovery of millions of novel viruses and an appreciation for the great diversity of viruses on Earth. Yet, with knowledge of only "who is there?" we fall short in our ability to infer the impacts of viruses on microbes at population, community, and ecosystem-scales. To do this, we need a more explicit understanding "who do they infect?" Here, we developed a novel machine learning model (ML), Virus-Host Interaction Predictor (VHIP), to predict virus-host interactions (infection/non-infection) from input virus and host genomes. This ML model was trained and tested on a high-value manually curated set of 8849 virus-host pairs and their corresponding sequence data. The resulting dataset, 'Virus Host Range network' (VHRnet), is core to VHIP functionality. Each data point that underlies the VHIP training and testing represents a lab-tested virus-host pair in VHRnet, from which meaningful signals of viral adaptation to host were computed from genomic sequences. VHIP departs from existing virus-host prediction models in its ability to predict multiple interactions rather than predicting a single most likely host or host clade. As a result, VHIP is able to infer the complexity of virus-host networks in natural systems. VHIP has an 87.8% accuracy rate at predicting interactions between virus-host pairs at the species level and can be applied to novel viral and host population genomes reconstructed from metagenomic datasets. Author summary: The ecology and evolution of microbial communities are deeply influenced by viruses. Metagenomics analysis, the non-targeted sequencing of community genomes, has led to the discovery of millions of novel viruses. Yet, through the sequencing process, only DNA sequences are recovered, begging the question: which microbial hosts do those novel viruses infect? To address this question, we developed a computational tool to allow researchers to predict virus-host interactions from such sequence data. The power of this tool is its use of a high-value, manually curated set of 8849 lab-verified virus-host pairs and their corresponding sequence data. For each pair, we computed signals of coevolution to use as the predictive features in a machine learning model designed to predict interactions between viruses and hosts. The resulting model, Virus-Host Interaction Predictor (VHIP), has an accuracy of 87.8% and can be applied to novel viral and host genomes reconstructed from metagenomic datasets. Because the model considers all possible virus-host pairs, it can resolve complete virus-host interaction networks and supports a new avenue to apply network thinking to viral ecology. [ABSTRACT FROM AUTHOR]

How can inter-individual variability be quantified? Measuring many features per experiment raises the question of choosing them to recapitulate high-dimensional data. Tackling this challenge on spindle elongation phenotypes, we showed that only three typical elongation patterns describe spindle elongation in C. elegans one-cell embryo. These archetypes, automatically extracted from the experimental data using principal component analysis (PCA), accounted for more than 95% of inter-individual variability of more than 1600 experiments across more than 100 different conditions. The two first archetypes were related to spindle average length and anaphasic elongation rate. The third archetype, accounting for 6% of the variability, was novel and corresponded to a transient spindle shortening in late metaphase, reminiscent of kinetochore function-defect phenotypes. Importantly, these three archetypes were robust to the choice of the dataset and were found even considering only non-treated conditions. Thus, the inter-individual differences between genetically perturbed embryos have the same underlying nature as natural inter-individual differences between wild-type embryos, independently of the temperatures. We thus propose that beyond the apparent complexity of the spindle, only three independent mechanisms account for spindle elongation, weighted differently in the various conditions. Interestingly, the spindle-length archetypes covered both metaphase and anaphase, suggesting that spindle elongation in late metaphase is sufficient to predict the late anaphase length. We validated this idea using a machine-learning approach. Finally, given amounts of these three archetypes could represent a quantitative phenotype. To take advantage of this, we set out to predict interacting genes from a seed based on the PCA coefficients. We exemplified this firstly on the role of tpxl-1 whose homolog tpx2 is involved in spindle microtubule branching, secondly the mechanism regulating metaphase length, and thirdly the central spindle players which set the length at anaphase. We found novel interactors not in public databases but supported by recent experimental publications. Author summary: When quantifying the cell, scientists need to accurately measure cell-to-cell variability as it carries unique information about the underlying mechanisms. In this article, we focused on the spindle length as a proxy for correct mitosis. We used the nematode one-cell embryo, an established model organism, to investigate (stem-like) cell divisions. Through a data-only approach of spindle dynamics over time, we automatically extracted the most informative variability descriptors. We recalled two known ones: spindle length and anaphase elongation rate. We uncovered a new one –late-metaphase shortening –present in all conditions. Such a phenotype was previously confined to cells with defective chromosome attachments. These three descriptors account for 95% of variability, suggesting that the complex spindle choreography relies only on a few core mechanisms. Furthermore, we showed that the final spindle length at anaphase, important to set the daughter cell fates, is already determined in late metaphase, despite a complete spindle reassembly between both phases. Interestingly, the same descriptors explain variability in genetically perturbed and non-treated conditions. This suggests that no novel mechanism appears in defective cells. Only mechanism contributions are changing. Finally, we propose a tool to predict genes co-involved in a mechanism from a known gene to support candidate approaches. [ABSTRACT FROM AUTHOR]

Humans tend to give more weight to information confirming their beliefs than to information that disconfirms them. Nevertheless, this apparent irrationality has been shown to improve individual decision-making under uncertainty. However, little is known about this bias' impact on decision-making in a social context. Here, we investigate the conditions under which confirmation bias is beneficial or detrimental to decision-making under social influence. To do so, we develop a Collective Asymmetric Reinforcement Learning (CARL) model in which artificial agents observe others' actions and rewards, and update this information asymmetrically. We use agent-based simulations to study how confirmation bias affects collective performance on a two-armed bandit task, and how resource scarcity, group size and bias strength modulate this effect. We find that a confirmation bias benefits group learning across a wide range of resource-scarcity conditions. Moreover, we discover that, past a critical bias strength, resource abundance favors the emergence of two different performance regimes, one of which is suboptimal. In addition, we find that this regime bifurcation comes with polarization in small groups of agents. Overall, our results suggest the existence of an optimal, moderate level of confirmation bias for decision-making in a social context. Author summary: When we give more weight to information that confirms our existing beliefs, it typically has a negative impact on learning and decision-making. However, our study shows that a moderate confirmation bias can actually improve decision-making when multiple reinforcement learning agents learn together in a social context. This finding has important implications for policymakers who engage in fighting against societal polarization and the spreading of misinformation. It can also inspire the development of artificial, distributed learning algorithms. Based on our research, we recommend not directly targeting confirmation bias but instead focusing on its underlying factors, such as group size, individual incentives, and the interactions between bias and the environment (such as filter bubbles). [ABSTRACT FROM AUTHOR]

Joint species distribution modelling (JSDM) is a widely used statistical method that analyzes combined patterns of all species in a community, linking empirical data to ecological theory and enhancing community-wide prediction tasks. However, fitting JSDMs to large datasets is often computationally demanding and time-consuming. Recent studies have introduced new statistical and machine learning techniques to provide more scalable fitting algorithms, but extending these to complex JSDM structures that account for spatial dependencies or multi-level sampling designs remains challenging. In this study, we aim to enhance JSDM scalability by leveraging high-performance computing (HPC) resources for an existing fitting method. Our work focuses on the Hmsc R-package, a widely used JSDM framework that supports the integration of various dataset types into a single comprehensive model. We developed a GPU-compatible implementation of its model-fitting algorithm using Python and the TensorFlow library. Despite these changes, our enhanced framework retains the original user interface of the Hmsc R-package. We evaluated the performance of the proposed implementation across various model configurations and dataset sizes. Our results show a significant increase in model fitting speed for most models compared to the baseline Hmsc R-package. For the largest datasets, we achieved speed-ups of over 1000 times, demonstrating the substantial potential of GPU porting for previously CPU-bound JSDM software. This advancement opens promising opportunities for better utilizing the rapidly accumulating new biodiversity data resources for inference and prediction. [ABSTRACT FROM AUTHOR]

Neural activity in the cortex exhibits a wide range of firing variability and rich correlation structures. Studies on neural coding indicate that correlated neural variability can influence the quality of neural codes, either beneficially or adversely. However, the mechanisms by which correlated neural variability is transformed and processed across neural populations to achieve meaningful computation remain largely unclear. Here we propose a theory of covariance computation with spiking neurons which offers a unifying perspective on neural representation and computation with correlated noise. We employ a recently proposed computational framework known as the moment neural network to resolve the nonlinear coupling of correlated neural variability with a task-driven approach to constructing neural network models for performing covariance-based perceptual tasks. In particular, we demonstrate how perceptual information initially encoded entirely within the covariance of upstream neurons' spiking activity can be passed, in a near-lossless manner, to the mean firing rate of downstream neurons, which in turn can be used to inform inference. The proposed theory of covariance computation addresses an important question of how the brain extracts perceptual information from noisy sensory stimuli to generate a stable perceptual whole and indicates a more direct role that correlated variability plays in cortical information processing. Author summary: Understanding how the brain represents and processes perceptual information through neuronal firing patterns is at the heart of neuroscience. The prevailing idea suggests that the information is primarily encoded in mean firing rates, whereas correlations among neurons may play a secondary role. However, given that firing variability is ubiquitously observed in cortical neurons, one wonders if correlated noise may play a more central role in neural computation than previously thought. Here, we propose that perceptual information can be encoded in part or even entirely in the correlated variability of spiking neurons. Through a combination of theoretical modeling and machine learning approaches, we construct neural network models capable of processing correlated variability in a task-driven way. We demonstrate that the trained network is able to learn to extract covariance-encoded perceptual information to generate stimulus-selectivity in their mean firing rates, thanks to the nonlinear coupling of statistical moments of their activity. Information-theoretic analysis reveals a near-lossless transfer of perceptual information from the covariance of upstream neurons to the mean firing rate of downstream neurons. Our work offers new insights into the role of correlated variability in cortical processing and hints towards a task-driven paradigm for studying cortical computation with biologically plausible neural network models. [ABSTRACT FROM AUTHOR]

Molecules of the Major Histocompatibility Complex (MHC) present short protein fragments on the cell surface, an important step in T cell immune recognition. MHC-I molecules process peptides from intracellular proteins; MHC-II molecules act in antigen-presenting cells and present peptides derived from extracellular proteins. Here we show that the sequence-dependent energy landscapes of MHC-peptide binding encode class-specific nonlinearities (epistasis). MHC-I has a smooth landscape with global epistasis; the binding energy is a simple deformation of an underlying linear trait. This form of epistasis enhances the discrimination between strong-binding peptides. In contrast, MHC-II has a rugged landscape with idiosyncratic epistasis: binding depends on detailed amino acid combinations at multiple positions of the peptide sequence. The form of epistasis affects the learning of energy landscapes from training data. For MHC-I, a low-complexity problem, we derive a simple matrix model of binding energies that outperforms current models trained by machine learning. For MHC-II, higher complexity prevents learning by simple regression methods. Epistasis also affects the energy and fitness effects of mutations in antigen-derived peptides (epitopes). In MHC-I, large-effect mutations occur predominantly in anchor positions of strong-binding epitopes. In MHC-II, large effects depend on the background epitope sequence but are broadly distributed over the epitope, generating a bigger target for escape mutations due to loss of presentation. Together, our analysis shows how an energy landscape of protein-protein binding constrains the target of escape mutations from T cell immunity, linking the complexity of the molecular interactions to the dynamics of adaptive immune response. Author summary: T cell immunity involves the binding of short peptides to the intracellular MHC recognition machinery. Understanding how the peptide-MHC binding energy depends on the peptide sequence is key to computationally predict immune recognition and immune escape evolution, for example, of pathogens and cancer cells. By analyzing experimental binding affinity data, we find that peptide-MHC interactions are described by nonlinear energy landscapes that depend on the recognition pathway: smooth and easy to learn for MHC class I, rugged and difficult to learn for MHC class II. The different forms of nonlinearity—or epistasis—suggest the most suitable strategy to infer these landscapes. In addition, the nonlinearity has direct implications on the energy and fitness of single mutations, generating distinct distributions of possible targets for immune escape. Together, this work establishes links between biophysical origin, nonlinear structure, learnability from data, and biological implications of peptide-MHC interactions, which can be applicable also to other protein interaction landscapes. [ABSTRACT FROM AUTHOR]

MicroRNAs (miRNAs) are small non-coding RNAs that regulate gene expression post-transcriptionally. In animals, this regulation is achieved via base-pairing with partially complementary sequences on mainly 3' UTR region of messenger RNAs (mRNAs). Computational approaches that predict miRNA target interactions (MTIs) facilitate the process of narrowing down potential targets for experimental validation. The availability of new datasets of high-throughput, direct MTIs has led to the development of machine learning (ML) based methods for MTI prediction. To train an ML algorithm, it is beneficial to provide entries from all class labels (i.e., positive and negative). Currently, no high-throughput assays exist for capturing negative examples. Therefore, current ML approaches must rely on either artificially generated or inferred negative examples deduced from experimentally identified positive miRNA-target datasets. Moreover, the lack of uniform standards for generating such data leads to biased results and hampers comparisons between studies. In this comprehensive study, we collected methods for generating negative data for animal miRNAs–target interactions and investigated their impact on the classification of true human MTIs. Our study relies on training ML models on a fixed positive dataset in combination with different negative datasets and evaluating their intra- and cross-dataset performance. As a result, we were able to examine each method independently and evaluate ML models' sensitivity to the methodologies utilized in negative data generation. To achieve a deep understanding of the performance results, we analyzed unique features that distinguish between datasets. In addition, we examined whether one-class classification models that utilize solely positive interactions for training are suitable for the task of MTI classification. We demonstrate the importance of negative data in MTI classification, analyze specific methodological characteristics that differentiate negative datasets, and highlight the challenge of ML models generalizing interaction rules from training to testing sets derived from different approaches. This study provides valuable insights into the computational prediction of MTIs that can be further used to establish standards in the field. Author summary: Gene expression regulation is fundamental for all organisms' development, homeostasis, and environmental adaptation. microRNAs (miRNAs) play a central role in post-transcriptional gene regulation by binding to target mRNAs and repressing their translation or mediating their degradation. Technical challenges in experimental miRNA target identification led to growing interest in computational target prediction. While machine learning (ML) models have shown success in this area, they rely heavily on artificially generated negative examples due to limited experimental data. The diversity of methods for generating negative interactions and the lack of a uniform standardized approach introduce bias and hinder the comparison of results across different studies. In this study, we collected methods for generating negative data for animal miRNAs–target interactions and analyzed their impact on classifying true interactions in humans. Using an ML approach, we evaluated miRNA–target prediction performance within and across negative datasets. We also analyzed unique features distinguishing between the negative datasets to understand the performance results better. Our study shows that negative data is essential for accurately classifying miRNA–target interactions and that ML models struggle to apply what they have learned from the training set when faced with new data derived from different approaches. This study particularly appeals to researchers interested in miRNA–target classification. It emphasizes the need for standardized methods to enhance comparability between studies. [ABSTRACT FROM AUTHOR]

Synthetic lethality (SL) and synthetic viability (SV) are commonly studied genetic interactions in the targeted therapy approach in cancer. In SL, inhibiting either of the genes does not affect the cancer cell survival, but inhibiting both leads to a lethal phenotype. In SV, inhibiting the vulnerable gene makes the cancer cell sick; inhibiting the partner gene rescues and promotes cell viability. Many low and high-throughput experimental approaches have been employed to identify SLs and SVs, but they are time-consuming and expensive. The computational tools for SL prediction involve statistical and machine-learning approaches. Almost all machine learning tools are binary classifiers and involve only identifying SL pairs. Most importantly, there are limited properties known that best describe and discriminate SL from SV. We developed MAGICAL (Multi-class Approach for Genetic Interaction in Cancer via Algorithm Learning), a multi-class random forest based machine learning model for genetic interaction prediction. Network properties of protein derived from physical protein-protein interactions are used as features to classify SL and SV. The model results in an accuracy of ~80% for the training dataset (CGIdb, BioGRID, and SynLethDB) and performs well on DepMap and other experimentally derived reported datasets. Amongst all the network properties, the shortest path, average neighbor2, average betweenness, average triangle, and adhesion have significant discriminatory power. MAGICAL is the first multi-class model to identify discriminatory features of synthetic lethal and viable interactions. MAGICAL can predict SL and SV interactions with better accuracy and precision than any existing binary classifier. Author summary: Targeted therapy aims to selectively target cancer cells without damaging the normal ones. Synthetic lethality is a negative genetic interaction in which alteration of both genes leads to cell death and mediates drug sensitivity. In contrast, synthetic viability is a positive genetic interaction in which gene alteration rescues the cell sickness induced by alteration in the vulnerable gene and promotes cell viability, leading to drug resistance. Hence, identifying these genetic interactions is crucial to fostering selective treatment and improving the patient's health. We have designed MAGICAL, a multi-class classifier for predicting genetic interactions, a machine-learning model that can predict SL and SV based on the network properties. We aim to address how these genetic interactions get affected when the placement of the nodes (genes) in the network changes. As genetic interaction in cancer has a key role in precision oncology/targeted therapy, this work would enable researchers to understand how these interactions foster better treatment. [ABSTRACT FROM AUTHOR]

Circular RNAs (circRNAs) play vital roles in transcription and translation. Identification of circRNA-RBP (RNA-binding protein) interaction sites has become a fundamental step in molecular and cell biology. Deep learning (DL)-based methods have been proposed to predict circRNA-RBP interaction sites and achieved impressive identification performance. However, those methods cannot effectively capture long-distance dependencies, and cannot effectively utilize the interaction information of multiple features. To overcome those limitations, we propose a DL-based model iCRBP-LKHA using deep hybrid networks for identifying circRNA-RBP interaction sites. iCRBP-LKHA adopts five encoding schemes. Meanwhile, the neural network architecture, which consists of large kernel convolutional neural network (LKCNN), convolutional block attention module with one-dimensional convolution (CBAM-1D) and bidirectional gating recurrent unit (BiGRU), can explore local information, global context information and multiple features interaction information automatically. To verify the effectiveness of iCRBP-LKHA, we compared its performance with shallow learning algorithms on 37 circRNAs datasets and 37 circRNAs stringent datasets. And we compared its performance with state-of-the-art DL-based methods on 37 circRNAs datasets, 37 circRNAs stringent datasets and 31 linear RNAs datasets. The experimental results not only show that iCRBP-LKHA outperforms other competing methods, but also demonstrate the potential of this model in identifying other RNA-RBP interaction sites. Author summary: The interaction between circRNAs and RBPs is one of the main activities of circRNAs. CircRNAs participate in the occurrence and development of diseases by interacting with RBPs. Identifying circRNA-RBP interaction sites have become a fundamental step for exploring the role of circRNA in the occurrence and progression of diseases. Many computational methods have been proposed to predict circRNA-RBP interaction sites. Nevertheless, they still have several limitations. For long nucleotide sequence data of circRNA, traditional CNN or LSTM cannot effectively capture long-distance dependencies (relationships between non-adjacent nucleotides in a circRNA). Furthermore, existing methods fail to effectively utilize the interaction information of multiple features, and insufficient consideration of interaction information leads to biased circRNA-RBP interaction relationships. To overcome these limitations, we propose iCRBP-LKHA, based on a large convolutional kernel and hybrid channel-spatial attention for identifying circRNA-RBP interaction sites. We compared its performance with state-of-the-art DL-based methods on 37 circRNAs datasets, 37 circRNAs stringent datasets and 31 linear RNAs datasets. Experimental results not only show that iCRBP-LKHA outperforms competing methods, but also demonstrate the potential of this model in identifying other RNA-RBP interaction sites. [ABSTRACT FROM AUTHOR]

There remains a clinical need for better approaches to rapid drug susceptibility testing in view of the increasing burden of multidrug resistant tuberculosis. Binary susceptibility phenotypes only capture changes in minimum inhibitory concentration when these cross the critical concentration, even though other changes may be clinically relevant. We developed a machine learning system to predict minimum inhibitory concentration from unassembled whole-genome sequencing data for 13 anti-tuberculosis drugs. We trained, validated and tested the system on 10,859 isolates from the CRyPTIC dataset. Essential agreement rates (predicted MIC within one doubling dilution of observed MIC) were above 92% for first-line drugs, 91% for fluoroquinolones and aminoglycosides, and 90% for new and repurposed drugs, albeit with a significant drop in performance for the very few phenotypically resistant isolates in the latter group. To further validate the model in the absence of external MIC datasets, we predicted MIC and converted values to binary for an external set of 15,239 isolates with binary phenotypes, and compare their performance against a previously validated mutation catalogue, the expected performance of existing molecular assays, and World Health Organization Target Product Profiles. The sensitivity of the model on the external dataset was greater than 90% for all drugs except ethionamide, clofazimine and linezolid. Specificity was greater than 95% for all drugs except ethambutol, ethionamide, bedaquiline, delamanid and clofazimine. The proposed system can provide quantitative susceptibility phenotyping to help guide antimicrobial therapy, although further data collection and validation are required before machine learning can be used clinically for all drugs. Author summary: Tuberculosis is a leading cause of morbidity and mortality globally, killing over 1.5 million people each year. Tuberculosis drug susceptibility testing is used to determine which antibiotics should be used to manage the disease, particularly in view of the rise in multidrug resistant tuberculosis. The current standard for testing, binary phenotypes, only capture changes in minimum inhibitory concentration when these cross a critical concentration, even though other changes may be clinically relevant. We developed a machine learning system to predict minimum inhibitory concentration directly from sequencing data for 13 antibiotics. We find that the model has an accuracy of 92% for first-line drugs, 91% for fluoroquinolones and aminoglycosides, and 90% for new and repurposed drugs, with a significant drop in performance for resistant isolates in the latter group. We validated the model on an external dataset with binary phenotypes. The sensitivity of the model on the external dataset was greater than 90% for all drugs except ethionamide, clofazimine and linezolid. Specificity was greater than 95% for all drugs except ethambutol, ethionamide, bedaquiline, delamanid and clofazimine. The proposed system could help guide antimicrobial therapy. However, further data and validation are required before machine learning can be used clinically for all drugs. [ABSTRACT FROM AUTHOR]

Bacteriophages (phages) are viruses that infect bacteria. Many of them produce specific enzymes called depolymerases to break down external polysaccharide structures. Accurate annotation and domain identification of these depolymerases are challenging due to their inherent sequence diversity. Hence, we present DepoScope, a machine learning tool that combines a fine-tuned ESM-2 model with a convolutional neural network to identify depolymerase sequences and their enzymatic domains precisely. To accomplish this, we curated a dataset from the INPHARED phage genome database, created a polysaccharide-degrading domain database, and applied sequential filters to construct a high-quality dataset, which is subsequently used to train DepoScope. Our work is the first approach that combines sequence-level predictions with amino-acid-level predictions for accurate depolymerase detection and functional domain identification. In that way, we believe that DepoScope can greatly enhance our understanding of phage-host interactions at the level of depolymerases. [ABSTRACT FROM AUTHOR]

This article discusses the importance of incorporating ethics and societal impact considerations into data science research. It outlines 10 rules for building and maintaining a responsible data science workflow, including considering potential harms, questioning data inputs for biases and privacy concerns, and maintaining transparency and diversity in research teams. The article emphasizes the need for transparency, reproducibility, and documentation in data science research, as well as the use of open-source tools for detecting biases and unfairness. It acknowledges the challenges involved in implementing a responsible workflow but highlights the benefits for both the research team and the wider community. [Extracted from the article]

The classification of B cell lymphomas—mainly based on light microscopy evaluation by a pathologist—requires many years of training. Since the B cell receptor (BCR) of the lymphoma clonotype and the microenvironmental immune architecture are important features discriminating different lymphoma subsets, we asked whether BCR repertoire next-generation sequencing (NGS) of lymphoma-infiltrated tissues in conjunction with machine learning algorithms could have diagnostic utility in the subclassification of these cancers. We trained a random forest and a linear classifier via logistic regression based on patterns of clonal distribution, VDJ gene usage and physico-chemical properties of the top-n most frequently represented clonotypes in the BCR repertoires of 620 paradigmatic lymphoma samples—nodular lymphocyte predominant B cell lymphoma (NLPBL), diffuse large B cell lymphoma (DLBCL) and chronic lymphocytic leukemia (CLL)—alongside with 291 control samples. With regard to DLBCL and CLL, the models demonstrated optimal performance when utilizing only the most prevalent clonotype for classification, while in NLPBL—that has a dominant background of non-malignant bystander cells—a broader array of clonotypes enhanced model accuracy. Surprisingly, the straightforward logistic regression model performed best in this seemingly complex classification problem, suggesting linear separability in our chosen dimensions. It achieved a weighted F1-score of 0.84 on a test cohort including 125 samples from all three lymphoma entities and 58 samples from healthy individuals. Together, we provide proof-of-concept that at least the 3 studied lymphoma entities can be differentiated from each other using BCR repertoire NGS on lymphoma-infiltrated tissues by a trained machine learning model. Author summary: Lymphoma, a complex group of malignant blood cancers, poses a significant diagnostic challenge due to its diverse subtypes. Yet, precise classification is crucial for tailored treatment. In our research, we developed a machine learning algorithm and conducted comprehensive validation to discern distinct B cell lymphoma subtypes. We therefore leveraged B cell repertoires of lymphoma-infiltrated tissue, as ascertained through next-generation sequencing. Our data offers three key insights: We detail the creation and training of our machine learning algorithm, explaining how we selected features and designed the model. We demonstrate the algorithm's diagnostic precision using sequencing data from a test-set of patient samples. Moreover, through a deep dive into the most distinguishing aspects of our algorithm, we unveil distinctive disease-related patterns present within the malignant B cell and its surrounding environment. This analysis showed that both the malignant lymphoma cell, but also healthy bystander immune cells contribute to the distinctive architecture that characterizes a specific lymphoma subtype. We hope our work will contribute towards creating tools to diagnose lymphoma more easily and accurately ultimately leading to better outcomes for patients with this type of cancer. [ABSTRACT FROM AUTHOR]

Computational fluid dynamics (CFD) can be used for non-invasive evaluation of hemodynamics. However, its routine use is limited by labor-intensive manual segmentation, CFD mesh creation, and time-consuming simulation. This study aims to train a deep learning model to both generate patient-specific volume-meshes of the pulmonary artery from 3D cardiac MRI data and directly estimate CFD flow fields. This proof-of-concept study used 135 3D cardiac MRIs from both a public and private dataset. The pulmonary arteries in the MRIs were manually segmented and converted into volume-meshes. CFD simulations were performed on ground truth meshes and interpolated onto point-point correspondent meshes to create the ground truth dataset. The dataset was split 110/10/15 for training, validation, and testing. Image2Flow, a hybrid image and graph convolutional neural network, was trained to transform a pulmonary artery template to patient-specific anatomy and CFD values, taking a specific inlet velocity as an additional input. Image2Flow was evaluated in terms of segmentation, and the accuracy of predicted CFD was assessed using node-wise comparisons. In addition, the ability of Image2Flow to respond to increasing inlet velocities was also evaluated. Image2Flow achieved excellent segmentation accuracy with a median Dice score of 0.91 (IQR: 0.86–0.92). The median node-wise normalized absolute error for pressure and velocity magnitude was 11.75% (IQR: 9.60–15.30%) and 9.90% (IQR: 8.47–11.90), respectively. Image2Flow also showed an expected response to increased inlet velocities with increasing pressure and velocity values. This proof-of-concept study has shown that it is possible to simultaneously perform patient-specific volume-mesh based segmentation and pressure and flow field estimation using Image2Flow. Image2Flow completes segmentation and CFD in ~330ms, which is ~5000 times faster than manual methods, making it more feasible in a clinical environment. Author summary: Computational fluid dynamics is an engineering tool that can be used in a clinical setting to non-invasively model blood flow through blood vessels, such as the pulmonary artery. This information can be used to inform treatment planning for patients, especially those with cardiovascular conditions. However, its routine use is limited by a labor-intensive process, requiring substantial expertise and computational resources. Recently, machine learning has offered solutions to automate parts of this process, but no single model has addressed the entire workflow. Therefore, we created Image2Flow, a machine learning model capable of generating a patient-specific representation of a pulmonary artery while predicting blood flow through the vessel. Image2Flow can generate highly accurate pulmonary artery representations and reasonably accurate blood flow predictions. Notably, it performs this whole process in 330ms, which is ~5000x faster than the manual methods. With its speed and accuracy, Image2Flow holds substantial promise for facilitating computational fluid dynamics in clinical settings. [ABSTRACT FROM AUTHOR]

Tumor microenvironments (TMEs) contain vast amounts of information on patient's cancer through their cellular composition and the spatial distribution of tumor cells and immune cell populations. Exploring variations in TMEs between patient groups, as well as determining the extent to which this information can predict outcomes such as patient survival or treatment success with emerging immunotherapies, is of great interest. Moreover, in the face of a large number of cell interactions to consider, we often wish to identify specific interactions that are useful in making such predictions. We present an approach to achieve these goals based on summarizing spatial relationships in the TME using spatial K functions, and then applying functional data analysis and random forest models to both predict outcomes of interest and identify important spatial relationships. This approach is shown to be effective in simulation experiments at both identifying important spatial interactions while also controlling the false discovery rate. We further used the proposed approach to interrogate two real data sets of Multiplexed Ion Beam Images of TMEs in triple negative breast cancer and lung cancer patients. The methods proposed are publicly available in a companion R package funkycells. Author summary: Spatial data on the tumor microenvironment (TME) are becoming more prevalent. Existing methods to interrogate such data often have several limitations: (1) they can rely on estimating the spatial relationships among cells by examining simple counts of cells within a single radius, (2) they may not come with ways to evaluate the statistical significance of any findings, or (3) they model individual interactions independently of other interactions. Our approach leverages techniques in spatial statistics and uses a benchmark ensemble machine learning method to address each of these deficiencies; it (1) uses K functions to encode the relative densities of cells over all radii up to a user-selected maximum radius, (2) employs permutation and cross-validation to evaluate the statistical significance of any findings on the spatial interactions in the TME, and (3) models multiple interactions simultaneously. Our approach is freely available with an R implementation called funkycells. In the analysis of two real data sets, we have seen that the method performs well, and gives the expected results. We think this will be a robust tool for researchers looking to interrogate TME data. [ABSTRACT FROM AUTHOR]

Immunization through repeated direct venous inoculation of Plasmodium falciparum (Pf) sporozoites (PfSPZ) under chloroquine chemoprophylaxis, using the PfSPZ Chemoprophylaxis Vaccine (PfSPZ-CVac), induces high-level protection against controlled human malaria infection (CHMI). Humoral and cellular immunity contribute to vaccine efficacy but only limited information about the implicated Pf-specific antigens is available. Here, we examined Pf-specific antibody profiles, measured by protein arrays representing the full Pf proteome, of 40 placebo- and PfSPZ-immunized malaria-naïve volunteers from an earlier published PfSPZ-CVac dose-escalation trial. For this purpose, we both utilized and adapted supervised machine learning methods to identify predictive antibody profiles at two different time points: after immunization and before CHMI. We developed an adapted multitask support vector machine (SVM) approach and compared it to standard methods, i.e. single-task SVM, regularized logistic regression and random forests. Our results show, that the multitask SVM approach improved the classification performance to discriminate the protection status based on the underlying antibody-profiles while combining time- and dose-dependent data in the prediction model. Additionally, we developed the new fEature diStance exPlainabilitY (ESPY) method to quantify the impact of single antigens on the non-linear multitask SVM model and make it more interpretable. In conclusion, our multitask SVM model outperforms the studied standard approaches in regard of classification performance. Moreover, with our new explanation method ESPY, we were able to interpret the impact of Pf-specific antigen antibody responses that predict sterile protective immunity against CHMI after immunization. The identified Pf-specific antigens may contribute to a better understanding of immunity against human malaria and may foster vaccine development. Author summary: Developing an effective malaria vaccine is challenging. Malaria is a life-threatening disease caused by the plasmodium parasite, which has a complex multi-stage life-cycle and expresses several thousand proteins in a highly coordinated manner. To date, our understanding of the immune mechanisms mediating protection against Plasmodium falciparum (Pf) is incomplete. Proteome microarrays have been used earlier by our clinical collaboration partners to identify Pf-specific antibody profiles of malaria-naïve volunteers during immunization with attenuated Pf sporozoites (PfSPZ). We reused this data to compare the ability of three supervised machine learning methods to identify predictive antibody profiles after immunization and before controlled human malaria infection (CHMI). We adapted a multitask support vector machine (SVM) approach to analyze time-dependent Pf-induced antibody profiles from several time points in a single prediction model. Our multitask SVM approach outperforms the studied standard approaches in classification performance. Additionally, we developed a new explanation method, named fEature diStance exPlainabilitY (ESPY), to interpret the impact of Pf-specific antigens. We applied ESPY on the multitask SVM model and identified diverse Pf-specific antigen sets after immunization and before CHMI. Furthermore, we showed that the identified Pf-induced antibody profiles vary among protected and non-protected individuals who had been exposed to different doses of PfSPZ. [ABSTRACT FROM AUTHOR]

Multi-factor screenings are commonly used in diverse applications in medicine and bioengineering, including optimizing combination drug treatments and microbiome engineering. Despite the advances in high-throughput technologies, large-scale experiments typically remain prohibitively expensive. Here we introduce a machine learning platform, structure-augmented regression (SAR), that exploits the intrinsic structure of each biological system to learn a high-accuracy model with minimal data requirement. Under different environmental perturbations, each biological system exhibits a unique, structured phenotypic response. This structure can be learned based on limited data and once learned, can constrain subsequent quantitative predictions. We demonstrate that SAR requires significantly fewer data comparing to other existing machine-learning methods to achieve a high prediction accuracy, first on simulated data, then on experimental data of various systems and input dimensions. We then show how a learned structure can guide effective design of new experiments. Our approach has implications for predictive control of biological systems and an integration of machine learning prediction and experimental design. Author summary: Using limited data to predict how biological systems respond to combinations of perturbations is important for better understanding of biological systems and for practical applications. Here we present a new method, structure-augmented regression, that efficiently learns biological response landscapes from sparse measurements. Our method exploits the property that many biological response landscapes have distinct lower-dimensional structures, which can be learned first to assist the subsequent quantitative predictions. We demonstrate that our algorithm outperforms existing algorithms on simulated and experimental data of various biological systems and dimensions. We further exploit the learned structure by suggesting new experiments that refines the learned structure to further improve the prediction accuracy. By integrating machine learning with experimental design, our approach has implications for the predictive control of biological systems. [ABSTRACT FROM AUTHOR]

Unit activity in particular deep neural networks (DNNs) are remarkably similar to the neuronal population responses to static images along the primate ventral visual cortex. Linear combinations of DNN unit activities are widely used to build predictive models of neuronal activity in the visual cortex. Nevertheless, prediction performance in these models is often investigated on stimulus sets consisting of everyday objects under naturalistic settings. Recent work has revealed a generalization gap in how predicting neuronal responses to synthetically generated out-of-distribution (OOD) stimuli. Here, we investigated how the recent progress in improving DNNs' object recognition generalization, as well as various DNN design choices such as architecture, learning algorithm, and datasets have impacted the generalization gap in neural predictivity. We came to a surprising conclusion that the performance on none of the common computer vision OOD object recognition benchmarks is predictive of OOD neural predictivity performance. Furthermore, we found that adversarially robust models often yield substantially higher generalization in neural predictivity, although the degree of robustness itself was not predictive of neural predictivity score. These results suggest that improving object recognition behavior on current benchmarks alone may not lead to more general models of neurons in the primate ventral visual cortex. Author summary: Inspired by the neural circuits of the brain, deep neural networks (DNN) have been steadily improving in their ability to perform foundational visual tasks such as object recognition. Whereas, early models struggled with generalization to abstract visual domains such as line drawings and cartoons, recent advancement have approached near-human recognition capabilities. Moreover, the unit activity in these networks exhibit strong similarities with the activity of single-unit recordings along the primate ventral visual cortex. This capability of DNNs has provided visual neuroscientists with precise models for exploring the neural underpinnings of object recognition. Our research probes whether enhancements in neural networks' recognition of out-of-distribution objects correlate with improved predictability of brain activity in the visual cortex of monkeys to synthetic stimuli. We found that the out of distribution object recognition performance on natural image datasets is not a reliable measure of neural predictivity. However, DNN models that were trained to be more resilient to adversarially generated noise patterns as well as DNN ensembles, consistently yielded better generalization in neural predictivity. Altogether, our results suggest that improving object recognition behaviour on current benchmarks alone may not lead to more general models of neurons in the primate ventral visual cortex. [ABSTRACT FROM AUTHOR]

Animals can learn in real-life scenarios where rewards are often only available when a goal is achieved. This 'distal' or 'sparse' reward problem remains a challenge for conventional reinforcement learning algorithms. Here we investigate an algorithm for learning in such scenarios, inspired by the possibility that axo-axonal gap junction connections, observed in neural circuits with parallel fibres such as the insect mushroom body, could form a resistive network. In such a network, an active node represents the task state, connections between nodes represent state transitions and their connection to actions, and current flow to a target state can guide decision making. Building on evidence that gap junction weights are adaptive, we propose that experience of a task can modulate the connections to form a graph encoding the task structure. We demonstrate that the approach can be used for efficient reinforcement learning under sparse rewards, and discuss whether it is plausible as an account of the insect mushroom body. Author summary: Learning in situations where reward is only rarely encountered is difficult. It is hard to discover the right sequence of actions when most actions, most of the time, provide no apparent progress towards a goal. Inspired by a neural circuit in the insect brain, and using direct electrical connections between neurons as well as synaptic connections, we present a new algorithm for learning. The model represents the states of the world with nodes and an electrical connection between two nodes is strengthened when the two corresponding states occur consecutively. The connections between nodes can also become associated to output actions that correlate with (hence are assumed to cause) transitions between states. When a particular goal is chosen or associated with a reward, for example, the target location in a navigation task, a flow of electrical current through the nodes will find the shortest path from the present state to the goal state and trigger the appropriate actions. [ABSTRACT FROM AUTHOR]

The activities of most enzymes and drugs depend on interactions between proteins and small molecules. Accurate prediction of these interactions could greatly accelerate pharmaceutical and biotechnological research. Current machine learning models designed for this task have a limited ability to generalize beyond the proteins used for training. This limitation is likely due to a lack of information exchange between the protein and the small molecule during the generation of the required numerical representations. Here, we introduce ProSmith, a machine learning framework that employs a multimodal Transformer Network to simultaneously process protein amino acid sequences and small molecule strings in the same input. This approach facilitates the exchange of all relevant information between the two molecule types during the computation of their numerical representations, allowing the model to account for their structural and functional interactions. Our final model combines gradient boosting predictions based on the resulting multimodal Transformer Network with independent predictions based on separate deep learning representations of the proteins and small molecules. The resulting predictions outperform recently published state-of-the-art models for predicting protein-small molecule interactions across three diverse tasks: predicting kinase inhibitions; inferring potential substrates for enzymes; and predicting Michaelis constants KM. The Python code provided can be used to easily implement and improve machine learning predictions involving arbitrary protein-small molecule interactions. Author summary: Understanding how proteins interact with small molecules, such as drugs, is critical to advancing medical, biological, and biotechnological research. Our work introduces ProSmith, a machine learning framework that improves the prediction of protein-small molecule interactions. Protein-small molecule interactions can be predicted by using numerical representations of proteins and small molecules as input to machine learning prediction models. Previous methods typically generated separate numerical representations for the proteins and small molecules without considering their interactions. ProSmith, however, combines both protein sequence and small molecule structural information in the input of a single multimodal Transformer Network to generate a joint numerical representation. Unlike previous methods, this allows for a comprehensive exchange of information between protein and small molecule, capturing the complex relationships and interactions between these two types of molecules. ProSmith successfully predicts several biological interactions, including kinase inhibitions, potential enzyme-substrate pairs, and enzyme kinetic parameters KM. We provide Python code that can be easily adapted to improve predictions for any protein-small molecule interaction. [ABSTRACT FROM AUTHOR]

Clinical trials involving systemic neoadjuvant treatments in breast cancer aim to shrink tumors before surgery while simultaneously allowing for controlled evaluation of biomarkers, toxicity, and suppression of distant (occult) metastatic disease. Yet neoadjuvant clinical trials are rarely preceded by preclinical testing involving neoadjuvant treatment, surgery, and post-surgery monitoring of the disease. Here we used a mouse model of spontaneous metastasis occurring after surgical removal of orthotopically implanted primary tumors to develop a predictive mathematical model of neoadjuvant treatment response to sunitinib, a receptor tyrosine kinase inhibitor (RTKI). Treatment outcomes were used to validate a novel mathematical kinetics-pharmacodynamics model predictive of perioperative disease progression. Longitudinal measurements of presurgical primary tumor size and postsurgical metastatic burden were compiled using 128 mice receiving variable neoadjuvant treatment doses and schedules (released publicly at https://zenodo.org/records/10607753). A non-linear mixed-effects modeling approach quantified inter-animal variabilities in metastatic dynamics and survival, and machine-learning algorithms were applied to investigate the significance of several biomarkers at resection as predictors of individual kinetics. Biomarkers included circulating tumor- and immune-based cells (i.e., circulating tumor cells and myeloid-derived suppressor cells) as well as immunohistochemical tumor proteins (i.e., CD31 and Ki67). Our computational simulations show that neoadjuvant RTKI treatment inhibits primary tumor growth but has little efficacy in preventing (micro)-metastatic disease progression after surgery and treatment cessation. Machine learning algorithms that included support vector machines, random forests, and artificial neural networks, confirmed a lack of definitive biomarkers, which shows the value of preclinical modeling studies to identify potential failures that should be avoided clinically. Author summary: Using simulations from a mechanistic mathematical model compared with preclinical data from surgical metastasis models, we found that anti-tumor effects of neoadjuvant receptor tyrosine kinase inhibitor treatment can differ between the primary tumor and metastases in the perioperative setting. Model simulations with variable drug doses and scheduling of neoadjuvant treatment revealed a contrasting impact on initial primary tumor debulking and metastatic outcomes long after treatment has stopped and the primary tumor has been surgically removed. Using machine-learning algorithms, we identified the limited power of several circulating cellular and molecular biomarkers in predicting metastatic outcomes, uncovering a potential fast-track strategy for assessing future clinical biomarkers by pairing patient studies with identical studies in mice. [ABSTRACT FROM AUTHOR]

The use of deep learning (DL) is steadily gaining traction in scientific challenges such as cancer research. Advances in enhanced data generation, machine learning algorithms, and compute infrastructure have led to an acceleration in the use of deep learning in various domains of cancer research such as drug response problems. In our study, we explored tree-based models to improve the accuracy of a single drug response model and demonstrate that tree-based models such as XGBoost (eXtreme Gradient Boosting) have advantages over deep learning models, such as a convolutional neural network (CNN), for single drug response problems. However, comparing models is not a trivial task. To make training and comparing CNNs and XGBoost more accessible to users, we developed an open-source library called UNNT (A novel Utility for comparing Neural Net and Tree-based models). The case studies, in this manuscript, focus on cancer drug response datasets however the application can be used on datasets from other domains, such as chemistry. Author summary: Advancement in data science, machine learning (ML), and artificial intelligence (AI) methods has enabled extraction of meaningful information from large and complex datasets that has assisted in better understanding, diagnosing, and treating cancer. The understanding of the drug response domain in cancer research has been accelerated with developing ML models to aid in predicting the effectiveness of the drugs based on a specific genomic molecular feature. In this study we developed a novel robust framework called UNNT (A novel Utility for comparing Neural Net and Tree-based models) that trains and compares deep learning method such as CNN and tree-based method such as XGBoost on the user input dataset. We applied this software to single drug response problem in cancer to identify the best performing ML method based on the National Cancer Institute 60 (NCI60) dataset. In addition, we studied the computational aspects of training each of these models where our results show that neither is evidently superior on both CPUs and GPUs while training. This shows that when both models have similar error rates for a dataset the hardware available determines the model choice for training. [ABSTRACT FROM AUTHOR]

Neural Cellular Automata (NCA) are a powerful combination of machine learning and mechanistic modelling. We train NCA to learn complex dynamics from time series of images and PDE trajectories. Our method is designed to identify underlying local rules that govern large scale dynamic emergent behaviours. Previous work on NCA focuses on learning rules that give stationary emergent structures. We extend NCA to capture both transient and stable structures within the same system, as well as learning rules that capture the dynamics of Turing pattern formation in nonlinear Partial Differential Equations (PDEs). We demonstrate that NCA can generalise very well beyond their PDE training data, we show how to constrain NCA to respect given symmetries, and we explore the effects of associated hyperparameters on model performance and stability. Being able to learn arbitrary dynamics gives NCA great potential as a data driven modelling framework, especially for modelling biological pattern formation. Author summary: Pattern formation is ubiquitous in biological systems, across many length and time scales—from vegetation stripes in deserts, to the spots of a leopard's skin, and the fine detail of stem cell differentiation in an embryo. While many simple rules that create complex patterns are known, reverse engineering the mechanisms responsible from an observed pattern has generally remained a difficult problem. In this work we build on the connections between machine learning, cellular automata, and partial differential equations, to create models that can learn the underlying mechanisms that yield any desired emergent pattern. To do this we consider Neural Cellular Automata (a mix of neural networks and cellular automata). We describe how they are built and trained, and we show that they are capable of reproducing a wide range of complex emergent behaviours. Previous work focuses on learning stationary patterns, but we focus on patterns that evolve in time. We show these models can learn classic Turing patterns (biochemical models of spot and stripe formation), as well as learning to morph through an arbitrary sequence of images. We believe this hybrid between machine learning and mechanistic modelling has great potential for modelling biological growth, regeneration, and pattern formation. [ABSTRACT FROM AUTHOR]

Biomedical texts provide important data for investigating drug-drug interactions (DDIs) in the field of pharmacovigilance. Although researchers have attempted to investigate DDIs from biomedical texts and predict unknown DDIs, the lack of accurate manual annotations significantly hinders the performance of machine learning algorithms. In this study, a new DDI prediction framework, Subgraph Enhance model, was developed for DDI (SubGE-DDI) to improve the performance of machine learning algorithms. This model uses drug pairs knowledge subgraph information to achieve large-scale plain text prediction without many annotations. This model treats DDI prediction as a multi-class classification problem and predicts the specific DDI type for each drug pair (e.g. Mechanism, Effect, Advise, Interact and Negative). The drug pairs knowledge subgraph was derived from a huge drug knowledge graph containing various public datasets, such as DrugBank, TwoSIDES, OffSIDES, DrugCentral, EntrezeGene, SMPDB (The Small Molecule Pathway Database), CTD (The Comparative Toxicogenomics Database) and SIDER. The SubGE-DDI was evaluated from the public dataset (SemEval-2013 Task 9 dataset) and then compared with other state-of-the-art baselines. SubGE-DDI achieves 83.91% micro F1 score and 84.75% macro F1 score in the test dataset, outperforming the other state-of-the-art baselines. These findings show that the proposed drug pairs knowledge subgraph-assisted model can effectively improve the prediction performance of DDIs from biomedical texts. Author summary: Drug-drug interactions occur when two or more drugs react with each other, potentially leading to adverse drug reactions that can manifest as toxicity, reduced treatment efficacy, and, in extreme cases, patient mortality. With the growing number of approved drugs, detecting negative drug-drug interactions has become a crucial concern in pharmacovigilance. However, manually mining these interactions from the vast and rapidly expanding pool of published biomedical texts, such as clinical trial reports and literature, is arduous and challenging. To address this issue, various machine learning methods have been developed and applied to text mining for both known and unknown drug-drug interactions. Most of these methods primarily predict interactions by considering only drug molecular properties and text characters, neglecting valuable information about interactions between different entities, such as drugs, pathways, and side effects. In response, our approach seeks to enhance the performance of drug-drug interaction prediction by leveraging knowledge graphs constructed from a variety of biomedical entities. Our experimental results validate this approach, demonstrating that incorporating interactive information from knowledge graphs leads to the development of a more efficient model for predicting drug-drug interactions. [ABSTRACT FROM AUTHOR]

Animal psychophysics can generate rich behavioral datasets, often comprised of many 1000s of trials for an individual subject. Gradient-boosted models are a promising machine learning approach for analyzing such data, partly due to the tools that allow users to gain insight into how the model makes predictions. We trained ferrets to report a target word's presence, timing, and lateralization within a stream of consecutively presented non-target words. To assess the animals' ability to generalize across pitch, we manipulated the fundamental frequency (F0) of the speech stimuli across trials, and to assess the contribution of pitch to streaming, we roved the F0 from word token-to-token. We then implemented gradient-boosted regression and decision trees on the trial outcome and reaction time data to understand the behavioral factors behind the ferrets' decision-making. We visualized model contributions by implementing SHAPs feature importance and partial dependency plots. While ferrets could accurately perform the task across all pitch-shifted conditions, our models reveal subtle effects of shifting F0 on performance, with within-trial pitch shifting elevating false alarms and extending reaction times. Our models identified a subset of non-target words that animals commonly false alarmed to. Follow-up analysis demonstrated that the spectrotemporal similarity of target and non-target words rather than similarity in duration or amplitude waveform was the strongest predictor of the likelihood of false alarming. Finally, we compared the results with those obtained with traditional mixed effects models, revealing equivalent or better performance for the gradient-boosted models over these approaches. Author summary: The sorts of listening challenges faced by real-world listeners are rarely captured by most laboratory-based auditory paradigms, particularly those testing animal models. However, many labs are attempting to utilize more realistic experiments, and more complicated behavioral paradigms require more sophisticated approaches to analyzing the resulting data. Here, we used a new behavioral paradigm to test the ability of ferret listeners to identify target speech sounds and assess their ability to generalize across changes in pitch. To make sense of the resulting dataset, we used machine learning to understand how trained ferrets perform this task. Gradient-boosted regression and decision trees are well-established machine learning methods that do not require users to predetermine interaction effects and are accompanied by visualization tools that allow insights into how multiple factors ultimately shape behavior. We compare the use of gradient-boosted models to more standard regression approaches and that this machine learning approach is ideal for analyzing behavioral data in animal models. [ABSTRACT FROM AUTHOR]

Early identification of safe and efficacious disease targets is crucial to alleviating the tremendous cost of drug discovery projects. However, existing experimental methods for identifying new targets are generally labor-intensive and failure-prone. On the other hand, computational approaches, especially the machine learning-based frameworks, have shown remarkable application potential in drug discovery. In this work, we propose Progeni, a novel machine learning-based framework for target identification. In addition to fully exploiting the known heterogeneous biological networks from various sources, Progeni integrates literature evidence about the relations between biological entities to construct a probabilistic knowledge graph. Graph neural networks are then employed in Progeni to learn the feature embeddings of biological entities to facilitate the identification of biologically relevant target candidates. A comprehensive evaluation of Progeni demonstrated its superior predictive power over the baseline methods on the target identification task. In addition, our extensive tests showed that Progeni exhibited high robustness to the negative effect of exposure bias, a common phenomenon in recommendation systems, and effectively identified new targets that can be strongly supported by the literature. Moreover, our wet lab experiments successfully validated the biological significance of the top target candidates predicted by Progeni for melanoma and colorectal cancer. All these results suggested that Progeni can identify biologically effective targets and thus provide a powerful and useful tool for advancing the drug discovery process. Author summary: The selection of a disease target, a biological entity with which potential drugs interact to treat diseases, is often the first step of a drug discovery project. Known to be extremely costly, drug discovery projects spend tremendous resources on failed drug instances, most of which result from inadequate choices of targets. To help address the costliness problem of drug discovery, we developed a machine learning-based framework that identifies biologically effective targets using a probabilistic knowledge graph built from both biological network data and literature evidence. Our method not only outperforms state-of-the-art baseline methods on the target prediction task, but can also identify targets with high biological relevance, as shown by the strong support of the literature for the predicted target candidates and wet lab experiments that validate the significance of the predicted target candidates for melanoma and colorectal cancer, These results suggest that our method can identify effective targets and therefore benefit drug discovery. [ABSTRACT FROM AUTHOR]

A key element for successful blood transfusion is compatibility of the patient and donor red blood cell (RBC) antigens. Precise antigen matching reduces the risk for immunization and other adverse transfusion outcomes. RBC antigens are encoded by specific genes, which allows developing computational methods for determining antigens from genomic data. We describe here a classification method for determining RBC antigens from genotyping array data. Random forest models for 39 RBC antigens in 14 blood group systems and for human platelet antigen (HPA)-1 were trained and tested using genotype and RBC antigen and HPA-1 typing data available for 1,192 blood donors in the Finnish Blood Service Biobank. The algorithm and models were further evaluated using a validation cohort of 111,667 Danish blood donors. In the Finnish test data set, the median (interquartile range [IQR]) balanced accuracy for 39 models was 99.9 (98.9–100)%. We were able to replicate 34 out of 39 Finnish models in the Danish cohort and the median (IQR) balanced accuracy for classifications was 97.1 (90.1–99.4)%. When applying models trained with the Danish cohort, the median (IQR) balanced accuracy for the 40 Danish models in the Danish test data set was 99.3 (95.1–99.8)%. The RBC antigen and HPA-1 prediction models demonstrated high overall accuracies suitable for probabilistic determination of blood groups and HPA-1 at biobank-scale. Furthermore, population-specific training cohort increased the accuracies of the models. This stand-alone and freely available method is applicable for research and screening for antigen-negative blood donors. Author summary: Blood transfusion is one of the most common clinical procedures in the hospitals and the key element for safe transfusion is compatibility between the recipient and donor red blood cell antigens. Precise antigen matching reduces the risk for sensitization and other adverse transfusion outcomes. Here we describe a stand-alone and freely available random forest classification method and models for determining red blood cell and human platelet antigens from array technology-based genotyping data. We investigate the performance of models trained with Finnish blood donor biobank data and further validate the method with a large Danish cohort. The results demonstrate high overall accuracy, and the method is suitable for biobank-scale research and screening of antigen-negative donors. The implementation is possible in the local computing environment without sensitive data uploads and requires only a moderate level of bioinformatic skills. [ABSTRACT FROM AUTHOR]

Post-translational modifications (PTMs) of proteins play a vital role in their function and stability. These modifications influence protein folding, signaling, protein-protein interactions, enzyme activity, binding affinity, aggregation, degradation, and much more. To date, over 400 types of PTMs have been described, representing chemical diversity well beyond the genetically encoded amino acids. Such modifications pose a challenge to the successful design of proteins, but also represent a major opportunity to diversify the protein engineering toolbox. To this end, we first trained artificial neural networks (ANNs) to predict eighteen of the most abundant PTMs, including protein glycosylation, phosphorylation, methylation, and deamidation. In a second step, these models were implemented inside the computational protein modeling suite Rosetta, which allows flexible combination with existing protocols to model the modified sites and understand their impact on protein stability as well as function. Lastly, we developed a new design protocol that either maximizes or minimizes the predicted probability of a particular site being modified. We find that this combination of ANN prediction and structure-based design can enable the modification of existing, as well as the introduction of novel, PTMs. The potential applications of our work include, but are not limited to, glycan masking of epitopes, strengthening protein-protein interactions through phosphorylation, as well as protecting proteins from deamidation liabilities. These applications are especially important for the design of new protein therapeutics where PTMs can drastically change the therapeutic properties of a protein. Our work adds novel tools to Rosetta's protein engineering toolbox that allow for the rational design of PTMs. Author summary: Machine learning (ML) is changing the world of protein design, from structure prediction methods like AlphaFold to fixed-backbone design methods like ProteinMPNN. ML methods have made much progress in various aspects of protein computational biology, both complementing and, in some cases, surpassing traditional macromolecular modeling methods such as those combined in libraries like the Rosetta software suite. However, a lack of compatibility and flexibility can hinder interoperability with existing methods, preventing the full potential of these new solutions from being realized. Here, we first present a new machine learning tool for predicting post-translational modifications (PTMs), which play an important role in the stability and function of proteins, and then highlight how the implementation of this tool in the existing Rosetta toolbox can facilitate new applications. To this end, we combine PTM prediction with protein design, maximizing or minimizing the predicted probability of a post-translational modification occurring at a specific site. As one example, we predict the N-linked glycosylation of influenza hemagglutinin, which has applications in both understanding the evolution of viral strains over time, and engineering additional glycosylation sites to mask unwanted epitopes of vaccine candidates. [ABSTRACT FROM AUTHOR]

Discovering mathematical equations that govern physical and biological systems from observed data is a fundamental challenge in scientific research. We present a new physics-informed framework for parameter estimation and missing physics identification (gray-box) in the field of Systems Biology. The proposed framework—named AI-Aristotle—combines the eXtreme Theory of Functional Connections (X-TFC) domain-decomposition and Physics-Informed Neural Networks (PINNs) with symbolic regression (SR) techniques for parameter discovery and gray-box identification. We test the accuracy, speed, flexibility, and robustness of AI-Aristotle based on two benchmark problems in Systems Biology: a pharmacokinetics drug absorption model and an ultradian endocrine model for glucose-insulin interactions. We compare the two machine learning methods (X-TFC and PINNs), and moreover, we employ two different symbolic regression techniques to cross-verify our results. To test the performance of AI-Aristotle, we use sparse synthetic data perturbed by uniformly distributed noise. More broadly, our work provides insights into the accuracy, cost, scalability, and robustness of integrating neural networks with symbolic regressors, offering a comprehensive guide for researchers tackling gray-box identification challenges in complex dynamical systems in biomedicine and beyond. Author summary: Our study addresses the fundamental challenge of uncovering mathematical rules governing physical and biological systems from real-world data. We introduce a novel framework, AI-Aristotle, designed for parameter estimation and identifying hidden physics (gray-box) in Systems Biology. AI-Aristotle combines the powerful eXtreme Theory of Functional Connections (X-TFC), Physics-Informed Neural Networks (PINNs), and symbolic regression (SR) techniques to discover parameters and uncover hidden relationships. Our work offers guidance to researchers addressing gray-box identification challenges in complex dynamic systems, including applications in biomedicine and beyond. [ABSTRACT FROM AUTHOR]

Synthetic biology dictates the data-driven engineering of biocatalysis, cellular functions, and organism behavior. Integral to synthetic biology is the aspiration to efficiently find, access, interoperate, and reuse high-quality data on genotype-phenotype relationships of native and engineered biosystems under FAIR principles, and from this facilitate forward-engineering strategies. However, biology is complex at the regulatory level, and noisy at the operational level, thus necessitating systematic and diligent data handling at all levels of the design, build, and test phases in order to maximize learning in the iterative design-build-test-learn engineering cycle. To enable user-friendly simulation, organization, and guidance for the engineering of biosystems, we have developed an open-source python-based computer-aided design and analysis platform operating under a literate programming user-interface hosted on Github. The platform is called teemi and is fully compliant with FAIR principles. In this study we apply teemi for i) designing and simulating bioengineering, ii) integrating and analyzing multivariate datasets, and iii) machine-learning for predictive engineering of metabolic pathway designs for production of a key precursor to medicinal alkaloids in yeast. The teemi platform is publicly available at PyPi and GitHub. Author summary: Bioengineering holds fantastic perspectives and is poised to change how we produce foods, materials, and medicines. However, rapid progress is limited by a lack of mechanistic knowledge in even the simplest model organisms, such as bacteria and yeast. Thus, to compensate, we often have to construct and study a large number of engineered cells and select the cells with the greatest potential for the given objective function. The targeted construction of engineered cells is often described as an iterative process of design, build, test, and learn (the DBTL cycle). Literate programming is a paradigm that encourages the combination of text and computer code with the potential to describe all workflows covered by the DBTL cycle. The purpose of the present work is to give a first estimate of the extent to which we can accelerate the individual steps in the DBTL cycle by using end-to-end literate programming workflows. To achieve this, we established an open-source platform called teemi, and used it to optimize production of a key precursor to medicinal alkaloids in yeast. We expect that teemi will enable higher DBTL throughput with fewer errors, better integration of IT tools with laboratory resources, and more effective knowledge capture. [ABSTRACT FROM AUTHOR]