Search

Introduction: Artificial intelligence (AI) is exhibiting tremendous potential to reduce the massive costs and long timescales of drug discovery. There are however important challenges currently limiting the impact and scope of AI models. Areas covered: In this perspective, the authors discuss a range of data issues (bias, inconsistency, skewness, irrelevance, small size, high dimensionality), how they challenge AI models, and which issue-specific mitigations have been effective. Next, they point out the challenges faced by uncertainty quantification techniques aimed at enhancing and trusting the predictions from these AI models. They also discuss how conceptual errors, unrealistic benchmarks and performance misestimation can confound the evaluation of models and thus their development. Lastly, the authors explain how human bias, whether from AI experts or drug discovery experts, constitutes another challenge that can be alleviated by gaining more prospective experience. Expert opinion: AI models are often developed to excel on retrospective benchmarks unlikely to anticipate their prospective performance. As a result, only a few of these models are ever reported to have prospective value (e.g. by discovering potent and innovative drug leads for a therapeutic target). The authors have discussed what can go wrong in practice with AI for drug discovery. We hope that this will help inform the decisions of editors, funders investors and researchers working in this area.

Virtual Screening (VS) of vast compound libraries guided by Artificial Intelligence (AI) models is a highly productive approach to early drug discovery. Data splitting is crucial for better benchmarking of such AI models. Traditional random data splits produce similar molecules between training and test sets, conflicting with the reality of VS libraries which mostly contain structurally distinct compounds. Scaffold split, grouping molecules by shared core structure, is widely considered to reflect this real-world scenario. However, here we show that the scaffold split also overestimates VS performance. The reason is that molecules with different chemical scaffolds are often similar, which hence introduces unrealistically high similarities between training molecules and test molecules following a scaffold split. Our study examined three representative AI models on 60 NCI-60 datasets, each with approximately 30,000 to 50,000 molecules tested on a different cancer cell line. Each dataset was split with three methods: scaffold, Butina clustering and the more accurate Uniform Manifold Approximation and Projection (UMAP) clustering. Regardless of the model, model performance is much worse with UMAP splits from the results of the 2100 models trained and evaluated for each algorithm and split. These robust results demonstrate the need for more realistic data splits to tune, compare, and select models for VS. For the same reason, avoiding the scaffold split is also recommended for other molecular property prediction problems. The code to reproduce these results is available at https://github.com/ScaffoldSplitsOverestimateVS

Artificial Neural Networks (ANN) have been popularized in many science and technological areas due to their capacity to solve many complex pattern matching problems. That is the case of Virtual Screening, a research area that studies how to identify those molecular compounds with the highest probability to present biological activity for a therapeutic target. Due to the vast number of small organic compounds and the thousands of targets for which such large-scale screening can potentially be carried out, there has been an increasing interest in the research community to increase both, processing speed and energy efficiency in the screening of molecular databases. In this work, we present a classification model describing each molecule with a single energy-based vector and propose a machine-learning system based on the use of ANNs. Different ANNs are studied with respect to their suitability to identify biochemical similarities. Also, a high-performance and energy-efficient hardware acceleration platform based on the use of stochastic computing is proposed for the ANN implementation. This platform is of utility when screening vast libraries of compounds. As a result, the proposed model showed appreciable improvements with respect previously published works in terms of the main relevant characteristics (accuracy, speed and energy-efficiency)., Comment: 14 pages, 9 Figures, 3 Tables. Paper submitted to an IEEE journal

While deep learning models become more widespread, their ability to handle unseen data and generalize for any scenario is yet to be challenged. In medical imaging, there is a high heterogeneity of distributions among images based on the equipment that generates them and their parametrization. This heterogeneity triggers a common issue in machine learning called domain shift, which represents the difference between the training data distribution and the distribution of where a model is employed. A high domain shift tends to implicate in a poor generalization performance from the models. In this work, we evaluate the extent of domain shift on four of the largest datasets of chest radiographs. We show how training and testing with different datasets (e.g., training in ChestX-ray14 and testing in CheXpert) drastically affects model performance, posing a big question over the reliability of deep learning models trained on public datasets. We also show that models trained on CheXpert and MIMIC-CXR generalize better to other datasets., Comment: 10 pages, 3 figures

Recent advances in deep learning methods have come to define the state-of-the-art for many medical imaging applications, surpassing even human judgment in several tasks. Those models, however, when trained to reduce the empirical risk on a single domain, fail to generalize when applied to other domains, a very common scenario in medical imaging due to the variability of images and anatomical structures, even across the same imaging modality. In this work, we extend the method of unsupervised domain adaptation using self-ensembling for the semantic segmentation task and explore multiple facets of the method on a small and realistic publicly-available magnetic resonance (MRI) dataset. Through an extensive evaluation, we show that self-ensembling can indeed improve the generalization of the models even when using a small amount of unlabelled data., Comment: 15 pages, 9 figures

Predicting the response of a specific cancer to a therapy is a major goal in modern oncology that should ultimately lead to a personalised treatment. High-throughput screenings of potentially active compounds against a panel of genomically heterogeneous cancer cell lines have unveiled multiple relationships between genomic alterations and drug responses. Various computational approaches have been proposed to predict sensitivity based on genomic features, while others have used the chemical properties of the drugs to ascertain their effect. In an effort to integrate these complementary approaches, we developed machine learning models to predict the response of cancer cell lines to drug treatment, quantified through IC50 values, based on both the genomic features of the cell lines and the chemical properties of the considered drugs. Models predicted IC50 values in a 8-fold cross-validation and an independent blind test with coefficient of determination R2 of 0.72 and 0.64 respectively. Furthermore, models were able to predict with comparable accuracy (R2 of 0.61) IC50s of cell lines from a tissue not used in the training stage. Our in silico models can be used to optimise the experimental design of drug-cell screenings by estimating a large proportion of missing IC50 values rather than experimentally measure them. The implications of our results go beyond virtual drug screening design: potentially thousands of drugs could be probed in silico to systematically test their potential efficacy as anti-tumour agents based on their structure, thus providing a computational framework to identify new drug repositioning opportunities as well as ultimately be useful for personalized medicine by linking the genomic traits of patients to drug sensitivity., Comment: 26 pages, 7 figures, including supplemental information, presented by Michael Menden at the 5th annual RECOMB Conference on Regulatory and Systems Genomics with DREAM Challenges; accepted in PLOS ONE