Search

Learning compact but highly accurate models that help in human decision-making is challenging. Most such scoring systems were constructed by human experts using some heuristics. In this contribution, we propose a principled method with theoretical guarantees to learn interpretable simple rules. We introduce Vanishing Boosted Weights (VBW) approach which is a corrective fine-tuning procedure on decision stumps. It is a simple method which surprisingly was never investigated. We propose its extension, Corrective Federated Averaging VBW, that is practical in a federated learning scenario. We illustrate by our numerical experiments both on simulated and real data that the novel approaches are competitive compared to the state-of-the-art methods, and outperform them.

Cascade learning with abstention and individualised feature selection is a class of models in high demand in personalised medical applications. The cascade consists of sequential classifiers and rejectors, where the classifiers estimate confidence of prediction, and the rejectors evaluate an expected cost-to-go of features not selected yet. The number of models is exponential in the number of features and, therefore, the challenge is to find efficient heuristics for the NP-hard problem. The state-of-the-art is based on complex deep neural networks. We introduce an efficient and robust approach based on a probabilistic graphical model representing a unified probabilistic classifier that can be applied at any stage of a multi-stage sequential model. As for the rejector, we build it on the probabilistic-based neural network that incorporates the very same probabilistic model to treat unobserved feature values. We illustrate the efficiency of the proposed method on several data sets, and compare our results to the state-of-the-art.

Background Current clinical routines rely more and more on “omics” data such as flow cytometry data from host and microbiota. Cohorts variability in addition to patients’ heterogeneity and huge dimensions make it difficult to understand underlying structure of the data and decipher pathologies. Patients stratification and diagnostics from such complex data are extremely challenging. There is an acute need to develop novel statistical machine learning methods that are robust with respect to the data heterogeneity, efficient from the computational viewpoint, and can be understood by human experts. Results We propose a novel approach to stratify cell-based observations within a single probabilistic framework, i.e., to extract meaningful phenotypes from both patients and cells data simultaneously. We define this problem as a double clustering problem that we tackle with the proposed approach. Our method is a practical extension of the Latent Dirichlet Allocation and is used for the Double Clustering task (LDA-DC). We first validate the method on artificial datasets, then we apply our method to two real problems of patients stratification based on cytometry and microbiota data. We observe that the LDA-DC returns clusters of patients and also clusters of cells related to patients’ conditions. We also construct a graphical representation of the results that can be easily understood by humans and are, therefore, of a big help for experts involved in pre-clinical research.

The general articulation of the issue and its relation to the important research and practice tasks. In Ukraine, the implementation of measures to improve energy security does not fully take into account the potential available at regional and local levels. Use of the method of hierarchy analysis, which is a component of classification of regions by the level of energy security, is an effective tool for narrowing the range of ambiguity of goals in the process of managerial decision-making by means of cluster analysis. The analysis of recent publications regarding the issues this article deals with; identification of parts of a common problem that has not been previously addressed. The research of different aspects of clustering in the energy sphere has been given considerable attention by such scholars as V. Rosen, I. Nedin, O. Sukhodolya, A. Prahovnik, O. Boyko-Boychuk, O. Vlasyuk, N. Karaeva. In domestic practice, cluster analysis is widely used in addition to the study of energy security problems. The purpose (objective) of this article. The purpose of the article is to use a hierarchical analysis method for clustering regions by the level of use of energy security potential at regional and local levels under the unification rule – full connection and the chosen proximity method – Euclidean distance. The keyresults and the background. The energy security level of each region is simulated with their further division into 5 groups (clusters) according to the level of use of the energy security potential, which is available at regional and local levels with a minimum composition of the main components of energy security. Conclusions and the prospects for further research. Five homogeneous classes identified with the level of use of energy security potential, which are regarded at regional and local levels as not satisfactory, close to satisfactory, moderately satisfactory, sufficiently satisfactory and satisfactory, should be the basis for further adjustment of sustainable development of the regions and improvement of the regional and national energy policy levels.

In a number of real life applications, scientists do not have access to temporal data, since budget for data acquisition is always limited. Here we challenge the problem of causal inference between groups of heterogeneous non-temporal observations obtained from multiple sources. We consider a family of probabilistic algorithms for causal inference based on an assumption that in case where X causes Y, P(X) and P(Y vertical bar X) are statistically independent. For a number of real world applications, deep learning methods were reported to achieve the most accurate empirical performance, what motivates us to use deep Boltzmann machines to approximate the marginal and conditional probabilities of heterogeneous observations as accurate as possible. We introduce a novel algorithm to infer causal relationships between blocks of variables. The proposed method was tested on a benchmark of multivariate cause-effect pairs. We show by our experiments that our method achieves the state-of-the-art empirical accuracy, and sometimes outperforms the state-of-the-art methods. An important part of our contribution is an application of the proposed algorithm to an original medical data set, where we explore relations between alimentary patters, human gut microbiome composition, and health status.

International audience; Adipose tissue (AT) fibrosis in obesity compromises adipocyte functions and responses to intervention-induced weight loss. It is driven by AT progenitors with dual fibro/adipogenic potential, but pro-fibrogenic pathways activated in obesity remain to be deciphered. To investigate the role of macroautophagy/autophagy in AT fibrogenesis, we used Pdgfra-CreErt2 transgenic mice to create conditional deletion of Atg7 alleles in AT progenitor cells (atg7 cKO) and examined sex-dependent, depot-specific AT remodeling in high-fat diet (HFD)-fed mice. Mice with atg7 cKO had markedly decreased extracellular matrix (ECM) gene expression in visceral, subcutaneous, and epicardial adipose depots compared to Atg7lox/lox littermates. ECM gene program regulation by autophagy inhibition occurred independently of changes in the mass of fat tissues or adipocyte numbers of specific depots, and cultured preadipocytes treated with pharmacological or siRNA-mediated autophagy disruptors could mimic these effects. We found that autophagy inhibition promotes global cell-autonomous remodeling of the paracrine TGF-BMP family landscape, whereas ECM gene modulation was independent of the autophagic regulation of GTF2IRD1. The progenitor-specific mouse model of ATG7 inhibition confirms the requirement of autophagy for white/beige adipocyte turnover, and combined to in vitro experiments, reveal progenitor autophagy dependence for AT fibrogenic response to HFD, through the paracrine remodeling of TGF-BMP factors balance. Abbreviations: CQ: chloroquine; ECM: extracellular matrix; EpiAT: epididymal adipose tissue; GTF2IRD1: general transcription factor II I repeat domain-containing 1; HFD: high-fat diet; KO: knockout; OvAT: ovarian adipose tissue; PDGFR: platelet derived growth factor receptor; ScAT: subcutaneous adipose tissue; TGF-BMP: transforming growth factor-bone morphogenic protein.

Machine learning (ML) methods are becoming the state-of-the-art in numerous domains, including material sciences. In this manuscript, we demonstrate how ML can be used to efficiently predict several properties in solid-state chemistry applications, in particular, to estimate the heat of formation of a given complex crystallographic phase (here, the σ − phase, t P 30 , D 8 b ). Based on an independent and unprecedented large first principles dataset containing about 10,000 σ − compounds with n = 14 different elements, we used a supervised learning approach to predict all the ∼ 500,000 possible configurations. From a random set of ∼ 1000 samples, predictions are given within a mean absolute error of 23 meV at−1 ( ∼ 2 kJ mol−1) on the heat of formation and ∼ 0.06 A on the tetragonal cell parameters. We show that deep neural network regression results in a significant improvement in the accuracy of the predicted output compared to traditional regression techniques. We also integrated descriptors having physical nature (atomic radius, number of valence electrons), and we observe that they improve the model precision. We conclude from our numerical experiments that the learning database composed of the binary-compositions only, plays a major role in predicting the higher degree system configurations. Our results open a broad avenue to efficient high-throughput investigations of the combinatorial binary computations for multicomponent complex intermetallic phase prediction.

Context Bariatric surgery (BS) induces major and sustainable weight loss in many patients. Factors predicting poor weight-loss response (PR) need to be identified to improve patient care. Quantification of subcutaneous adipose tissue (scAT) fibrosis is negatively associated with post-BS weight loss, but whether it could constitute a predictor applicable in clinical routine remains to be demonstrated. Objective To create a semiquantitative score evaluating scAT fibrosis and test its predictive value on weight-loss response after Roux-en-Y gastric bypass (RYGB). Methods We created a fibrosis score of adipose tissue (FAT score) integrating perilobular and pericellular fibrosis. Using this score, we characterized 183 perioperative scAT biopsy specimens from severely obese patients who underwent RYGB (n = 85 from a training cohort; n = 98 from a confirmation cohort). PR to RYGB was defined as

Information from metagenomic data from human microbiome may improve diagnosis and prognosis for multiple human diseases. However, to achieve a prediction based on bacterial abundance information remains a challenge. Indeed, the number of features being much higher than the number of samples, we face difficulties related to high dimensional data processing, as well as overfitting. In this study, we investigate several convolutional neural network architectures for synthetic images and some experimental techniques to generate and train these synthetic images. We also explore supervised learning for visualizing high dimensional data that use data on genus, species and higher taxonomic level information. In addition, some dimensionality reduction approaches are examined on very high dimensional data such as gene families abundance. We evaluated our approach on six different metagenomic datasets including five types of diseases with more than 1000 samples. Our method displays promising results and can be used in different omics data settings, including integrative ones.

International audience; This paper addresses the problem of dimension reduction of noisy data, more precisely the challenge to determine the dimension of the subspace where the observed signal lives in. Based on results from random matrix theory, two novel estimators of the signal dimension are proposed in this paper. Consistency of the estimators is proved in the modern asymptotic regime, where the number of parameters grows proportionally with the sample size. Experimental results show that the novel estimators are robust to noise and, moreover, they give highly accurate results in settings where standard methods fail. We apply the novel dimension estimators to several life sciences benchmarks in the context of classification, and illustrate the improvements achieved by the new methods compared to the state-of-the-art approaches.

An important question in microbiology is whether treatment causes changes in gut flora, and whether it also affects metabolism. The reconstruction of causal relations purely from non-temporal observational data is challenging. We address the problem of causal inference in a bivariate case, where the joint distribution of two variables is observed. We consider, in particular, data on discrete domains. The state-of-the-art causal inference methods for continuous data suffer from high computational complexity. Some modern approaches are not suitable for categorical data, and others need to estimate and fix multiple hyper-parameters. In this contribution, we introduce a novel method of causal inference which is based on the widely used assumption that if $X$ X causes $Y$ Y , then $P(X)$ P ( X ) and $P(Y|X)$ P ( Y | X ) are independent. We propose to explore a semi-supervised approach where $P(Y|X)$ P ( Y | X ) and $P(X)$ P ( X ) are estimated from labeled and unlabeled data respectively, whereas the marginal probability is estimated potentially from much more (cheap unlabeled) data than the conditional distribution. We validate the proposed method on the standard cause-effect pairs. We illustrate by experiments on several benchmarks of biological network reconstruction that the proposed approach is very competitive in terms of computational time and accuracy compared to the state-of-the-art methods. Finally, we apply the proposed method to an original medical task where we study whether drugs confound human metagenome.

OBJECTIVE Roux-en-Y gastric bypass (RYGB) induces type 2 diabetes remission (DR) in 60% of patients at 1 year, yet long-term relapse occurs in half of these patients. Scoring methods to predict DR outcomes 1 year after surgery that include only baseline parameters cannot accurately predict 5-year DR (5y-DR). We aimed to develop a new score to better predict 5y-DR. RESEARCH DESIGN AND METHODS We retrospectively included 175 RYGB patients with type 2 diabetes with 5-year follow-up. Using machine learning algorithms, we developed a scoring method, 5-year Advanced-Diabetes Remission (5y-Ad-DiaRem), predicting longer-term DR postsurgery by integrating medical history, bioclinical data, and antidiabetic treatments. The scoring method was based on odds ratios and variables significantly different between groups. This score was further validated in three independent RYGB cohorts from three European countries. RESULTS Compared with 5y-DR patients, patients who had relapsed after 5 years exhibited more severe type 2 diabetes at baseline, lost significantly less weight during the 1st year after RYGB, and regained more weight afterward. The 5y-Ad-DiaRem includes baseline (diabetes duration, number of antidiabetic treatments, and HbA1c) and 1-year follow-up parameters (glycemia, number of antidiabetic treatments, remission status, 1st-year weight loss). The 5y-Ad-DiaRem was accurate (area under the receiver operating characteristic curve [AUROC], 90%; accuracy, 85%) at predicting 5y-DR, performed better than the Diabetes Remission score (DiaRem) and the Advanced-DiaRem (AUROC, 81% and 84%; accuracy, 79% and 78%, respectively), and correctly reclassified 13 of 39 patients misclassified with the DiaRem. The 5y-Ad-DiaRem robustness was confirmed in the independent cohorts. CONCLUSIONS The 5y-Ad-DiaRem accurately predicts 5y-DR and appears relevant to identify patients at risk for relapse. Using this score could help personalize patient care after the 1st year post-RYGB to maximize weight loss, limit weight regains, and prevent relapse.

An important question in microbiology is whether treatment causes changes in gut flora, and whether it also affects metabolism. The reconstruction of causal relations purely from non-temporal observational data is challenging. We address the problem of causal inference in a bivariate case, where the joint distribution of two variables is observed. The state-of-the-art causality inference methods for continuous data suffer from high computational complexity. Some modern approaches are not suitable for categorical data, and others need to estimate and fix multiple hyper-parameters.

Score learning aims at taking advantage of supervised learning to estimate interpretable models which facilitate decision making. Ideally, a scoring system is based on simple arithmetic operations, is sparse, and can be easily explained by human experts. In this contribution, we introduce an original methodology to simultaneously learn interpretable binning mapped to a class variable, and the weights associated with these bins contributing to the score. We show by numerical experiments on benchmark data sets that our approach is competitive compared to the state-of-the-art methods. We illustrate by a real medical problem of type 2 diabetes remission prediction that a scoring system learned automatically is comparable to one manually constructed by clinicians.

Not all people with type 2 diabetes who undergo bariatric surgery achieve diabetes remission. Thus it is critical to develop methods for predicting outcomes that are applicable for clinical practice. The DiaRem score is relevant for predicting diabetes remission post-Roux-en-Y gastric bypass (RYGB), but it is not accurate for all individuals across the entire spectrum of scores. We aimed to develop an improved scoring system for predicting diabetes remission following RYGB (the Advanced-DiaRem [Ad-DiaRem]). We used a retrospective French cohort (n = 1866) that included 352 individuals with type 2 diabetes followed for 1 year post-RYGB. We developed the Ad-DiaRem in a test cohort (n = 213) and examined its accuracy in independent cohorts from France (n = 134) and Israel (n = 99). Adding two clinical variables (diabetes duration and number of glucose-lowering agents) to the original DiaRem and modifying the penalties for each category led to improved predictive performance for Ad-DiaRem. Ad-DiaRem displayed improved area under the receiver operating characteristic curve and predictive accuracy compared with DiaRem (0.911 vs 0.856 and 0.841 vs 0.789, respectively; p = 0.03); thus correcting classification for 8% of those initially misclassified with DiaRem. With Ad-DiaRem, there were also fewer misclassifications of individuals with mid-range scores. This improved predictive performance was confirmed in independent cohorts. We propose the Ad-DiaRem, which includes two additional clinical variables, as an optimised tool with improved accuracy to predict diabetes remission 1 year post-RYGB. This tool might be helpful for personalised management of individuals with diabetes when considering bariatric surgery in routine care, ultimately contributing to precision medicine.

Background The last decades witnessed an explosion of large-scale biological datasets whose analyses require the continuous development of innovative algorithms. Many of these high-dimensional datasets are related to large biological networks with few or no experimentally proven interactions. A striking example lies in the recent gut bacterial studies that provided researchers with a plethora of information sources. Despite a deeper knowledge of microbiome composition, inferring bacterial interactions remains a critical step that encounters significant issues, due in particular to high-dimensional settings, unknown gut bacterial taxa and unavoidable noise in sparse datasets. Such data type make any a priori choice of a learning method particularly difficult and urge the need for the development of new scalable approaches. Results We propose a consensus method based on spectral decomposition, named Spectral Consensus Strategy, to reconstruct large networks from high-dimensional datasets. This novel unsupervised approach can be applied to a broad range of biological networks and the associated spectral framework provides scalability to diverse reconstruction methods. The results obtained on benchmark datasets demonstrate the interest of our approach for high-dimensional cases. As a suitable example, we considered the human gut microbiome co-presence network. For this application, our method successfully retrieves biologically relevant relationships and gives new insights into the topology of this complex ecosystem. Conclusions The Spectral Consensus Strategy improves prediction precision and allows scalability of various reconstruction methods to large networks. The integration of multiple reconstruction algorithms turns our approach into a robust learning method. All together, this strategy increases the confidence of predicted interactions from high-dimensional datasets without demanding computations. Electronic supplementary material The online version of this article (doi:10.1186/s12859-016-1308-y) contains supplementary material, which is available to authorized users.

International audience; —Feature selection is used to preserve significant properties of data in a compact space. In particular, feature selection is needed in applications, where information comes from multiple heterogeneous high dimensional sources. Data integration, however, is a challenge in itself. In our contribution, we introduce a feature selection framework based on powerful visualisation capabilities of self-organising maps, where the deep structure can be learned in a supervised or unsupervised manner. For a supervised version of the deep SOM, we propose to carry out inference with a linear SVM. A forward-backward procedure helps to converge to an optimal feature set. We show by experiments on real large-scale biomedical data set that the proposed methods embed data in a new compact meaningful representation, allow to visualise biomedical signatures, and also lead to a reasonable classification accuracy compared to the state-of-the-art methods.

Prior knowledge integration aims at taking advantage of cheap plentiful data to improve the efficiency of supervised learning procedures. For parametric models, however, it is a challenging task. In this contribution, we introduce a novel simple methodology to incorporate prior knowledge into a supervised objective function. We propose to introduce the prior knowledge in the form of joint probability of observations and labels. We discuss the nature of features in discriminative and generative models and hence, differences in priors integration. We illustrate the efficiency of the proposed method both by synthetic data sets and by our results on a realistic large-scale sequence labeling task.