ABSTRACT. In this talk I will introduce the Restricted Boltzmann Machines (RBM): a simple Machine learning model with a bipartite graph architecture, learned only on sequence data. I will focus on the applications of RBM on the predictions of SARS-CoV-2 evolution. By integrating pre-pandemic evolutionary constraints gathered from SARS-CoV-2 far homologous sequences, with large-scale Deep mutational Scans (DMS) data we model how viral fitness, ACE2 binding, and immune escape pressures jointly sculpt the mutational landscape. In contrast to structure-based models that are restricted in scalability, our sequence-based energy framework enables broad exploration of evolutionary trajectories while remaining valuable for experimental validation. Experimental validation of model predictions includes the test of 22 synthetic RBD variants with up to 21 mutations from the wild-type. Half of these variants maintained expression and ACE2 binding, and some successfully escaped most of the 9 antibodies tested.

Bibliography

[1] Minimal epistatic networks from integrated sequence and mutational protein data. S. Cocco, L. Posani, R. Monasson. PNAS doi: 10.1073/pnas.2312335121 (2024) Supplementary Material

[2] Learning the differences: a transfer-learning approach to predict antigen immunogenicity and T-cell receptor specificity.B. Bravi, A. Di Gioacchino, J. Fernandez-de-Cossio-Diaz, A. M. Walczak, T. Mora, S. Cocco, and R. Monasson. eLife 12:e85126 (2023)

[3] Machine learning for evolutionary-based and physics-inspired protein design: Current and future synergies. C. Malbranke, D. Bikard, S. Cocco, R. Monasson, J. Tubiana. Current Opinion in Structural Biology 80:102571 (2023)

[4]  Mutational paths in protein-sequence landscapes: from sampling to mean-field characterization. E. Mauri, S. Cocco, R. Monasson. Physical Review Letters 130, 158402 (2023) 

[5] An evolution-based model for designing chorismate mutase enzymes . W.P. Russ, M. Figliuzzi, C. Stocker, P. Barrat-Charlaix, M. Socolich, P. Kast, D. Hilvert, R. Monasson, S. Cocco, M. Weigt, R. Ranganathan. Science 369, 440-5 (2020) 

[6] Learning protein constitutive motifs from sequence data . J. Tubiana, S. Cocco, R. Monasson. eLife 2019;8:e39397 (2019). See also the press release.

ABSTRACT. While of primary importance in both the biomedical and therapeutic fields, peptides suffer from a relative lack of dedicated tools to predict efficiently and accurately their 3D structures despite a crucial step in understanding their physiopathological function or designing new drugs. In recent years, deep-learning methods have enabled a major breakthrough for the protein 3D structures prediction approaches, allowing to predict protein 3D structures with a near-experimental accuracy for nearly any protein sequence. This present study aims at confronting some of these new methods (AlphaFold2, RoseTTAFold2 and ESMFold) for the peptides 3D structure prediction problem, and to evaluate their performance. All methods produced high quality results, but their overall performance is lower as compared the prediction of proteins 3D structure. We also identified a few structural features that impede the ability to produce high-quality peptide structure predictions. These findings point out the discrepancy that still exists between the protein and peptide 3D structure prediction methods, and underline a few cases where the generated peptide structures should be used very cautiously.

ABSTRACT. In RNA tertiary structures, non canonical interactions form dense interaction networks called structural motifs, which are recurrent in tertiary structures, and which highly contribute to the tridimensional shape of the molecules. Being able to detect these motifs in RNAs represented as graphs without any geometric information constitutes a step towards tertiary structure prediction. This is a difficult problem due to the variability of structural motifs in a same family. This paper presents a simple way to formally describe variable structural motifs, and algorithms for searching for instances of these motifs in RNA structures abstracted as graphs. Experimental results show that this approach is very good in terms of specificity, and fairly good in terms of sensitivity, thus taking us a step towards predicting such motifs in RNA graphs. A user friendly web server and a freely downloadable python resource allow to search for given variable motifs in RNA graphs.

ABSTRACT. A growing body of evidence shows that RNA function depends not only on primary and secondarystructures but also on its 3D conformation. As the experimental determinations are costly anduncertain processes, computational prediction methods are essential. A critical task in suchprediction is identifying substructures that can be modeled independently before assembling theglobal fold. In proteins, these are “structural domains” - yet no equivalent concept exists for RNA.

In this work, we present RNA3DClust, an adaptation of the Mean Shift algorithm for partitioning RNA3D structures into compact, distinct regions, akin to protein domains. To evaluate the method, webuilt a reference dataset of annotated RNA 3D domains and developed a custom scoring scheme. Wealso show that RNA3DClust’s segmentations align with biologically and evolutionarily defineddomains. Finally, with the emerging interest in long non-coding RNAs (lncRNAs), which likely containfolded substructures, we created a second dataset using predicted lncRNA models. RNA3DClust’sresults on these models further demonstrate its potential for RNA domain analysis.

ABSTRACT. Climate change is intensifying summer droughts in Europe, significantly affecting plant growth and crop yields.The flowering of maize happens during this high-risk period, and the water deficit stress results in kernel abortion among other consequences, seriously impacting yield. Understanding the genomic basis of maize responses to water deficit is therefore crucial for agricultural adaptation. Plants regulate gene expression in response to environmental changes through signaling pathways, where transcription factors (TFs) interact with cis-regulatory elements (CREs). Two main categories of CREs exist: (i) proximal CREs, or promoters, and (ii) distal CREs (dCREs), which include enhancers and silencers. dCREs can regulate multiple genes across various cell types and influence gene expression in a complex, context-dependent manner. While promoters have been extensively studied, dCREs remain underexplored in maize due to genome assembly challenges and the absence of specific epigenetic markers. However, recent technological advances, including long-read sequencing and the discovery that most maize dCREs are unmethylated, facilitate their identification and functional characterization.

Traditional gene regulatory network (GRN) analyses rely on co-expression networks, which assume that TF effects are directly correlated with their expression levels. To address this limitation, we propose a co-regulation approach that considers the physical interaction potential of TFs with CREs and the correlation between the expression of target genes regulated by the same TF.

Our study focuses on identifying dCREs involved in maize responses to water deficit in the reference inbred line B73. We constructed an initial regulatory network using methylation data and genome annotation, which was refined using message-passing-based inference to generate tissue- and condition-specific networks. Comparative analysis revealed differential gene regulation across tissues, particularly in leaves, where genes associated with photosynthesis and stress responses were significantly affected. These findings highlight the utility of CRE-based networks for identifying key regulatory elements in maize drought responses.

ABSTRACT. Acute Promyelocytic Leukaemia (APL) arises from an aberrant chromosomal translocation involving the Retinoic Acid Receptor Alpha (RARA) gene, predominantly with the Promyelocytic Leukaemia (PML) or Promyelocytic Leukaemia Zinc Finger (PLZF) genes. The resulting oncoproteins block the haematopoietic differentiation program promoting aberrant proliferative promyelocytes. Retinoic Acid (RA) therapy is successful in most of the PML::RARA patients, while PLZF::RARA patients frequently become resistant and relapse. Recent studies pointed to various underlying molecular components, but their precise contributions remain to be deciphered. We developed a logical network model integrating signalling, transcriptional and epigenetic regulatory mechanisms, which captures key features of the APL cell responses to RA depending on the genetic background. The explicit inclusion of the histone methyltransferase EZH2 allowed the assessment of its role in the resistance mechanism, distinguishing between its canonical and non-canonical activities. The model dynamics was thoroughly analysed using tools integrated in the public software suite maintained by the CoLoMoTo consortium (https://colomoto.github.io/). The model serves as a solid basis to assess the roles of novel regulatory mechanisms, as well as to explore novel therapeutical approaches in silico.

ABSTRACT. Rheumatoid arthritis is a complex disease marked by joint pain, stiffness, swelling, and chronic synovitis, arising from the dysregulated interaction between synoviocytes and immune cells. Its unclear etiology makes finding a cure challenging. The concept of digital twins, used in engineering, can be applied to healthcare to improve diagnosis and treatment for complex diseases like rheumatoid arthritis. In this work, we pave the path towards a digital twin of the arthritic joint by building a large, modular biochemical reaction map of intra- and intercellular interactions. This network, featuring over 1000 biomolecules, is then converted to one of the largest executable Boolean models for biological systems to date. Validated through existing knowledge and gene expression data, our model is used to explore current treatments and identify new therapeutic targets for rheumatoid arthritis.

ABSTRACT. In 2025, the Hub of Bioinformatics and Biostatistics of the Institut Pasteur (the Hub) celebrates its 10th anniversary. Created to centralize bioinformaticians and biostatisticians scattered across campus, the Hub aims to pool resources and promote collaboration. It has become a key group supporting computational biology research through expert guidance, training, tool development, and building a community of specialists. Over the last decade, the Hub has collaborated with 200 research units on over 800 projects, combining its diverse expertise. This work resulted in more than 500 peer-reviewed publications, nearly 50 computational tools, and 30 deployed websites and portals. Additionally, the Hub established a dedicated PhD program and provided over 3,000 trainings. It tailored courses and mentoring programs for the Pasteur network and became integrated into the French bioinformatics landscape, co-organising events like the IFB Bioinformatics School and workshops on single-cell data analysis. The hub adopts a hub-and-spoke model with over 40 core members divides into five groups. Another 30 engineers are assigned to research units and transmissible disease surveillance labs for five-years renewable detachments, contributing 20% of their time to the Hub activities. Systems like weekly open-desks, project management tools, and steering committees ensure effective collaboration and address diverse scientific needs. The Hub has fulfilled its initial goals, supporting research and enabling knowledge exchange among computational biologists. It reduces isolated workloads and offers flexible career opportunities for research engineers. Its first decade highlights valuable lessons in governance, operational models, methodology development and outreach. The Hub looks forward for addressing challenges like innovation, career growth and time management as computational biology continues to evolve rapidly, while preparing for impactful next ten years.

ABSTRACT. The importance of protein amyloidogenesis, associated with various diseases and functional roles, has driven the creation of computational predictors of amyloidogenicity. The accuracy of these predictors, particularly those utilizing artificial intelligence technologies, heavily depends on the quality of the data. We built Cross-Beta DB, a database containing high-quality data on known cross-β amyloids formed under natural conditions. We used it to train and benchmark several machine-learning (ML) algorithms to predict amyloid-forming potential of proteins. We developed the Cross-Beta predictor using an Extra trees ML algorithm, which outperforms other amyloid predictors with the highest F1 score (0.852) and accuracy (0.844) compared to existing methods. The development of the Cross-Beta DB database and a new ML-based Cross-Beta predictor may enable the creation of personalized risk profiles for neurodegenerative diseases and other amyloidoses—especially as genome sequencing becomes more affordable.

ABSTRACT. Background

Childhood trauma, including abuse or neglect, has profound effects on mental health, increasing susceptibility to psychiatric disorders. Bipolar disorder, marked by extreme mood swings encompassing manic and depressive episodes, disrupts daily functioning. Despite the growing interest in molecular psychiatry, the etiology of bipolar disorder remains unclear, with no established blood biomarkers [1]. This gap of knowledge is partially due to the complexity and heterogeneity of the disorder. Additionally, environmental factors, particularly early-life trauma, are suspected to play a significant role in the onset and progression of bipolar disorder [2,3,4].

Recent advances in Next-Generation Sequencing (NGS) have generated extensive genomic data, yet the integration of multi-omic data with advanced machine learning techniques remains underutilized in psychiatric research[5]. As seen in cancer research, the application of multi-omics approaches that combine genetic, transcriptomic, and epigenomic data with machine learning holds potential for advancing our understanding of psychiatric disorders.

Material and Methods

This study aims to determine the minimum sample size required to accurately predict trauma exposure and identify potential biomarkers of childhood trauma in peripheral blood samples from bipolar patients. We utilized transcriptomics (RNA-seq), and epigenomics (miRNA-seq and DNA methylation) datasets from a cohort of bipolar disorder (n = 274) patients, all of whom were assessed using the Childhood Trauma Questionnaire (CTQ). After quality control and preprocessing the final dataset included 200 individuals with DNAm data, 122 individuals with mRNA and miRNA data, and 102 individuals with data from all three omics modalities.

We derived train/test subsets by gradually increasing the sample size in the training set. Using an advanced joint reduction dimension method, named Regularized Generalized Canonical Correlation Analysis (RGCCA) [6], we evaluated the prediction error rates as a function of sample size in the training set.

Results and Discussion

The analysis revealed that N80% = 81 individuals in the training set (i. e. 80% train-test split), for at least 2 modalities over 3 and from 3 components per block, achieved the best prediction performances. However, almost no feature survived the multiple testing procedure when assessing model stability, suggesting that further investigations are needed to obtain a biologically interpretable sparse model.

ABSTRACT. NORDic is an open-source package designed for network-based transcriptional analysis for drug repurposing, (1) enabling the inference of disease-specific gene regulatory networks, (2) identifying key genes in the regulation (master regulators), (3) simulating in silico the drug effect in pathological profiles, and (4) screening automatically drugs based on those simulations. This demo applies two functionalities of NORDic to proteome and transcriptome datasets from resilient and post-traumatic stress disorder (PTSD)- diagnosed individuals. Indeed, PTSD is associated with widespread protein and gene dysregulations in blood and across multiple brain regions [1,2]. However, traditional differential expression analyses often fail to capture the complex interactions between groups of genes governing these molecular changes. Here, the aim is to model a gene regulatory network predictive of PTSD risk and to identify master regulators of resilience to PTSD. We start this demo with the Network Identification (NI) function, that builds a dynamic regulatory model (a boolean network) by integrating input and publicly available biological sources in an automated fashion. Then, we apply the Prioritization of Master Regulators (PMR) function. This function identifies disease-specific master regulators based on the topology and dynamics of the network, redefining gene centrality as the primary target to drive pathological transcriptional profiles toward resilient ones. Future extensions will explore the integration of other omic layers (e.g., microRNAs, DNA methylation) to refine disease models further.

ABSTRACT. Information integration into Semantic Knowledge Graphs (SKG) is gaining traction as a way to integrate large sets of biomedical knowledge bases and databases. However, building up such het- erogeneous SKG requires tools allowing both ease of use and FAIRness. Recently, we introduced a set of tools allowing a fully automated and FAIR creation of SKGs, from a configuration that is easy to read and write by an end-user. This demonstration will show how to use the OntoWeaver to produce a SKG integrating clinical data from patients with high-grade serous ovarian cancer, and including informa- tion on genome changes collected as part of the DECIDER project. The built SKG can then be queried to gather “evidence paths” linking patient-specific alterations to actionable drugs, enabling the discovery of optimal personalized treatment options, together with the supporting literature knowledge and data. While our demonstration focuses on high-grade serous ovarian cancer, OntoWeaver’s modular design and configuration-driven approach can be readily extended to other malignancies or disease contexts, offering a generalizable solution for large-scale biomedical knowledge integration.

ABSTRACT. Metagenomic sequencing provides profound insights into microbial communities, but it is often compromised by technical biases, including cross-sample contamination. This phenomenon arises when microbial content is inadvertently exchanged among concurrently processed samples, distorting microbial profiles and compromising the reliability of metagenomic data and downstream analyses. Existing detection methods often rely on negative controls, which are inconvenient and do not detect contamination within real samples. Meanwhile, strain-level bioinformatics approaches fail to distinguish contamination from natural strain sharing and lack sensitivity. To fill this gap, we introduce CroCoDeEL, a decision-support tool for detecting and quantifying cross-sample contamination. Leveraging linear modeling and a pre-trained supervised model, CroCoDeEL identifies specific contamination patterns in species abundance profiles. It requires no negative controls or prior knowledge of sample processing positions, offering improved accuracy and versatility. Benchmarks across three public datasets demonstrate that CroCoDeEL accurately detects contaminated samples and identifies their contamination sources, even at low rates (<0.1%), provided sufficient sequencing depth. Notably, we discovered critical contamination cases in highly cited studies, calling some of their results into question. Our findings suggest that cross-sample contamination is a widespread yet underexplored issue in metagenomics and emphasize the necessity of systematically integrating contamination detection into sequencing quality control. Future work will consist in developping an innovative approach to remove the contamination signal detected by CroCoDeEL.

ABSTRACT. With the massive sequencing of genomes, transcriptomes and environmental samples, the quantity of sequences is increasing exponentially, forcing us to develop new, faster and more efficient methods. We propose LAGOON-MCL, a Nextflow pipeline for annotating and comparing protein sequences. First, sequences are annotated using Pfam [1] and user-supplied data (e.g. taxonomy) and compared to AlphaFold Protein Structure Database (AlphaFold DB) [2] to obtain structural information. Next, a graph clustering algorithm is applied to a sequence similarity network built from a pairwise alignment, enabling sequences to be compared and putative protein families to be constructed labeled with the annotations.

The pipeline was tested on 101 dinoflagellate transcriptomes (7 million sequences). 22% of sequences have a similarity in Pfam and 69% in AlphaFold DB. They are grouped into 365,006 clusters, 17% contain at least one sequence annotated with Pfam, and AlphaFold DB enables 19% more families to be annotated; 68% of sequences are in annotated families and 67% of unannotated families have less than 3 sequences. The pipeline is adaptable to other organisms and data, offering a versatile solution for sequence annotation and comparison.

ABSTRACT. Selection of test instances for string algorithms

The consumption of resources, such as electricity and materials used in the manufacture of computers, is increasing sharply with the development of the digital economy and human activities. The question is how to limit this consumption without reducing activities that may be socially useful or necessary.

In bioinformatics, we develop a lot of software, and sometimes many software for the same task, the same computational question. Think, for instance, about genome assembly or read mapping.

When it comes to software development, it's advisable to develop the most comprehensive series of tests to ensure the validity of a piece of software. In the development cycle, we iterate, frequently or automatically, the execution of tests in order to verify the software's correctness at each step. Tests may also be run to evaluate computational speed. This raises scientific questions about the usefulness or redundancy of certain test instances.

Let's take the case of a program that searches for a word in a text, such as your favorite genome. To test the program in all situations, we can run tests on all words of length k, for k equals 2, then 3, then 4, ..., up to, say 31 (footnote{The value k=31 is common for some DNA sequence analyses.}). But clearly the number of instances increases exponentially with length k, and quickly becomes prohibitive. From the point of view of computer science, we can rephrase our initial question as follows: 

Can we identify redundant test instances? Can we generate only useful test instances?

I will address these questions during this talk, illustrating them with two complex sequence processing algorithms and showing the impact of instance selection. 

This principle of test organization is called Equivalence Class Partitioning ([[https://en.wikipedia.org/wiki/Equivalence_partitioning][ECP]]) in the domain of software development. The study of these equivalence classes for a given algorithm can be complex.

Given the frequency with which tests are run during software development, and the number of software in bioinformatics, this type of approach can help reducing the ecological impact of our developments.

In addition to the advantage in terms of resource usage, the partitioning approach can also inform us about the average program execution time on a class of instances of a given size.  This generic approach points to avenues of research for many core algorithms, avenues that may foster interactions with other areas of computer science.