I will give an overview of my recent, ongoing and future work in machine learning for biological sequences. This will cover in particular: - Microbial GWAS: tools to identify genetic determinants of phenotypic traits such as antimicrobial resistances. The importance of accessory genes in microbes make the usual focus on SNPs inappropriate. I am developing solutions based on k-mers, i.e., the presence of short sequences in the genomes. - Prediction from biological sequences: neural networks that take a biological sequence as input and predict a property of this sequence. This applies e.g. to regulatory genomics or fold prediction. I am working on regularization and on statistical inference on the features extracted by these networks. - Machine learning for evolutionary genomics: neural networks to infer parameters of sequence evolution models. For some complex models, likelihood maximization is too hard but sampling is easy. I use a large amount of simulated data to learn…

Cancer is a heterogeneous disease, where multiple, phenotypically distinct subpopulations co-exist. Tumour evolution is a result of a complex interplay of genetic and epigenetic factors. Predicting the different cancer phenotypes requires linking the molecular state of a clone to its fate with high precision. Single-cell lineage tracing techniques are being successfully applied to the study of cancer evolution. I will show how single-cell assays endowed with genetic barcodes can be exploited to extract transcriptional stability properties of cancer clones, link molecular states to phenotypes, and study tumour heterogeneity, using a breast cancer model as a case study. Next, I will show how a new framework, called GALILEO, capturing lineage tracing, transcriptomic, and chromatin accessibility information simultaneously and at single-cell resolution, can be used to gain deep insights on aggressive cancer phenotypes. By tracing cancer cells in mice, we observe that the highly tumorigenic niche properly contains two transcriptionally distinct clonal subpopulations. Importantly,…

A central problem in computational structural biology is to design effective algorithms dealing with high dimensional conformationalspaces, also shedding light on the intrinsic complexity of the problems tackled (identifying meta-stable states, computing free energies, computing kinetic constants). In this talk, I will review recent work on Monte Carlo Markov chains (MCMC) algorithms of the Hit-and-Run type, for two seemingly unrelated problems: computing the volume of high dimensional polytopes, and sampling the conformational space of long flexible molecules.  I will also discuss the connexion with densities of states (DoS) calculations,  to reliably estimate thermodynamic properties of biomolecular systems with hundreds of degrees of freedom. [1] Geometric constraints within tripeptides and the existence of tripeptide reconstructionsT. O'Donnell, and F. Cazalshttps://www.biorxiv.org/content/10.1101/2022.06.21.497005v2 [2] Enhanced conformational exploration of protein loops using a global parameterization of the backbone geometryT. O'Donnell, and F. Cazalshttps://www.biorxiv.org/content/10.1101/2022.06.21.497022v2 [3] Efficient computation of the the volume of a polytope in high-dimensions using Piecewise Deterministic…

Paramecium is a unicellular organism that swims in fresh water using cilia. When it is stimulated (mechanically, chemically, optically, thermally, etc.), it often swims backward then turns and swims forward again: this is called the avoiding reaction. This reaction is triggered by a calcium-based action potential. For this reason, it enjoyed a period of glory in the 1970s as a model organism for neuroscience – some authors called it the "swimming neuron". Paramecium has considerable experimental advantages: it can be easily mass cultured in days, it is accessible to intracellular electrophysiology, its genome has been sequenced and it benefits from many techniques such as RNA interference. Because Paramecium is both a cell and an organism, it becomes possible to develop quantitative models that link physiology and behavior, cellular plasticity and behavioral plasticity (including learning), theoretical neuroscience and systems biology. I will describe our first efforts to develop such models, based on electrophysiological…

Due to the advent of high-throughput technologies, high-dimensional "omics" data are produced at an increasing pace. In cancer biology, national and international consortia have profiled thousands of tumors at multiple molecular levels ("multi-omics") allowing to gather a comprehensive molecular picture of this disease. Moreover, multi-omics profiling approaches are currently being transposed at single-cell resolution, further increasing the information accessible from cancer samples. The current main challenge is to design appropriate methods to integrate this wealth of information and translate it into actionable biological knowledge. In this talk, I will discuss two maincomputational directions, both grounded on multi-view learning: (i) multilayer networks to integrate a large range of interactions and (ii) dimensionality reduction to extract biological knowledge simultaneously from multiple omics. First, I will present the development done by my team for  bulk data and then our ongoing research in single-cell.

It is well-known that there are interactions everywhere in the living world. This seems very much like old scientific news. However, I will argue that, when biological evolution is investigated in ways that focus on microbes and interactions, some rather unexpected, evolutionary meaningful, interactions can still be uncovered. In this talk, I will introduce some recently published works from my lab and go through three, quite distinct cases of such surprising interactions. First, I will show how the use of similarity networks sheds light on a diversity of remodelled genes (symbiogenetic, chimeric, etc.), which arise from interactions between molecular components within genomes. Then, I will describe unexpected interactions between phages and bacteria, suggested by the prediction of novel quorum-sensing systems. Finally, I will provide original evidence that protein-protein interaction networks associated with ageing in humans are possibly impacted by 'age-distorters', suggesting that dozens of viruses can alter or manipulate human ageing. I…

.