Our long term goal is to build predictive theories for biology, similar to what is done in physics.
We are working in close collaboration with biologists, building models of complex biological systems, from the sub-cellular to the evolutionary scale.
We combine tools from multiple disciplines from physics, mathematics, computer science, and machine learning such as :
dynamical systems theory (bifurcation analysis, phase portraits, phase oscillators, …)
bioinformatics
dimensional reduction techniques, neural networks
evolutionary algorithms
landscape descriptions
As theorists, we do not hesitate to explore multiple directions and topics. Below we describe a few recent systems and themes that we have explored.
Some key words : biophysics, bioinformatics, systems biology, developmental biology, evolution, quantitative immunology, non linear dynamics
Immune Recognition
How does the immune system recognize pathogens ? How can it “encode” information to trigger an immune response ? Using a first-principle, theory based approach combined with experiments, we have uncovered a universal recognition process that we call Adaptive Kinetic Proofreading. An instance of this model explains both how T cells can perform sensitive and specific antigen recognition, and why they can be antagonized by sub-threshold antigens.
In collaboration with the Altan-Bonnet group at the US National Cancer Institute, we recently used a robotic platform combined to machine learning to reverse-engineer the dynamics of immune (cytokine) response, uncovering a universal antigen encoding, with latent structure similar to Adaptive Kinetic Proofreading
Evolving Biological Networks
In a famous thought experiment, Stephen Jay Gould once proposed that, if one could go back in time and re-run evolution, the kingdom of life would be drastically different from what is observed today because of amplifications of random effects. But we also keep seeing similar solutions evolved for the same problems (think about wings of flying animals). Similar phenomena are observed at the gene network level. We have designed tools to simulate evolution of biological networks in the computer (including a software called phi-evo), and have applied it to multiple problems, from immune recognition to cell size control (the latter in collaboration with the Skotheim group at Stanford)
Controlling and Understanding Developmental Oscillators
A biological oscillator called the “Segmentation clock” is controlling vertebrae formation in vertebrate embryos. Non-linear physics predicts how typical oscillators can be controlled and entrained . Can we hack the segmentation clock and entrain it ? In this collaboration with the Aulehla lab at EMBL Heidelberg, we entrain the segmentation clock and characterize its response, developing a more quantitative theory of non-linear segmentation oscillation.
Geometry of Embryonic Development
Individual cells in developing embryos perform multiple computations, to decide which cellular types they will eventually become. We use dynamical systems theory to understand how cells take such decisions. We could in particular identify from first principles specific geometries in phase space associated to developmental processes, such as vertebrae formation. There, we propose that the well described segmentation clock transitions towards bistability through Infinite Period Bifurcations, explaining many observed features in embryos such as the propagation of genetic waves of expressions.
In a recent collaboration with the Gregor group, at Princeton, we leveraged simple machine learning techniques (auto-encoder, SVD) to characterize the geometry in latent space of fly gap genes expression. This allows to extract a simple model manifold, explaining in particular how positional information is efficiently encoded
Simplifying Genetic Networks
Evolved genetic networks are usually super complex, possibly suggesting that entangled “biological hairballs” are fundamentally irreducible. We leveraged recent advances is complex systems theory to propose an algorithm for network reduction, preserving network function. We applied it to multiple problems and showed how very complex networks can in fact be reduced to simple, core modules, that can be analytically studied.
Machine Learning and Biology
Non neural Biological systems perform computation and learn. We believe we can learn a lot on such Biological Intelligence from recent advances in artificial intelligence and machine learning. For instance, we were able to draw mathematical parallels between the phenomenon of immune antagonism and adversarial examples in artificial intelligence. This also suggests that some disease might be best understood as computational problems.