Monday, 21 July 2025
-
We will discuss models to probe and control how multicomponent, multiphasic condensates function.
Growth is a crucial feature of living systems, setting them apart from most inanimate physical systems. I will use statistical physics tools to quantify genome replication programs from DNA sequencing data extracted fron growing microbial populations. In the second part of my talk, by analyzing the spatial distribution of chromatophores on squid skin, I will show how tissue growth leads to a novel, unexpected physics of disordered packing.
TBA
-
Tuesday, 22 July 2025
TBA
I will describe high-throughput methods for mapping the genetic basis of complex traits in budding yeast. I will then explain how this data can be used as the basis for discovering lower-dimensional latent structure in the genotype-phenotype map, using either sparse structure discovery or attention models.
In the first part of the talk I will review a list of results on the macroecological structure of microbial communities. In the second part, I will discuss how microbial population can encode the statistical structure of the environment to anticipate changes of conditions.
Life today relies on numerous baroque mechanisms to maintain order: keeping itself far from the inanimate state of thermal equilibrium. Yet, while we understand how these mechanisms work, much less is known about how they come into existence in the first place. Here, we suggest that surprisingly little might be needed. We study the copying of information in different classes of biological macromolecules. In each, fast replication alone selects for the evolution of non-equilibrium error-correction. Our results rely only on geometric frustration effects intrinsic to the physics of multicomponent assembly. To test the theory, we develop a massively multiplexed Luria-Delbruck assay in yeast, capable of measuring thousands of mutation rates in a single experiment. Overall, our work sheds light on the origins of fidelity in biomolecular processes, and highlights the role of biophysical constraints in shaping the parameters of evolution.
-
Wednesday, 23 July 2025
Pseudo-bulking' over hundreds of single cells in scRNA-seq datasets is often essential to overcome biological and technical noise while extracting meaningful biological information. However, this process naturally leads to loss of resolution, and is especially problematic while analysing rare cell types within tissues. Here, taking the example of phase estimation of the mammalian circadian clock, I will show how a principled approach to dimensionality reduction via probabilistic latent space modeling can significantly reduce the number of cells over which coarse-graining has to be performed. Combining a deep-embedded clustering approach with spherical latent spaces, our method significantly improves over existing techniques, as I will demonstrate with initial results from various scRNA-seq and smFISH datasets.
TBA
Biological systems explore vast spaces of possibility, yet their function is often robust to certain variations while remaining exquisitely sensitive to others. A striking example is the hypervariable repertoire of T cell receptors, which underlies the specificity of the cellular immune response. Can we construct principled coarse-grained descriptions of this diversity that retain functionally relevant variation? In this talk, I will summarize some of our recent progress on this question. First, I will show how we can quantify, in bits, the information that different receptor regions provide about antigen specificity. Second, I will describe how contrastive learning can be used to infer which receptors recognise common targets. Finally, I will discuss how these statistical insights can advance the study of human adaptive immunity to infectious disease.
Many biological decision-making tasks require classifying high-dimensional chemical states. The biophysical and computational mechanisms that enable classification remain enigmatic. In this talk, using Markov jump processes as an abstraction of general biochemical networks, we reveal several unanticipated and universal limitations on the classification ability of generic biophysical processes
-
Thursday, 24 July 2025
Cells often rely on signalling molecules to make cell-fate decisions during development. We can see these fate-regulating decisions as computational tasks and measure the performance of these tasks with information-theoretic quantifiers. Yet, it is often unclear under what constraints a particular task is performed. Here, I will argue that we can nevertheless learn about the computational performance of a gene regulatory system. First, I will discuss an example in early fly development, where the computational task involves interpreting signals so that cell fates can correctly be distinguished. I will discuss how clustering of transcription factors can provide a way for how binding sites can achieve optimal information transfer, and achieve optimal bounds predicted from an information bottleneck approach. Second, I will focus on an example from the canonical Wnt pathway, where we explore synthetic signals and infer which signals would need to be present naturally for the computation to be close to optimality. We find that for appropriately chosen signals, the cellular response can be precise enough to allow reliable differentiation into two distinct states. As the precision in the pathway improves, more distinct states can be reliably distinguished.
Cell fate decisions emerge as a consequence of a complex set of gene regulatory networks. Models of these networks are known to have more parameters than data can determine. Recent work has instead mathematically formalized the concept of a Waddington landscape. These landscapes are minimally parameterized descriptions of cell fate decisions. We will describe how to construct these landscapes, with an example from early mammalian development. We will then show how they can be extended to describe Turing patterns.
TBA
Spatial profiling technologies like hyper-plex immunostaining in tissue, spatial transcriptomics etc have the potential to enable a multi-factorial, multi-modal characterization of the tissue microenvironment. Scalable, quantitative methods to analyze and interpret spatial patterns of protein staining and gene expression are required to understand cell-cell relationships in the context of local variations in tissue structure. Objective scoring methods inspired by recent advances in statistics and machine learning can serve to aid the interpretation of these datasets, as well as their integration with other, companion data like genomics. In this talk, we will discuss elements of spatial profiling from multiple studies as well as paradigms from statistics and machine learning in the context of these problems. This talk will also discuss the use of AI/ML and spatial analytics of the tumr microenvironment to derive spatial biomarkers of immunotherapy.
-
Friday, 25 July 2025
In this talk, I will present two vignettes: (1) how manifold learning allows for olfactory habituation in turbulent environments, and (2) a contrastive framework for identifying irreversible degrees of freedom in high-dimensional stochastic systems. Common themes of finding low-dimensional representations of high-dimensional time series data will be highlighted.
High-throughput DNA sequencing technologies are now fast and cost-effective. As a result, these technologies have been used to probe various biochemical activities of the genome, producing activity maps at a high resolution. The next step is to identify the sequence components at these regions critical for the activity. I will talk about our work on Bayesian mixture models targeted to identify diverse regulatory mechanisms directly from such data, which traditional approaches miss.
I review the roughly 10-year-old literature on application of deep learning techniques to DNA sequence. Much recent attention has been paid to single basepair resolution prediction of variant effect. In contrast, I focus on a simple classification problem, where DNA is classified into two categories (for example, binding or not binding a particular protein). We recently developed a neural network architecture that outperforms published methods significantly on this basic task. We show that in silico mutation analysis indicates that the classification is not strongly biased by single nucleotides but, instead, there are sequence patterns in an extended region that lead the classifier to the correct answer, even when the central signal (binding motif) is mutated out. This builds on earlier work where we showed the importance of flanking sequence in regulatory sites. Much as how we can recognize a photo of a tiger even when its face is removed, we suggest that there is significant surrounding signal in DNA even when the core motif is removed, though the in vivo function may require the core motif. Finally, I discuss the converse, harder problem of generating synthetic DNA sequence that has a desired biological function.
-
-
Monday, 28 July 2025
Proteins are partially disordered, creating entropy that encodes information for ligand selectivity and allosteric signaling. However, inferring entropy maps from protein dynamics has been difficult due to complex interactions between conformational degrees of freedom. I will discuss our recent progress in mapping shared entropy within protein structures to reveal how proteins compute using apparent disorder. Our approach, ciMIST, infers discrete whole-residue states from molecular dynamics simulations and converts joint residue probabilities into maximum likelihood tree networks of shared entropy. On the PDZ3 and ERalpha ligand-binding domains in multiple ligand and mutational states, ciMIST networks matched total entropy measured by calorimetry and residue-level entropy probed by NMR and hydrogen-deuterium exchange. ciMIST accurately predicts allosteric mutational hotspots in PDZ3 from saturation mutagenesis. ciMIST maps of the entire human steroid receptor family reconstructs family phylogeny, revealing entropic reprogramming of function within fixed protein folds. These results show that sparse entropy networks effectively map protein ensembles to function when static structure is insufficient.
Establishing causal relationships motivates many classical biological experiments, from single-gene knockdown perturbations to neuron patch-clamping. Yet modern biological experiments measure hundreds to thousands of genes, cells, or neurons simultaneously. Traditional causal inference methods cannot scale to such high dimensions, where a combinatorial explosion in potential confounding variables precludes isolating the effect of individual variables. I will describe my effort to use the physics of synchronization to identify causal interactions from high-dimensional time series datasets. Applying these approaches to gene expression measurements, we identify regulatory motifs and gene subcommunities inaccessible to other statistical inference tools. Our physics-driven framework scales to large datasets containing thousands of variables, facilitating systematic discovery of regulatory interactions in the high‑throughput era.
The physical principles governing biological heteropolymer interactions are fundamental to cellular organization and function. In this talk, we first present a physics-based theory derived from extensive molecular dynamics simulations that quantitatively captures the interactions driving protein phase separation and allows for the design of sequences with predictable behavior. We then extend this theme to chromatin, presenting experimental work on synthetic nucleosome arrays that isolates the direct physical impact of histone acetylation on chromatin architecture. Using single-molecule experiments and in vitro Hi-C, we demonstrate how this epigenetic mark modulates intra-chain interactions to control the polymer's conformational ensemble. Together, these studies provide a bottom-up, quantitative framework for understanding how sequence and chemical modifications tune heteropolymer interactions to shape biological structure and function.
I will give a brief introduction to using the cavity method using ecosystems as an example.
-
Tuesday, 29 July 2025
Proteins can fold spontaneously into well-defined three-dimensional structures and can carry out complex biochemical reactions such as binding, catalysis, and long-range information transfer. The precision required for these properties is achieved while also preserving evolvability – the capacity to adapt in response to fluctuating selection pressures in the environment. What is the basic design of proteins that supports all of these properties? Going beyond direct physical analysis, statistical analysis of genome sequences have, in recent years, provided a powerful and general approach to this problem. Using different methodologies, this approach has exposed direct structural contacts and collective functional modes within proteins, and has yielded generative models for protein design. In this talk, I will discuss these advances and will present approaches for probing the physical mechanisms implied by the evolution-based models. An important aspect is to understand how such mechanisms may be constrained by and originate from the dynamics of the evolutionary process. This work represents a step towards a theory for the physics of proteins that is consistent with evolution.
TBA
The large scale network of genes regulating E. coli dynamics will be discussed together with a method of breaking it down into tractable dynamical subsystems that have functional saliency.
-
Wednesday, 30 July 2025
TBA
TBA
TBA
Transformer models pretrained on large amounts of language data display a powerful feature known as in-context learning (ICL): the ability to parse new information presented in the context with no additional updates to the synaptic weights in the network. Recent work shows that ICL emerges when models are trained on a sufficiently diverse set of tasks and the transition from memorization to generalization is sharp with increasing task diversity. One interpretation is that a network's limited capacity to memorize favors generalization. I will present our analysis of this transition using a small transformer applied to a synthetic ICL task. Using theory and numerical experiments, we show that the sub-circuits that memorize and generalize can be viewed as largely independent. The relative rates at which these sub-circuits learn explains the transition from memorization to generalization, rather than capacity constraints. We uncover a memorization scaling law, which determines the task diversity threshold at which the network generalizes. The theory quantitatively explains a variety of other ICL-related phenomena, including the long-tailed distribution of when ICL is acquired, the bimodal behavior of solutions close to the task diversity threshold, the influence of contextual and data distributional statistics on ICL, and the transient nature of ICL.
-
Thursday, 31 July 2025
Mammalian development is characterized with transitions from homogeneous populations of precursor to heterogeneous population of specified cells. The cells in the population continuously sense and respond to time-varying signals that are secreted, and an accurate interpretation of these signals is necessary to determine cell phenotype in a robust and reproducible manner. We demonstrate that the cell phenotype is implicitly encoded in the temporal evolution of the system’s trajectory in signaling space, and it can be directly read-out and decoded in real-time by the immediate-early genes, much earlier than the network sets to a steady-state. We propose that signaling homeorhesis determines cell phenotype in dynamic environments, and demonstrate that organization at criticality is a necessary pre-requisite to guide the system’s trajectories through the changing landscape.
Elucidating the design principles of regulatory networks driving cellular decision-making is of fundamental importance in mapping and controlling cellular behaviour. Despite their size and complexity, large biological regulatory networks often lead to a limited number of cell-states/phenotypes. How this canalization is achieved remains largely elusive. Here, we investigated multiple different networks governing cell-state transition during cancer metastasis, and identified a latent design principle in their topology that limits their phenotypic repertoire – the presence of two "teams" of nodes engaging in a mutually inhibitory feedback loop. These "teams" are specific to these networks and directly shape the phenotypic landscape and consequently the cell-fate trajectories. Our analysis reveals that network topology alone can contain information about phenotypic distributions it can lead to, thus obviating the need to simulate them. We present experimental evidence of such "teams" in transcriptomic datasets across many contexts (cancer cell plasticity in breast cancer, melanoma, lung cancer etc.). Overall, we propose these "teams" as a network design principle that drive cell-fate canalization in diverse decision-making processes, and drastically reduce the dimensionality of the phenotypic space.
Mammalian development is characterized with transitions from homogeneous populations of precursor to heterogeneous population of specified cells. The cells in the population continuously sense and respond to time-varying signals that are secreted, and an accurate interpretation of these signals is necessary to determine cell phenotype in a robust and reproducible manner. We demonstrate that the cell phenotype is implicitly encoded in the temporal evolution of the system’s trajectory in signaling space, and it can be directly read-out and decoded in real-time by the immediate-early genes, much earlier than the network sets to a steady-state. We propose that signaling homeorhesis determines cell phenotype in dynamic environments, and demonstrate that organization at criticality is a necessary pre-requisite to guide the system’s trajectories through the changing landscape.
TBA
-
Friday, 01 August 2025
TBA
TBA
TBA
-
-