|
09:15 to 09:30 |
Professor Rajesh Gopakumar (Centre-Director, ICTS-TIFR) |
Introduction |
|
|
|
09:30 to 10:15 |
Rituparno Mondal (Raman Research Institute ) |
Learning in Soft Matter Over the past decade, there has been remarkable growth in the use of machine learning (ML) tools for solving a wide range of scientific problems. Similarly, ML methods have demonstrated significant potential in the study of soft matter systems. In this talk, I aim to present selected case studies that highlight the application of neural network–based machine learning and physical learning in soft matter problems. Specifically, I will discuss three problems: (i) using neural networks to infer interaction potentials from structural data, (ii) applying graph neural networks to predict force chains in granular materials, and (iii) using physical learning to train materials to perform complex functions.
|
|
|
|
10:15 to 11:00 |
Shrisruti Sriraman (NCBS-TIFR) |
Spatial transcriptomic study of invasion markers in oral squamous cell carcinoma Oral squamous cell carcinoma (OSCC) is one of the most prevalent malignancies in India and is associated with high rates of recurrence and poor clinical outcomes. Pattern of invasion (POI), a histopathological descriptor of the tumor-stromal interface, is a robust predictor of invasion risk and recurrence in OSCC. Despite the strong correlation between POI and clinical outcomes, the molecular basis of aggressive invasion remains incompletely characterised.
Using NanoString GeoMx spatial profiling on an in-house cohort of 19 OSCC patients, we compared tumour and stromal regions of high-risk and low-risk POI. Differential expression revealed upregulation of extracellular matrix remodelling, partial-EMT, and hypoxia programs alongside an immunosuppressive microenvironment in high-risk POI tumor regions. Further, transcription factor activity analysis revealed coordinated TGF-β,AP-1,NF-κB signalling coupled with suppressed antigen presentation and epithelial differentiation programs. These findings resulted in a comprehensive invasion signature gene expression list.
Stratification of HNSCC patients from the TCGA cohort using our invasion signature demonstrated significant differences in progression-free survival. Cross-validation of these signatures against independent HNSCC single-cell and spatial datasets corroborated the spatial organisation of these programs at the tumor invasive front. Somatic mutation profiling further suggested enrichment of TP53, CDKN2A, HRAS, and CASP8 alterations in high risk POI tumors.
Since next generation sequencing remains impractical for wide clinical use, we are extending this work by applying deep learning-based spatial gene expression imputation directly from whole slide images. These findings position our invasion signature as a predictive biomarker for risk stratification and translate our findings into a scalable, cost-effective clinical tool suited to the Indian healthcare context, where OSCC disproportionately burdens patients with limited access to advanced molecular diagnostics.
|
|
|
|
11:00 to 11:30 |
- |
Tea Break |
|
|
|
11:30 to 12:15 |
Shaon Chakrabarti (NCBS-TIFR) |
(Disentangled) Representation Learning of the Mammalian Circadian Clock While Deep Learning models are increasingly being used to learn low dimensional “cell-states" of high dimensional biological datasets, how meaningful these representations are remains an open question. In a carefully controlled experimental setting where ground truth gene-expression states of the mammalian circadian clock are know by design, I will first demonstrate that current methods learn biologically meaningless states as a consequence of high levels of transcriptional noise. I will then demonstrate how a careful combination of pseudo-bulking and biologically inspired regularisation techniques can be used in a Variational Auto-Encoder setting to learn accurate circadian clock states. I will then extend these ideas to a different context — extracting clock states from ambulatory wearable datasets from human subjects in free-living conditions. Using a disentangled representation learning approach, I will once again demonstrate how biologically inspired regularisation helps learn meaningful low dimensional clock states. Overall, the two examples will serve to highlight the importance of embedding biological priors in Deep Learning models to learn meaningful low dimensional representations.
|
|
|
|
12:15 to 13:00 |
Paras Chopra (Lossfunk, Bengaluru) |
How to Build a Theory of Consciousness We are all intimately familiar with consciousness, yet the science of it is at a rudimentary level. Physics had its Newton and Einstein, but we're still at the dark arts stage for consciousness. In this talk, we'll approach the question of what it takes to "solve" the mystery of consciousness from a philosophy of science point of view, traversing metaphysics and our empirical reality. Then we will talk about how we're approaching this project at our lab and finally lay down a few (wild) hypotheses for what we believe consciousness could be about.
|
|
|
|
13:00 to 14:00 |
- |
Lunch |
|
|
|
14:00 to 14:45 |
Osheen Sharma (NCBS-TIFR) |
From Cellular Phenotypes to Cell States: Understanding Drug Response Across Scales High-content imaging (HCI) provides a rich description of drug-induced cellular phenotypes, but the information recovered from these data depends strongly on how those phenotypes are represented and how well the resulting models generalize across biological and experimental contexts. In this talk, I will present findings from my work on computational phenotypic profiling in cancer models, with a focus on ovarian cancer-fibroblast co-cultures. I will show how classical CellProfiler features and deep image representations derived through weakly supervised transfer learning recover mechanism-of-action (MOA) signatures across drug doses and cell lines, and how their performance depends on cell type, region-of-interest definition, and drug concentration. These analyses also show that morphological profiling can capture drug-induced changes that are not evident from viability measurements alone. I will further discuss how integrating morphological profiles with chemical-structure representations affects MOA prediction and model generalization under random and chemical-similarity-based data splits. Finally, I will briefly present recent ongoing work using graph neural networks to represent cancer-fibroblast co-cultures as spatial cell graphs, enabling single-cell morphology to be analysed together with local neighbourhood structure and cell-cell relationships. Together, these studies highlight how representation choice, multimodal information and spatial context influence the interpretation of drug perturbations.
These questions also form the basis of my current work within CALIBRE, where I am extending phenotypic profiling from population-level measurements toward biologically structured cell states. One project uses time-lapse imaging and lineage reconstruction to study whether pre-treatment morphological states persist across cell divisions and are associated with later drug fate. A second project focuses on inferring cell-cycle state from single-cell RNA-seq data using biologically constrained generative models, providing a complementary view of how intrinsic cell state may contribute to heterogeneous drug response.
|
|
|
|
14:45 to 15:30 |
Akshay S H (ICTS-TIFR) |
Explosive Synchronization in Networks of Type-I Neurons. Synchronization occurs across a wide variety of physical and Biological systems. The transition to synchronization in these systems can be smooth and continuous, or abrupt and discontinuous, with hysteresis in the order parameter. This abrupt synchronization transition, termed Explosive Synchronization (ES), has been studied extensively in the Kuramoto oscillatory networks. In the context of neuroscience, ES has been hypothesized as the mechanism behind some brain dynamics. But the studies involving neuronal models are sparse and have model-specific conditions under which ES is observed. In this talk, we will look at ES in networks of a weakly heterogeneous and weakly coupled class of neurons. Leveraging the recently demonstrated mapping between the Kuramoto and QIF models, we will realize different scenarios known to show ES in the oscillatory networks.
|
|
|
|
16:00 to 17:00 |
Sushrut Thorat (Lossfunk) |
Neural Networks as Scientific Instruments Can neural networks help explain cognition, rather than merely mimic it? I will show how they can serve as controllable experimental systems for testing mechanistic theories of the brain. Using two examples — how early visual experience may produce human-like shape bias, and how learning drives representational drift — I’ll illustrate how neural networks can turn qualitative ideas into testable predictions for biology.
This is based on a couple of papers:
|
|
|
|
19:00 to 20:30 |
- |
Dinner |
|
|
icts
res