Samata Das

Latent Misfolding Estimation and Disease Prediction from Protein Entanglement Embeddings

August 2025 - Present

Build statistical and machine-learning models to infer latent protein misfolding states using sequence embeddings and structural embeddings, with the goal of predicting mutation-driven disease outcomes. The project integrates neural representations and logistic latent-variable models to identify entanglement regions that drive misfolding, quantify uncertainty, and provide biologically interpretable insights into how sequence and structural changes are associated with disease risk.

Inference on Multivariate Gaussian Processes via Deep Neural Networks for Astronomical Time Series Data Analysis

August 2025 - Present

Developing scalable likelihood-free inference methods for multi-band astronomical time-series data by using deep neural networks to approximate high-dimensional Gaussian-process likelihood surfaces. This work addresses the computational bottlenecks of classical GP models, enabling efficient analysis of LSST-scale datasets and improving the detection of faint astrophysical signals such as exoplanet variability, supernova precursors, and cross-band correlated events.

Partial Identifiability in Latent Misfolding Models for Protein Disease Prediction

October 2026 -

This work is going to develop a statistically principled framework for inferring latent protein misfolding from entanglement-related structural features when direct observations of misfolding are unavailable. By characterizing the conditions under which the misfolding mechanism is identifiable—and deriving bounds when full identification is impossible—the model provides transparent uncertainty quantification for both latent misfolding status and disease risk. Unlike black-box predictive models, this approach explains disease predictions through an interpretable causal pathway, Entanglement Features - Latent Misfolding - Disease Outcome.