PhD Dissertation Proposal: Mahbuba Tasmin, Integrating Biological Signal to improve Generalizability and Interpretability of Machine Learning in Genomics
Content
Speaker:
Abstract:
The rise of drug-resistant Mycobacterium tuberculosis remains a critical global health challenge, motivating rapid genotype-based approaches for antimicrobial resistance prediction. While machine learning and foundation models provide promising tools for mapping genomic variation to resistance phenotypes, their clinical utility is limited by two persistent challenges: scarce clinically labeled data and limited biological interpretability. Sequence-only models can achieve strong predictive performance, but they may fail to reliably recover known resistance-conferring variants and can rely on spurious genomic correlations. This thesis develops a biologically grounded machine learning framework that improves generalization and interpretability by integrating three-dimensional protein structure, evolutionary constraint, and molecular energetics into pathogen genotype–phenotype prediction and evaluates each contribution under conditions that reflect how such models would be used in practice.
The thesis is organized around five connected contributions. First, I establish BIG-TB, a benchmark of more than 17,000 M. tuberculosis genomes that evaluates both phenotype prediction and resistance- variant discovery, including under lineage-held-out evaluation, exposing the interpretability limits of current sequence-based models. Second, using a proteome-wide unsupervised analysis of more than 31,000 clinical isolates, I show that resistance-associated mutations cluster in three-dimensional protein space, including in non-essential genes, supporting structural proximity as a biologically meaningful signal for resistance. Third, I incorporate this structural information directly into predictive models through Fused Ridge Regression, where coefficients are regularized according to residue adjacency in 3D protein space to improve generalization when learning from scarce training instances with highly sparse and redundant features .
Building on these structural insights, the fourth contribution develops a multimodal forecasting framework for variants of uncertain significance. By combining 3D structural context, Rosetta-derived biophysical energetics, and ESM-2 evolutionary likelihoods, this model forecasts World Health Organization variant reclassifications, recovering the majority of variants subsequently reclassified as resistance-associated, and prioritizes candidate resistance mutations among those currently classified as uncertain. The framework reveals distinct biological regimes: resistance in essential genes is more strongly associated with structural and energetic constraints, whereas resistance in non-essential genes is more strongly associated with sequence-level evolutionary likelihood.
Finally, the thesis investigates whether biologically constrained data augmentation can relieve the labeled-data bottleneck, using single-amino-acid substitutions filtered by protein language model likelihoods, and establishes the evaluation standards under which such gains can be trusted. This work motivates the proposed research: determining which form of biological information — evolutionary homolog context, biophysical priors, or variant effects measured in homologous proteins — most improves prediction of emerging resistance mutations. Two projects I supervise extend the framework, testing whether molecular docking contributes information beyond structural proximity, and whether the forecasting model generalizes to newer anti-TB drugs evaluated against an independently collected phenotypic cohort. Together, this thesis integrates biological signal into sequence modeling to build interpretable, data-efficient AI systems for antimicrobial resistance forecasting.
Advisor:
Anna Green