Ritambhara Singh
Brown University, RI·Computer Science·Early-career
A person's health is recorded in many incompatible forms at once: clinical notes, imaging scans, laboratory results, genetic screens. Each captures part of the picture, but they differ in format, in how many samples exist, and in the time scales they cover, which makes combining them hard for both clinicians and researchers.
- Fields
Computer Science
Genetics
Biomedical Engineering
- Topics
deep learning
machine learning
gene regulation
- Grants
- 2 known grants, through Jul 2030
What they work on
- Predict gene expression by integrating histone modification profiles, 3D genome conformation, DNA sequence and single-nucleotide polymorphisms in deep learning models, including graph neural networks that treat genome folding as a network of contacts between distant DNA regions.
- Extract the regulatory signals behind each prediction using neural network interpretation methods that score input features, and compare scores between healthy and disease cell lines to flag misregulated genes.
- Develop a Bayesian inference-based interpretation method to improve explanations from graph neural networks, expressing confidence in which graph connections drive a prediction.
- Extend the gene regulation frameworks to single-cell chromatin accessibility and DNA methylation data, testing whether imputation methods remove enough noise to model regulation across developmental stages.
- Integrate gene-gene interaction networks into counterfactual explanation methods, which ask what minimal input change would flip a prediction, to identify candidate genes for therapeutic targeting.
- Embed drug-drug interaction knowledge into large language models to predict adverse drug effects, and design joint representation learning techniques that handle sample size imbalance across imaging, clinical notes and genetic screening data.
More about the research
The same mismatch appears inside cells, where gene activity is shaped by chemical marks on DNA packaging, the folded three-dimensional layout of the genome, and sequence variants, all measured by separate experiments. Work in this group builds machine learning methods that merge such mismatched data sources and, crucially, explain their predictions, so that a model's output becomes a testable biological or clinical hypothesis rather than an unexplained number.
Why it matters
Patient data arrive in fragments, and models that cannot explain themselves are hard to act on. Methods that combine imaging, clinical notes, laboratory and genetic data while flagging the genes or drug interactions behind a prediction could support earlier diagnosis, more personalized treatment choices, and fewer adverse drug effects. On the genomics side, pinpointing misregulated genes in diseased cells gives experimentalists concrete hypotheses to test.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Yanming Li
Baylor College of Medicine, TX·Genetics
Identify the aortic cell types in which TAAD-associated SNPs act, by integrating GWAS results with single-cell ATAC-seq maps of accessible chromatin from human aortic tissue.
Martin Zhang
Carnegie Mellon University, PA·Genetics
Use multimodal single-cell and spatial genomics data to pinpoint, with high precision, the cell types and cellular contexts most important for diseases identified through genome-wide association studies (GWAS).
Sai Zhang
Yale University, CT·Genetics
Develop a sequence-based deep learning model, trained on single-cell and individual-level data, that predicts gene expression from DNA sequence and estimates how genetic variants alter gene activity in each cell type.
The school
Brown University: 23 labs on PI Pages, $375M a year on research, 251 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.