An interdisciplinary group at the intersection of artificial intelligence, statistics, and cancer research — developing computational tools that support cancer detection, diagnostic refinement, and therapeutic decision-making.
Models, tools, and methods we develop and share across biomedical data analysis, molecules, biomedical tables, networks, and single-cell biology.
Language-model agents that analyse biomedical data step by step, with open-weight or proprietary models.
An AI agent that answers research questions from biomedical data. Give it a folder of data and a question; it analyses the data with an open-weight or proprietary model, records every step, and returns an answer. Evaluated on five biomedical benchmarks and 24 clinical prediction tasks against Codex CLI, Claude Code and Biomni A1, with the same models for every agent.
Explore JeusPretrained representations designed to transfer across molecules, cells, networks, and biomedical tasks.
A topology-first foundation model that describes each node by its structural role, aligns that description with a graph representation, and transfers the resulting knowledge to unseen biological and non-biological networks.
Explore Graph Foundation Model
A vision foundation model for single-cell biology. It renders each cell as a continuous image via optimal-transport gene cartography and learns one frozen encoder by masked-image pretraining on 72 million human cells — the most accurate zero-shot cell-type annotator on every held-out atlas we tested.
Explore scVision
A conversational framework that brings four molecular specialists together. A language-model orchestrator coordinates molecule generation, description, property prediction, and editing, using chemistry tools to examine structures and guide the next step.
Explore MolCrewMethods that describe, construct, and learn from relational structure in biological systems.
A deterministic compiler that states a network's structure in words. It runs classical graph algorithms, applies one fixed rule per structural role, and writes each role with the measurements that triggered it and what removing that part would do — so a frozen language model can reason about topology and counterfactuals with no task-specific training.
Explore BioGlyph
A knowledge-guided framework that turns each patient transcriptome into a pathway-structured molecular graph. Graph neural networks learn from patient-specific expression while preserving curated gene–gene organization for prediction and interpretation.
Explore Graph-in-Graph
A semantic-cartography framework that independently lays out network structure and node attributes, then stacks the size-matched channels into a fixed-size representation for vision models. The analysis preserves a path from feature-channel attribution back to the original node attributes.
Explore Graph2ImageModels that discover task-relevant organization directly from high-dimensional biomedical tables.
A differentiable framework that learns where tabular features belong in a two-dimensional map while it learns the prediction task. Retrain Dynomap on a labelled dataset, evaluate pooled out-of-fold performance, and inspect its learned topology and feature attribution.
Explore DynomapWe develop deep-learning and statistical methods that turn high-dimensional biomedical measurements — genomics, proteomics, imaging, and clinical data — into interpretable, transferable representations, and translate them toward clinical practice and biological discovery.
Tabular data hides the relationships between features. We reconfigure each sample into a spatially semantic 2D topographic map (TabMap) that keeps feature values as pixel intensities and encodes feature relationships as spatial distance — letting 2D convolutional networks extract association patterns while ranking features by importance.
Nat. Biomed. Eng. 2025Most deep-learning applications are regressions, whose features lie on a complex high-dimensional continuum that resists visualization. Our manifold discovery and analysis (MDA) method learns the manifold topology tied to a network's outputs, preserving local geometry to reveal a model's appropriateness, generalizability, and adversarial robustness.
Nat. Commun. 2023In radiation oncology and medical physics, we integrate diverse data types — medical imaging, clinical records, genomic profiles, and treatment parameters — to uncover complex patterns across modalities that are invisible in any single one, improving diagnosis, prognosis, and therapeutic decision-making.
Related publicationsWe are located in the Stanford Research Park in Palo Alto.
Free visitor parking is available on-site at 3145 Porter Drive.
Get directionsMd Tauhidul Islam, PhD
Principal Investigator · Room A202
tauhid@stanford.edu
We are looking for undergraduate and graduate students and postdocs to join the lab. If your expertise and interests match our projects, please email tauhid@stanford.edu.