Research Program
From reading the literature to trusting the result
Our work follows a discovery from end to end: AI that reasons over the literature, models that explain the biology, and quality systems that make the outcome safe to use in the clinic.
AI for Accelerated Scientific Workflows & Deep Research
Science is bottlenecked less by data than by the time it takes to read, connect, and test it. We build agentic systems that handle the reading and the bookkeeping, so scientists can spend their time on judgment.
Driving question
Can a team of AI agents propose a hypothesis worth a scientist’s afternoon, and show its work?
- tracks: 4
- mode: agentic
- review: human-in-the-loop
- access: open
Multi-agent scientific reasoning
Specialist agents (retriever, critic, planner) that argue over evidence and expose their reasoning chain instead of returning a single opaque answer.
agents · critique · provenance
Automated hypothesis generation
Proposing and ranking candidate hypotheses by novelty, plausibility, and testability against the existing literature.
rank · novelty × testability
Biomedical knowledge graph mining
Linking genes, compounds, phenotypes, and papers into graphs where missing edges become research leads.
graph · link prediction
Laboratory protocol synthesis
Turning a hypothesis into a step-by-step, reagent-aware protocol that a bench scientist reviews before anything runs.
protocol · human-in-the-loop
Computational Biology & Molecular Systems
Biology is a systems problem. We model it across scales, from the fold of a single protein to the wiring of a cellular network, and ask what changes when one variant does.
Driving question
Which molecular changes actually matter, and can we predict them before the wet lab does?
- tracks: 4
- scale: atom → network
- code: open source
Structural bioinformatics
Reasoning over protein structure, folding, and binding interfaces to explain function from shape.
structure · fold · binding
Molecular property prediction
Learned representations of molecules that estimate properties early, before costly synthesis and assays.
graph models · property regression
Biological network topology
Mapping interaction networks to find the hubs, modules, and bottlenecks that hold a system together.
network · centrality · modules
Genomic variant impact analysis
Estimating how individual genomic variants propagate into changes in protein function and phenotype.
variant · effect prediction
Translational Quality & Regulated AI
A model that works in a notebook is not yet a model that works in a clinic. This lab applies the discipline of regulated pharmaceutical quality systems, drawing on the methodologies our founding lead used leading the Quality vertical at GSK, to AI that touches patients.
Driving question
What does it take for an AI system to earn trust in a regulated, high-stakes setting?
- tracks: 3
- standard: quality-by-design
- partner: clinical advisory
High-reliability healthcare pipelines
Data and model pipelines built to fail loudly: monitored, versioned, and checked for drift before outputs reach a clinician.
reliability · monitoring · drift
Regulated ML quality systems
Validation plans, audit trails, and change control for machine learning, adapted from pharmaceutical quality practice.
validation · audit trail · change control
Clinical data validation
Checking clinical datasets for completeness, consistency, and lineage so models are trained on data that can be defended.
data quality · lineage
Clinical relevance is reviewed with our Chief Clinical Advisor, Dr. Sasidhar, MD (General Medicine), of Sasidhar Hospitals.
Meet the teamWorking on a related problem?
We collaborate with labs, clinicians, and grant partners across all three pillars.