Research program · Brigham and Women's Hospital · Harvard Medical School

From molecular interventions to system-level responses.

I develop AI methods that learn how a molecular intervention—a small molecule binding a protein or RNA—reshapes interactions, networks, cellular states, and ultimately phenotype, and I hold every model to one standard: it has to be trustworthy.

Five scales where biological states can be measured and modeled. Solid work sits at the first three; the last two are the next steps.

Core question

How does a molecular intervention propagate across biological scales?

Most biological AI models live within a single scale: a binding affinity, a modification site, a cell type. Living systems are coupled across scales, and the quantity that matters clinically is the response of the whole system. My work builds models at each scale that are explicit about mechanism, honest about uncertainty, and consistent with each other, so they can be composed into a world model of biological response.

Mechanistic representationLearning across molecular modalities Network propagationUncertainty and experimental validation

Interaction · Understand

Explaining how molecules recognize proteins and RNA

ExplainBind predicts protein–ligand binding through explicit non-covalent interactions, so every prediction exposes its physicochemical determinants instead of a bare score. Ongoing work extends the same mechanistic framing to RNA–ligand recognition.

ExplainBind (bioRxiv 2026)InteractBind benchmark (arXiv 2026)

Scale: Biomolecular interaction
Question: Which non-covalent interactions determine binding, and why?
Trustworthiness: Interpretability, mechanistic consistency
Status: Preprint; RNA–ligand extension in progress

Interaction · Predict

Interpretable prediction of RNA modification in biological context

RNA modifications sit between sequence and function. EvoRMD integrates biological context with evolutionary RNA language models to predict modification sites interpretably, a step toward modeling how interventions on RNA propagate to cellular state.

EvoRMD (Genome Biology 2026)

Overview of EvoRMD. Figure 1 from Wang et al., Genome Biology 2026 (CC BY-NC-ND).

Molecule · Understand / Predict

Learning molecular representations across complementary modalities

Chemical structure is observed through complementary views: graphs, spectra, text, and 3D geometry. Aligning them, asymmetrically and with relational learning, yields representations that transfer to property prediction and drug discovery while staying consistent with chemistry.

Asymmetric contrastive learning (JCIM 2025)Multimodal fusion (Commun. Chem. 2025)Mat-Instructions (IJCAI 2025)

Multimodal fusion with relational learning. Figure 1 from Zhou et al., Communications Chemistry 2025 (CC BY-NC-ND).

Molecule · Understand

Self-verifying structure elucidation from spectra

Determining a molecule's structure from its spectra is a problem where a model must be able to check its own answer. This line builds solvent-aware 2D NMR prediction, an annotated experimental benchmark, and reasoning-based elucidation from mass spectra, toward NMRScape: self-verifying structure determination.

TransPeakNet (Commun. Chem. 2025)2DNMRGym (LoG 2025)LLMs on mass spectra (J. Cheminform. 2026)

Model design and training strategy of TransPeakNet. Figure 1 from Li, Xu et al., Communications Chemistry 2025 (CC BY-NC-ND).

Network · Understand / Intervene

From molecular interactions to disease networks and drug repurposing

Molecular interactions act through networks. Connecting binding and interactome structure to disease modules provides the route from a single intervention to a system-level response, and the basis for principled drug repurposing. Ongoing work builds affinity-weighted protein–protein interaction networks.

Systems chemical biology (2025)Drug repurposing review (2026)

Affinity-weighted interaction networks. A drug binds its target; the perturbation propagates along weighted edges into a disease module. Schematic, work in progress.

Methods · Trustworthiness

Foundation-model methods that stay reliable on scientific data

Scientific data are scarce and shift between experiments, so the foundation models behind these systems must adapt cheaply and remain grounded. This work studies which parameter directions matter during adaptation and how to keep multimodal models faithful to their inputs.

Not All Directions Matter (ACL 2026)VIGIL (ECCV 2026)HyperAdaLoRA (ACL Findings 2026)Sensitivity-LoRA (EMNLP Findings 2025)

a, Structured, task-aware low-rank adaptation keeps only the parameter directions that matter. b, Counterfactual visual alignment trains multimodal models whose answers must change when the image changes. Schematic.

Open resources

Benchmarks, datasets, and models built around these questions

BenchmarkInteractBindRoughly 100k protein–ligand pairs with residue–atom interaction maps across six non-covalent interaction types, binding-site annotations, affinities, and similarity-controlled splits.Paper Dataset Code

Benchmark2DNMRGymAnnotated experimental 2D NMR dataset for atom-level molecular representation learning.Paper Dataset Code

Outlook

Closing the loop from molecule to patient

Cellular perturbation response

Extend mechanistic, uncertainty-aware prediction to single-cell and Perturb-seq data, so that interventions learned at the molecular and network scales can be tested against measured cell states.

Precision intervention

Invert the models: identify the molecular interventions that move a biological system toward a desired state, with explicit uncertainty and experimental validation in the loop. At the patient scale, phenotype is the readout: a first step is predicting TP53 mutation status in hepatocellular carcinoma directly from MRI (BMC Med. Imaging 2025).

Foundations

Earlier work in cobalt-catalyzed asymmetric radical reactions and total synthesis (Chem 2024, Chem 2021) is the physical-organic foundation for building mechanism-aware models.