Representation learning for biology
Self-supervised models for protein and nucleotide sequences.
The Dallago lab at Duke University.
We study what machine-learned representations capture about biology, where they break down, and how to turn them into reliable tools for discovery and molecular engineering.
Our research covers self-supervised sequence and structure models, atomistic generative models and agentic workflows for biomolecular design, GPU-accelerated search and complex prediction across proteomes, and benchmarks evaluated under realistic experimental splits.
Self-supervised models for protein and nucleotide sequences.
Atomistic generative models for proteins, binders, and multi-component biomolecules.
Automated systems that translate biological mechanisms into computational assays and iterative design campaigns.
GPU-accelerated homology search and quaternary structure prediction across cellular and viral proteomes.
Datasets and splits that measure generalization in real protein-engineering campaigns.
Preprint 2026 · NVIDIA + Duke
A generative model operating in a shared chemical and conformational latent space across standard and modified amino acids, glycans, DNA, and RNA. It combines recurrent latent refinement with constraint-guided search, with experimental validation across therapeutic binders, post-translational modification recognition, glycosylation control, and de novo enzymes.
Preprint 2026 · NVIDIA
An agent-assisted workflow that converts a biological objective or paper into a mechanism-specific computational assay and coordinates iterative binder design across multiple generators. In wet-lab testing, 15 of 18 designs bound calcium-loaded calmodulin with subnanomolar to low-nanomolar BLI affinities, and 13 of the 15 showed no detectable binding under calcium-depleted conditions.
Preprint 2026 · NVIDIA + Duke
Monomer and dimer predictions across 2,812 viral proteomes from 23 viral families relevant to human health. Structural clustering consolidated 8,028 high-confidence dimers into 1,598 interface clusters, 29.5% without detectable similarity to experimental interfaces in the PDB, informing glycoprotein vaccinology and viral protease specificity.
Preprint 2026 · NVIDIA + Duke
We predicted 31 million candidate protein complexes across 4,777 proteomes, calibrated confidence criteria against experimentally determined assemblies, and released 1.81 million high-confidence predictions through the AlphaFold Database, 31.3% of which extend beyond detectable PDB coverage.
Nature Methods 2025 · NVIDIA
GPU-optimized gapless prefiltering and gapped profile alignment for protein sequence search, reaching roughly 100 trillion cell updates per second while maintaining high throughput on consumer and low-wattage GPUs.
ICML 2026 Oral · NVIDIA + Duke
Seven curated fitness datasets and splits covering enzymes, protein–protein interactions, and light-sensitive proteins under realistic distribution shifts, where simple regularized baselines frequently match or outperform fine-tuned protein language models.
Chris is a Senior Research Scientist and Applied Research Team Lead in Digital Biology at NVIDIA, and a Visiting Assistant Professor at Duke University, where he leads this lab. Work listed on this site spans both roles, and publications are tagged with the affiliation under which they were carried out.
The Duke lab is independent of NVIDIA. Joining the Duke group does not imply employment at, or access to the resources of, NVIDIA. For NVIDIA roles, see the NVIDIA careers page; for the Duke lab, see Join.
We supervise students at Duke and collaborate with researchers elsewhere through jointly supervised theses and postdoctoral fellowships. Eligibility requirements and application instructions are on the Join page.