Duke University

EHR foundation model that reads each ICU hour as clinical text and rolls patient state forward autoregressively in a shared latent space.

Sequence-based discrete-diffusion framework that designs peptide binders with specified agonist or antagonist behavior against GPCR targets.

Foundation model for paired ECG Lead-II and PPG waveforms, fusing the two through cross-modal attention and adaptive residual vector quantization.

Transcriptomics-native single-cell foundation model that learns batch-invariant cell representations and probabilistically generates virtual cells.

Post-hoc method that restores monotonic scaling to ESM-2 embeddings, yielding Matryoshka-style nested representations for variant effect prediction.

De novo peptide sequencing model that aligns tandem mass spectra with protein language model embeddings through constrained optimization.

De novo design of heavy metal-binding peptides by classifier-guided diffusion over ESM-2 embeddings, with Cu and Zn binders validated in vitro.

Vision foundation model for MRI, pretrained on 6.9 million slices across 18 body locations for label-efficient segmentation and classification.

Protein-DNA binding prediction and binder design from sequence, aligning protein and DNA language model embeddings instead of co-folding a complex.

Subject-level disease prediction from scRNA-seq, pairing cell-type-grouped scGPT pretraining with a Reactome pathway-constrained decoder.

Surgical video foundation model pretrained by entropy-maximizing compression on 0.78M unlabeled frames from 2,122 minimally invasive procedures.

Non-coding RNA language model using masked discrete diffusion to unify sequence generation with representation learning, trained on 30M ncRNAs.

Single-cell ATAC-seq foundation model that builds cell representations from non-zero chromatin peaks via peak-to-gene alignment.

RNA backbone conformer assignment from low-resolution maps, using Bayesian posteriors over a learned library of 3D suite shape clusters.

Drug-target interaction model that compresses any compound into a 15-bit hierarchical code, so billion-compound libraries can be screened in seconds.

Peptide representation model that DoRA-tunes ChemBERTa on 100,000 modified and bioactive peptide SMILES for therapeutic property prediction.

Template-guided protein design model that miniaturizes, diversifies, or expands a natural protein by decoding a fixed-size probabilistic encoding.

De novo peptide binder design framework that targets specific motifs, including disordered regions and conserved epitopes, from target sequence alone.

Protein language models pretrained on Rosetta biophysics simulations rather than evolutionary data, then finetuned on small experimental assays.

Antimicrobial peptide generator running denoising diffusion in the continuous ESM-2 embedding space, validated in mouse infection models.