Columbia University

Protein-ligand co-folding model on the OpenFold3 architecture, trained on PDB structures through June 2025 with inference-time chemical steering.

Binary protein classification over frozen ESM3 embeddings, using transformer contextualization and learned attention pooling instead of fixed pooling.

Spatial multi-omics integration model aligning RNA, protein, metabolomics, and histology to map cell-state gradients and cell-cell interactions.

Causal multimodal transformer that embeds the do-operator in attention to predict single-cell gene expression under unseen genetic perturbations.

Generative model for self-tracked menstrual health data, encoding cycles and symptom logs as token sequences for synthetic cohorts and forecasting.

Transformer that infers whole-genome DNA methylation from gene expression, generalizing zero-shot to unmeasured CpG sites and unseen samples.

EEG foundation model for brain-computer interfaces, pairing masked pretraining with a mixture-of-experts transformer across electrode montages.

Neuro-oncology foundation model for brain tumor MRI, using distributionally robust pretraining for molecular subtyping and survival prediction.

EHR foundation model for structured OMOP timelines that unifies patient embeddings, zero-shot outcome prediction and synthetic record generation.

Single-cell transcriptomics foundation model that encodes gene regulatory network structure into self-attention through graph signal processing.

TCR-pMHC specificity prediction that folds frozen ESM-2 and AlphaFold2 representations into a three-body peptide-MHC-CDR3 energy tensor.

Refines AlphaFold2 predictions against cryo-EM, cryo-ET, and X-ray data by optimizing coevolutionary embeddings rather than atomic coordinates.

Sparse autoencoders on the ESM-2 residual stream that expose interpretable protein features, with an open visualizer for what each latent detects.

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

Cryo-EM heterogeneous reconstruction that models particles as one of K neural fields, resolving compositional and conformational states ab initio.

Histopathology foundation model aligned to spatial transcriptomics by a cross-modal ranking loss, embedding H&E patches without gene input.

Light-weight rigid protein-ligand docking model adapting the AlphaFold2 architecture, with a binding affinity module for virtual screening.

Protein language model trained from scratch on MD and normal-mode dynamics, representing residue fluctuation and co-movement from sequence.

Zero-shot mutation effect scoring for designed and viral proteins, mapping frozen ESM2 representations onto MD and normal-mode dynamic properties.

Manufacturing-aware generative sequence models whose parameters are DNA synthesis reaction conditions, so designs are made in vitro at petascale.

Neural ab initio reconstruction for cryo-EM and cryo-ET that jointly infers particle poses and a continuous landscape of conformational states.

EHR foundation model that writes patient histories as token sequences carrying explicit visit and day-interval tokens, invertible back to OMOP tables.

Spatial proteomics imputation from a 7-plex immunofluorescence panel, generating in silico CODEX expression for 33 more biomarkers per cell.

De novo protein backbone generation by SE(3) diffusion over rigid-body frames. Samples designable monomers up to 500 residues without pretraining.

BERT adaptation for structured OMOP electronic health records that inserts artificial time tokens between visits so intervals shape the embeddings.