New York University

Protein-sequence-conditioned mixture-of-experts language model that designs small molecules and emits an explicit synthesis route for each one.

Designs RNA and DNA aptamers against protein targets by backpropagating binding and anti-binding objectives through a frozen all-atom predictor.

Cell type annotation model mapping human single-cell and spatial transcriptomes onto one hierarchical typology of 381 types across 23 tissues.

Multimodal brain MRI foundation model pretrained on 1.55M T1w, T2w and FLAIR scans with a Mixture-of-Experts encoder and latent predictive objective.

Pan-cancer multi-omic foundation model encoding CpG-island DNA methylation and RNA-seq for zero-shot cancer classification and mutation prediction.

Multimodal conversational LLM for metabolite analysis, fusing a molecular-graph GNN and molecular-image CNN with a Vicuna-13B language backbone.

Gene Ontology prediction for malaria parasite proteins, trained on SAR supergroup structure embeddings with calibrated uncertainty and FDR control.

Multimodal transformer predicting alternative splicing outcomes across C. elegans neuron subtypes, reaching Spearman ρ = 0.88 on held-out exons.

Text-guided medical image synthesis across OCT, fundus, X-ray, CT and MRI. Synthetic data lifts downstream clinical tasks by 12-17%.

Antibody language model trained on human clonal families, proposing mutations that mimic in vivo affinity maturation for binding and stability.

Protein language models that split translated algal genomes into real genes and contaminants, classifying the dark proteome without homology search.

Generative masked protein language model with an interpretable concept layer, letting designers set 718 biophysical and annotation concepts directly.

Breast cancer recurrence risk read from an H&E slide and six routine clinical variables, using a frozen pan-cancer pathology encoder.

Pan-cancer digital pathology encoder: a ViT-L pretrained with DINOv2 on 400 million H&E patches drawn from 45,000 whole-slide images.

Equivariant graph neural networks pretrained by denoising, used frozen as geometric featurizers for analyzing conformational dynamics in simulations.

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

Transformer neural speech decoder for intracranial ECoG and sEEG contacts at arbitrary cortical positions, generalizing to unseen participants.

Promptable foundation model for universal medical image segmentation, fine-tuned from SAM on 1.57M image-mask pairs across 10 imaging modalities.

Inverse folding model predicting protein sequence from backbone coordinates, trained on 12 million AlphaFold2-predicted structures.