Caltech

Protein structure tokenizer that encodes a whole structure globally, with each successive token adding detail for adaptive-length representations.

Flow-based protein structure tokenizer, a diffusion autoencoder replacing SE(3)-invariant components with global coordinates and standard attention.

Synthesizable 3D molecule generation that jointly samples building blocks, reactions, and atomic coordinates, returning a synthesis route per design.

Catalytic residue prediction from enzyme sequence alone, pairing frozen ESM2 embeddings with contrastive learning. Reaches F1 0.86 on Uni3175.

Functional ultrasound reconstruction with neural operators, rebuilding power Doppler brain images from 10-15% of the usual compounded frames.

Spatial genomics foundation model that predicts a cell's transcriptome from its neighbours and forward-simulates tissue response to perturbation.

De novo peptide design across non-canonical amino acid space, using guided diffusion over receptor-ligand interfaces to reach D-amino acid chemistry.

Cell phenotyping model for spatial proteomics using a language-informed vision transformer to classify cell types zero-shot across marker panels.

Conditional adapters over a frozen ProGen2 generate enzymes steered by EC number, taxonomy, or free-text function, including rare and unseen classes.

Universal cell segmentation model adapting Meta's SAM to segment mammalian cells, yeast, and bacteria across imaging modalities without retraining.