MIT CSAIL

Protein dynamics model predicting per-residue covariances and residue-residue coupling from a static structure for fast RMSF and ensemble generation.

De novo protein backbone generator trained on low-confidence AlphaFold structures as corrupted data, reaching 86% designability at 700 residues.

Protein backbone generation from voxelized 3D densities, pairing a 3D CNN autoencoder with latent flow matching for shape and motif conditioning.

Open model that jointly predicts biomolecular structure and small-molecule binding affinity, approaching FEP+ accuracy in seconds on a single GPU.

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

2D/3D X-ray to volume registration for surgical navigation, aligning live fluoroscopy to preoperative CT or MRI after a five-minute fine-tune.

Protein-protein interaction language model that embeds whole sets of interacting chains through cross-chain attention, not one sequence at a time.

Cell Painting image generation conditioned on a control well image and a compound's structure, covering cell lines and chemicals never trained on.

Molecular property prediction from 3D structure for assays with dozens of labels, pretrained by extreme denoising plus DFT and LLM auxiliary labels.

Multimodal LLM for inverse molecular design, interleaving text and graph generation with a diffusion transformer and A* retrosynthetic planning.

Molecular dynamics surrogate that generates whole trajectories, switching between forward simulation, transition path sampling and upsampling.

Interactive foundation model for biomedical image segmentation, prompted with scribbles, clicks, and bounding boxes to segment unseen structures.

Keypoint-based foundation model for brain MRI registration, pretrained on over 100,000 3D volumes for rigid, affine, and deformable alignment.

Protein language model that predicts which of twelve subcellular compartments and biomolecular condensates a human protein partitions into.

Blind protein-ligand docking that transfers to binding domains absent from training, scoring 22.6% top-1 on DockGen and 50% on PoseBusters.

De novo protein backbone generation by SE(3) flow matching, with motif-scaffolding built in. Samples a designable backbone in seconds on one GPU.

Medical image segmentation model that solves unseen segmentation tasks in context from a few labeled examples, with no retraining or fine-tuning.

Rigid protein-protein docking by diffusion over the rigid-body pose, with a confidence model ranking sampled complexes. Median C-RMSD 4.85 on DIPS.

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

Blind protein-ligand docking by diffusion over ligand translations, rotations and torsion angles, reaching 38% top-1 accuracy on PDBBind.

Rigid protein-protein docking model that predicts a complex from two unbound structures in one pass, with no candidate sampling or refinement.

Chemical language model pretrained on up to 77 million PubChem SMILES strings for molecular property prediction on the MoleculeNet benchmarks.