Mila

Multimodal diffusion model that co-designs protein sequence and 3D structure around cofactors and small molecules for de novo heme enzyme design.

Generative foundation model that produces atom-level molecular dynamics trajectories for protein monomers and protein-ligand complexes.

Protein inverse folding model aligning ProteinMPNN by multi-objective preference optimization to improve developability without losing fold fidelity.

Diffusion model predicting single-cell responses to genetic or drug perturbations, generating over distributions to capture population variability.

Family of autoregressive genomic foundation models that reconcile k-mer tokenization with single-nucleotide resolution at contexts up to 98k bp.

Self-supervised foundation model for 3D brain MRI, learning transferable anatomical representations from unlabeled scans for disease classification.

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

Protein conformation ensemble generation aligned to force-field energies, calibrating an AlphaFold 3-style diffusion model against MD thermodynamics.

AMPLIFY post-trained to absorb 3D structure from protein graph networks, lifting trRosetta contact precision from 0.253 to 0.320 at P@L/5.

ESM-2 post-trained to absorb 3D structure from protein graph networks, raising contact precision on a withheld CASP16 set by 59%.

Structure-based drug design model pairing an autoregressive transformer for ligand graphs with a diffusion head for 3D binding-pose coordinates.

De novo small molecule generation that assembles drug-like graphs atom by atom, pretrained on cheap property proxies and finetuned per objective.

Protein structure tokenizer that encodes 3D backbones as discrete VQ-VAE tokens, fixing the codebook under-utilization that caps their vocabulary.

Antibody affinity maturation framework that steers flow-matching structure generation with a binding predictor, then mutates CDRs by inverse folding.

Long-context generative genomic foundation model with a 98k-nucleotide window, trained on 386 billion bases of eukaryotic DNA for sequence design.

Structure-based drug discovery transformer that handles protein-ligand docking and pocket-aware 3D molecule design in one pretrained model.

Geometric foundation model matching enzymes to the reactions they catalyze, trained on 1.5 million structure-informed enzyme-reaction pairs.

De novo enzyme design conditioned on the reaction to be catalysed: substrate and product SMILES in, catalytic pocket, enzyme, and docked complex out.

Protein conformational ensemble generator that fine-tunes ESM3 with masked discrete diffusion over structure tokens instead of 3D coordinates.

Enzyme catalytic pocket design conditioned on a reaction: substrate and product in, pocket backbone, sequence, and EC class out.

Generative model of bacterial gene content that expands a handful of chosen KEGG modules into the full gene complement a viable cell would need.

Zero-shot protein conformation sampling that perturbs an input structure and anneals it back with a score network trained only on crystal structures.

Joint sequence-structure protein representation framework that fuses ESM-2 language model embeddings with GearNet geometric graph neural networks.

Geometric relational graph neural network that encodes 3D protein structures through geometry-aware message passing and self-supervised pretraining.

RNA virtual screening that reads a binding site's base-pairing graph and predicts the chemical fingerprint of its ligand to rank compound libraries.