All-atom protein–ligand co-folding model that feeds frozen pretrained ligand features into an AlphaFold 3-style backbone to place binding poses.
No providers recorded yet. Browse all providers
An all-atom structure model can recover the fold of a protein–ligand complex, score well on a complex-wide metric, and still put the drug in the wrong place: rotated in the pocket or shifted a few ångströms. For structure-based drug design that pose is the part that matters. Rather than asking the structure model to learn small-molecule chemistry from scratch, InteraFold-Chem borrows atom-level features from a separately pretrained molecular model and injects them into the network's ligand tokens.
InteraFold-Chem was released by MindRank (德睿智药), a clinical-stage AI drug-discovery company in Shanghai, in a September 2026 technical report from the MindRank Foundation Model Team. It is the first model in Intera ("Intelligence Era"), the company's planned series of life-science foundation models, and the first member of its InteraFold structure sub-family. It belongs to the same co-folding class as AlphaFold 3, Protenix-v2, Chai-1 and Boltz-2. The report describes its backbone as "AlphaFold 3–inspired" and does not name the base checkpoint or the pretrained molecular model that supplies the ligand features.
Two independent branches start from one shared checkpoint: InteraFold-Chem continues structural training on protein–ligand data, while InteraFold-Chem-RL adapts only the diffusion module with policy-gradient reinforcement learning, conditioned on a reference protein as a soft template.
The backbone follows the AlphaFold 3 layout: protein and ligand embeddings, MSA features, optional protein templates, a Pairformer trunk, an all-atom diffusion denoiser and confidence heads. The ligand interface adds ΔS = M ⊙ scatter[σ(G) ⊙ (HW + b)] to the single representation, where H holds pooled ligand-atom features and M is the ligand mask; only atom features are used, not pair features. Continued training updates the backbone, Pairformer and fusion module jointly under the inherited structural objective (coordinate, geometry, local-distance and confidence terms); the RL branch freezes every non-diffusion module and down-weights protein-atom errors. Neither uses the reference ligand at inference. Training data are protein–small-molecule complexes from the Protein Data Bank, organized through PLINDER and PoseBusters. The reported results use an early continued-training checkpoint.
Evaluation uses a curated, non-public internal cohort of 445 protein–small-molecule complexes, run at 4 recycles and 20 diffusion steps for every model. InteraFold-Chem reached 54.4% Success@2Å (ligand RMSD after protein-backbone alignment), a median ligand RMSD of 1.426 Å and complex lDDT of 0.924. The baselines scored 48.5%, 2.316 Å and 0.910 for Protenix-v2, 46.1%, 2.500 Å and 0.852 for Chai-1, and 22.9%, 7.753 Å and 0.600 for Boltz-2. InteraFold-Chem-RL scored 52.1% and 1.745 Å, with a lower complex lDDT of 0.779. The report bounds these numbers itself: the cohort is not released or shown to be disjoint from training data, the settings are not each model's recommended configuration, the RL branch alone receives a protein template, and no uncertainty intervals or PoseBusters validity rates are reported. It is an internal comparison, not a public benchmark, and does not establish generalization to unseen targets or ligands.
InteraFold-Chem is built for structure-based drug discovery: predicting how a candidate molecule binds its target so that medicinal chemists can reason about structure–activity relationships and propose modifications for synthesis and assay. In one retrospective case, a lytic transglycosylase–inhibitor complex (PDB 8GFD), it placed the ligand at 1.10 Å RMSD while the baselines missed by more than 8 Å. Code and weights have not been released. MindRank has announced a hosted InteraFold-Chem service for the fourth quarter of 2026. The report names binding-affinity prediction and ranking as future extensions, not current capabilities.
InteraFold-Chem is a concrete recipe for adding chemical pretraining to a co-folding model without retraining it from scratch: a small, zero-initialized, ligand-masked adapter that leaves the structure-generation pathway intact. It frames a question the field has mostly left implicit: whether a co-folder's ligand representation is the bottleneck for pose accuracy. The evidence so far is a company technical report with no preprint or peer review, and the report says independent, matched-condition studies are needed before its gains can be taken to hold more broadly. The model anchors MindRank's planned Intera series, which the company says will extend InteraFold to antibody–antigen and peptide–protein complexes.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.