IntFold

IntFold is an all-atom biomolecular structure prediction model released in July 2025 by IntelliGen AI, a Hong Kong startup working on AI for drug discovery. Like AlphaFold 3, whose architecture and training recipe it follows, it predicts the joint structure of proteins, nucleic acids, and small-molecule ligands from sequence, and it arrived alongside a wave of open all-atom co-folding models including Boltz-1, Chai-1, Protenix, and HelixFold3.

What separates IntFold from those reproductions is controllability. Instead of treating folding as a single fixed mapping from sequence to coordinates, IntFold inserts lightweight LoRA adapters into a frozen embedding trunk, so one base model can be steered toward a particular conformational state of a target family, toward a user-supplied binding pocket or epitope, or toward a numerical binding-affinity readout. The technical report pairs this with a training-free, model-agnostic ranking scheme that selects the most self-consistent structure from a set of diffusion samples.

The project is now developed and distributed under the name IntelliFold: code at IntelliGen-AI/IntelliFold under Apache 2.0, weights on HuggingFace, a pip package, and a hosted prediction server. A second generation, IntelliFold-2 (February 2026), adds latent-space scaling in the Pairformer blocks, stochastic atomization for atom attention, policy-guided diffusion sampling, and difficulty-aware loss reweighting, shipping as Flash, v2, and Pro variants with inference built on AlphaFold 3's Apache-2.0 JAX engine. Both reports are preprints.

#Key Features

• Adapter-based specialization: A LoRA adapter trained on roughly 200 CDK2 complexes recovered 4 of 5 rare allosteric closed-state conformations that the general model missed entirely, while retaining correct predictions on all 35 orthosteric structures in the test set.
• Constraint-guided folding: A constraint adapter with a dedicated embedder for known interacting residues raises the PoseBusters success rate from 79.5% to 89.7% and more than doubles antibody-antigen success, from 37.6% to 69.0%.
• Binding affinity estimation: A post-hoc module of four additional Pairformer blocks predicts protein-ligand affinity, correlating with CASP16 L1000 experimental values at a Pearson coefficient of 0.53 against 0.47 for Boltz-2.
• Consensus-based ranking: Generating 25 predictions (5 samples across 5 seeds) and keeping the one with the highest mean pairwise DockQ to the rest improves antibody-antigen success by about 3% over random selection, without additional training.
• Custom attention kernel: FlashAttentionPairBias, a Triton kernel built on FlashAttention-2, broadcasts the pair bias on the fly rather than materializing the full tensor, cutting peak memory and forward-pass latency relative to standard implementations.

#Technical Details

IntFold comprises a Pairformer-based embedding trunk, a diffusion module, a confidence head, and optional adapters, following the AlphaFold 3 supplementary methods. Training used PDB structures with a cutoff of 30 September 2021, supplemented by three distillation sets: AlphaFold Database monomers filtered at 30% sequence identity and global pLDDT above 85, disordered-region predictions from AlphaFold-Multimer v2.3, and antibody-antigen complexes derived from PLAbDab. The affinity module was trained on roughly three million log-scaled activity measurements spanning 27,000 proteins and one million ligands, merged from ChEMBL, BindingDB, GalaxyDB, BioLiP, and PubChem. MSAs come from a ColabFold/MMseqs2 pipeline, with Jackhmmer searches of UniProt supplying species-paired alignments; RNA is folded without MSAs.

On FoldBench, IntFold reaches a mean LDDT of 0.88 for protein monomers and a 72.9% protein-protein interface success rate, both matching AlphaFold 3 and ahead of Chai-1 at 68.5%. For antibody-antigen interfaces it reaches 37.6% (43.2% for the fine-tuned IntFold+ variant) against AlphaFold 3's 47.9%, and for protein-ligand interfaces 58.5% (61.8% for IntFold+) against 64.9%. On PoseBusters v2 targets deposited after the training cutoff it scores 76.1%, ahead of Protenix at 72.6%. Protein-DNA and protein-RNA interface success rates are 74.1% and 58.9% respectively, above Boltz-1 in both cases, and RNA monomer LDDT of 0.63 edges out AlphaFold 3's 0.61. Parameter counts and training compute are not disclosed.

#Applications

The controllable design targets structure-based drug discovery, where the useful structure is often a specific state rather than the most probable one. Medicinal chemists can push a kinase into the inhibitor-bound conformation relevant to a campaign; antibody engineers can supply a known epitope and obtain a docked complex that unconstrained folding gets wrong; screening workflows can rank compounds with the affinity module instead of exporting poses to a separate scoring tool. The hosted server serves users without GPU access, and the pip package supports local batch prediction.

#Impact

IntFold entered an already crowded field of open AlphaFold 3 reimplementations and differentiated itself on steerability rather than raw accuracy, an emphasis IntelliFold-2 continued with reported gains on antibody-antigen and protein-ligand co-folding. The Apache 2.0 license on both code and weights permits commercial use, a meaningful distinction from AlphaFold 3's restricted terms. One caveat: the public weight release covers general structure prediction, and the allosteric, constraint, and affinity adapters described in the technical report are not among the released checkpoints; the developers direct users to the hosted server for the full implementation.