Use IntFold Online

Commercially Available IntFold No-Code Web Server

IntFold: Controllable Foundation Model for Biomolecular Structure Prediction

IntFold is a controllable foundation model for general and specialized biomolecular structure prediction that reaches accuracy comparable to AlphaFold 3 on the FoldBench benchmark. It was developed by the IntFold Team at IntelliGen AI (Leon Qiao, Wayne Bai, He Yan, Gary Liu, Nova Xi, Xiang Zhang and Siqi Sun) and released as an arXiv preprint in July 2025. The authors' code is hosted at github.com/IntelliGen-AI/IntFold; the paper does not state a license.

All-atom structure predictors such as AlphaFold 3 can model proteins, nucleic acids and small molecules together, but a single general model is rarely tuned for specialized questions such as alternative conformations, user-defined constraints or binding affinity. IntFold addresses this by pairing an AlphaFold 3-style backbone with lightweight adapters.

Property

Detail

Task

All-atom biomolecular structure prediction with optional steering

Inputs

Sequences, with optional MSAs and templates

Architecture

Embedding trunk, diffusion module, confidence head (pLDDT, pTM) and optional adapters

Adapters

Allosteric-state guidance, structural constraints, binding affinity

Ranking

Training-free: 25 predictions (5 samples across 5 seeds) ranked by mean all-vs-all DockQ

How IntFold Works: Attention Kernel, Adapters and Ranking

  1. Custom attention kernel: FlashAttentionPairBias, written in Triton and based on FlashAttention 2, broadcasts the pair bias on the fly to reduce memory use.

  2. Per-layer LoRA adapters: steer predictions toward specific 3D states, such as CDK2 allosteric conformations, and a constraint embedder handles epitope or pocket constraints.

  3. Affinity module: four additional Pairformer blocks on top of the trunk score protein-ligand binding affinity.

  4. Similarity-based ranking: the predicted structure with the highest mean DockQ against the other predictions is selected, with no extra training.

IntFold+ for antibody-antigen complexes

IntFold+ is fine-tuned from the base model with added antibody-antigen data, no MSA pairing for these complexes, and pLDDT-based masking.

IntFold Benchmark Results on FoldBench and PoseBusters

On FoldBench, the paper reports success rates by interaction type, citing AlphaFold 3 numbers from its original benchmark.

FoldBench category

IntFold

AlphaFold 3

Protein-protein

72.9%

72.9%

Antibody-antigen

37.6% (IntFold+ 43.2%)

47.9%

Protein-ligand

58.5% (IntFold+ 61.8%)

64.9%

Protein-DNA

74.1%

79.18%

Protein-RNA

58.9%

62.3%

On a post-cutoff PoseBusters v2 subset, IntFold reached 76.1% versus 72.6% for Protenix. In a comparison with Boltz-2 on FoldBench targets released after January 2024, IntFold scored 40.27 versus 25.00 on antibody-antigen and 58.17 versus 53.90 on protein-ligand. For CDK2, the general model predicted none of 5 allosteric states while the fine-tuned adapter model got 4 of 5. Structural constraints raised PoseBusters success from 79.5% to 89.7% and antibody-antigen success from 37.6% to 69.0%.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

  • A user-friendly graphical interface for setting up and launching experiments

  • A robust API for integration into existing research pipelines

  • An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with IntFold on Tamarind Bio

  • General complex prediction: model proteins, DNA, RNA and ligands with accuracy the authors report as comparable to AlphaFold 3.

  • Antibody-antigen modeling: use epitope constraints or IntFold+ to improve antibody-antigen predictions.

  • Conformational states: steer predictions toward specific allosteric states, as shown for CDK2.

  • Binding affinity estimates: score protein-ligand binding with the affinity adapter, where the paper reports a Pearson correlation of 0.53 on CASP16 L1000.

How to Use IntFold on Tamarind Bio

  1. Log in and open the tool: sign in at tamarind.bio and select IntFold.

  2. Provide sequences: enter the protein, DNA or RNA sequences, and ligands if supported, for the system you want to model.

  3. Choose the mode: select the base model or an adapter such as constraints or IntFold+ for antibody-antigen targets, if exposed.

  4. Add constraints: provide epitope or pocket constraints if the option is exposed.

  5. Run the job: submit and wait for the structure predictions to complete.

  6. Download outputs: retrieve the ranked structures and confidence scores (pLDDT, pTM).

  7. Interpret confidence: favor top-ranked models and treat antibody-antigen results with extra care.

  8. Use downstream: visualize, dock, or use the models to guide design and experiments.

Things to Keep in Mind

  • Triangle attention scales roughly cubically with size, which limits training crop size and speed on very large assemblies.

  • Accuracy on challenging targets, especially antibody-antigen complexes, still has room to improve.

  • IntFold stays below AlphaFold 3 on several FoldBench categories, such as protein-ligand and protein-DNA.

  • The paper does not state a license, and a detailed ablation of IntFold+ is left for future work.

Source: The IntFold Team, Leon Qiao, Wayne Bai, He Yan, Gary Liu, Nova Xi, Xiang Zhang and Siqi Sun, "IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction", arXiv:2507.02025 (2025), https://arxiv.org/pdf/2507.02025. Code: https://github.com/IntelliGen-AI/IntFold.

Supporting 10,000+ scientists around the world,

from leading biotechs, and global biopharma