IntelliFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction

⚡ IntelliFold Server • 📄 IntelliFold 2 Release Note • 📄 IntelliFold Technical Report

📰 News

• 2026-06-15: IntelliFold-v2 now runs on the AlphaFold 3 JAX engine. AlphaFold 3's JAX inference is fast, and as of v3.0.3 it is released under the Apache-2.0 license — so we leverage their repository to accelerate IntelliFold-2 inference. See Usage to install and run.
• 2026-02-07: We are excited to present [IntelliFold 2]. This version represents a major architectural update and is one of the first open-source models to outperform AlphaFold3 on Foldbench.

📊 Benchmarking

To comprehensively evaluate the performance of IntelliFold 2, we conducted a rigorous evaluation on FoldBench. We compared IntelliFold against several leading methods, including Boltz-1,2, Chai-1, Protenix and Alphafold3.

For more details on the benchmarking process and results, please refer to our release note IntelliFold 2 Release Note and IntelliFold Technical Report.

🔍 Usage

Setup

If your environment already has alphafold3 installed, just add this wrapper from PyPI:

pip install intellifold

Otherwise install the vendored AlphaFold 3 engine first. Clone the repo, build the engine and its CCD data with build_data (reads libcifpp's bundled components.cif — no network, ~30s), then install the wrapper:

git clone https://github.com/IntelliGen-AI/IntelliFold
cd IntelliFold
pip install ./third_party/alphafold3 && build_data
pip install intellifold        # or `pip install .` from the clone

jax[cuda12] ships its own CUDA libraries and uses the GPU out of the box. If it instead falls back to CPU (cuSPARSE ... not found), a system CUDA on your LD_LIBRARY_PATH is shadowing the bundled ones — put the bundled libraries first:

export LD_LIBRARY_PATH=$(ls -d "$(python -c 'import nvidia,os;print(os.path.dirname(nvidia.__file__))')"/*/lib | paste -sd:):$LD_LIBRARY_PATH

Inference

The input is an AlphaFold 3-style JSON. On the first run the pre-converted IntelliFold-v2 weights are downloaded from Hugging Face into ./model_v2 (later runs just load them).

Quick start — the example fold_input.json already contains its MSAs and templates, so skip the data pipeline with --norun_data_pipeline:

wget https://huggingface.co/intelligenAI/intellifold/resolve/main/fold_input.json
intellifold predict fold_input.json --model-dir=model_v2 --output-dir results -- --norun_data_pipeline

Search MSAs yourself — if your JSON has no MSAs, first download the sequence databases, then point --db_dir at them (AF3 runs its data pipeline to build MSAs/templates):

bash third_party/alphafold3/fetch_databases.sh /path/to/databases
intellifold predict fold_input.json --model-dir=model_v2 --output-dir results -- --db_dir=/path/to/databases

Batch a directory across every GPU — predicts every JSON in the directory (one worker per GPU, self-balancing local queue, resumable — rerun to continue):

intellifold predict ./my_inputs/ --model-dir=model_v2 --gpus all --output-dir results -- --norun_data_pipeline

Steer toward physically valid poses (optional, off by default) — add --steering to nudge the diffusion sampler at every denoising step with the gradient of a set of differentiable physical/chemical potentials (bond lengths & angles, internal clashes, chirality, double-bond / ring planarity, protein–ligand van-der-Waals overlap, covalent connections, and symmetric-chain repulsion); the potential set is adapted from Boltz. This improves the physical plausibility of predicted small-molecule poses without retraining — on our PoseBusters-v2 evaluation it raises the PoseBusters validity rate from ~72% to ~93% (and the joint "valid and pocket-aligned ligand RMSD < 2 Å" rate from ~62% to ~75%) while leaving docking accuracy essentially unchanged. Without --steering the sampler is unchanged.

intellifold predict fold_input.json --model-dir=model_v2 --output-dir results \
    -- --norun_data_pipeline --steering

Steering tuning flags (also after --): --steering_num_gd_steps (default 20, gradient-descent iterations per denoising step) and --steering_weight_scale (default 1.0, global multiplier on all potential weights).

⚠️ Steering is slower. It runs num_gd_steps extra gradient evaluations inside every denoising step, and because the per-target constraint set has a target-specific shape, each input triggers its own XLA compilation (the no-steering path reuses bucketed compilations across inputs). Expect a noticeably longer wall-clock per target — only enable it when you want the improved physical validity; for large batches keep it off unless you specifically need PoseBusters-clean poses.

-- passes everything after it straight through to AlphaFold 3 (its own flags) — e.g. --norun_data_pipeline, --db_dir=/path/to/databases, --num_diffusion_samples=5, --steering. Set HF_ENDPOINT (e.g. hf-mirror.com) for a download mirror.

How the IntelliFold JAX weights were produced

The hosted intellifold_v2.bin.zst + intellifold_v2_fourier.npz were converted from the IntelliFold-v2 PyTorch checkpoint with convert_ifv2_to_jax.py. To convert your own .pt (needs torch, pip install '.[convert]'):

python convert_ifv2_to_jax.py --schema intellifold/af3_schema.pkl --v2-pt intellifold_v2.pt --out-dir model_v2

🌐 IntelliFold Server

We highly recommend using the IntelliFold Server for the most accurate, complete, and convenient biomolecular structure predictions. It requires no installation and provides an intuitive web interface to submit your sequences and visualize results directly in your browser. The server runs the full, optimized, latest IntelliFold implementation for optimal performance.

📜 Citation

If you use IntelliFold in your research, please cite our paper:

@article{qiao2026intellifold,
  title={IntelliFold-2: Surpassing AlphaFold 3 via Architectural Refinement and Structural Consistency},
  author={Qiao, Lifeng and Yan, He and Liu, Gary and Guo, Gaoxing and Sun, Siqi},
  journal={bioRxiv},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}

@article{team2025intfold,
  title={IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction},
  author={Team, The IntFold and Qiao, Leon and Bai, Wayne and Yan, He and Liu, Gary and Xi, Nova and Zhang, Xiang and Sun, Siqi},
  journal={arXiv preprint arXiv:2507.02025},
  year={2025}
}

🔗 Acknowledgements

• This repository runs IntelliFold-v2 on the AlphaFold 3 JAX inference engine by Google DeepMind (Apache-2.0), vendored at third_party/alphafold3/ (v3.0.3). The wrapper's run_jax_inference.py is a modified copy of AF3's; see NOTICE.
• The implementation of fast layernorm operators is inspired by OneFlow and FastFold, following Protenix's usage.
• Many components in intellifold/openfold/ are adapted from OpenFold, with substantial modifications and improvements by our team (except for the LayerNorm part).
• This repository, the implementation of Inference Data Pipeline(Data/Feature Processing and MSA generation tasks) referred to Boltz-1, and modify some codes to adapt to the input of our model.
  • The template pipeline implementation in the Inference Data Pipeline of this repository refers to Protenix, with additional adjustments and modifications to fit our model.
• The optional diffusion steering (physical guidance) feature adapts the inference-time potential set from Boltz, reimplemented for the AlphaFold 3 JAX sampler.

⚖️ License

The IntelliFold project, including code and model parameters, is made available under the Apache 2.0 License, it is free for both academic research and commercial use.

This repo vendors the Apache-2.0 AlphaFold 3 (v3.0.3+) in third_party/alphafold3/ — not the earlier CC BY-NC-SA (non-commercial) v3.0.1/v3.0.2. The AF3 model weights are not included and are non-commercial; this repo ships only IntelliGen-AI's own IntelliFold-v2 weights. See NOTICE.

📬 Contact Us

If you have any questions or are interested in collaboration, please feel free to contact us at contact@intfold.com.