HelixFold-Multimer: High‑Performance Antigen‑Antibody and Peptide‑Protein Complex Structure Prediction

In the field of large‑molecule drug discovery, antibody and peptide therapeutics have attracted increasing attention due to their high specificity and excellent safety profiles.

Accurate prediction of antigen‑antibody and peptide‑protein complex structures is crucial for guiding downstream research, yet current computational methods—including physics‑based energy functions and deep neural network approaches such as AlphaFold 3—still leave considerable room for improvement.

To address this challenge, the Baidu PaddleHelix team developed the HelixFold‑Multimer model. By optimizing data, network architecture, and training strategies, HelixFold‑Multimer achieves industry‑leading performance on both antigen‑antibody and peptide‑protein complex prediction tasks, surpassing existing methods, including the recently released AlphaFold 3, on key metrics such as DockQ and success rate.

Antigen‑Antibody Prediction Results

Using 141 antigen‑antibody complexes released between 2023‑01‑25 and 2023‑08‑09 (130 regular antibodies and 11 scFv), HelixFold‑Multimer attained a mean DockQ of 0.41, median 0.38, and a success rate of 58.8%, compared with AlphaFold 3’s mean DockQ of 0.29, median 0.07, and success rate of 38.9%.

When epitope information from experimental methods (e.g., Deep Mutation Scanning) is incorporated, HelixFold‑Multimer (with epitope info) further improves to a mean DockQ of 0.49, median 0.54, and success rate of 77.1%, nearly doubling AlphaFold 3’s performance.

A case study on the HIV‑targeting antibody 8saq (binding the V3 glycan of HIV) shows HelixFold‑Multimer achieving DockQ = 0.81 and Interface RMSD < 1 Å, whereas AlphaFold 3 produced a DockQ of only 0.03.

Peptide‑Protein Prediction Results

Testing on 61 peptide‑protein complexes released from the PDB between 2022‑01‑16 and 2022‑12‑12, HelixFold‑Multimer achieved a mean DockQ of 0.378 (median 0.295) and a success rate of 68.9%, outperforming AlphaFold 3’s mean DockQ of 0.359 (median 0.260) and success rate of 59.0%.

HelixFold‑Multimer is the latest Baidu Wenxin large‑model family work in computational biology. The model has already been applied in large‑molecule drug pipelines, identifying nanomolar‑level active candidates for two targets. The technical report is available on arXiv, and an online service is open for public testing.

Online demo links: • Peptide‑protein prediction: https://paddlehelix.baidu.com/app/drug/protein-complex/forecast • Antigen‑antibody prediction: https://paddlehelix.baidu.com/app/drug/KYKT/forecast

Technical report: https://arxiv.org/abs/2404.10260v2

Contact: [email protected]