Will AlphaFold 3 Earn a Nobel? Meet China’s HelixFold 3 Rival

AlphaFold 3: A Leap Forward in Biomolecular Modeling

In December 2020 AlphaFold 2 was released, enabling accurate 3‑D protein structure prediction within hours using deep learning, vastly accelerating biomedical research. However, it could not model interactions between proteins and other molecules.

In May 2024 DeepMind launched AlphaFold 3, which predicts structures and interactions for proteins, small‑molecule ligands, DNA, RNA and ions, delivering experimental‑level accuracy in minutes.

HelixFold 3: China’s First Domestic Model Matching AlphaFold 3

Although AlphaFold 3 is not open‑source and limits daily API calls, Baidu’s HelixFold 3, released in August 2024, reproduces its performance for protein‑ligand, nucleic‑acid and protein‑protein interactions.

HelixFold 3 expands AI‑driven design to challenging scenarios such as covalent binding, chemical modification, PROTACs, molecular glues and metalloenzymes.

High‑Performance Cloud Service

Baidu’s PaddleHelix team, together with Baidu Intelligent Cloud’s CHPC platform, provides an online high‑performance computing service for HelixFold 3, allowing researchers to run large‑scale structure predictions, screen candidate molecules, and accelerate drug discovery at low cost.

The service also supports APIs and no‑code workflows for small‑molecule, peptide/protein and mRNA drug projects.

Performance Comparison

Small‑Molecule Ligands

On the PoseBusters benchmark, HelixFold 3 achieves success rates and stereochemical quality comparable to AlphaFold 3, even without providing the true protein structure.

Nucleic Acids

Evaluated on RNA samples from CASP15 and 41 RNA and 41 DNA structures from the PDB, HelixFold 3 matches AlphaFold 3’s accuracy and surpasses specialized models such as RoseTTAFold2NA.

Proteins and Complexes

For protein‑protein complexes, HelixFold 3 slightly outperforms AlphaFold‑Multimer and approaches AlphaFold 3, though a performance gap remains.

Confidence Scores

HelixFold 3 provides confidence scores that correlate strongly with structural accuracy across ligand‑protein, protein‑protein, RNA and DNA predictions.