SimpleFold: Folding Proteins is Simpler than You Think: General Transformer Architecture Challenges Domain-Specific Design in Protein Folding
In September 2025, the Apple team proposed SimpleFold, the first flow-matching-based protein folding model, using only general Transformer modules without complex domain-specific designs such as triangular updates or explicit pairwise representations. The model has 3B parameters, trained on approximately 9 million distilled structures and PDB data, achieving performance comparable to state-of-the-art methods like AlphaFold on standard folding benchmarks, and performing better in ensemble predictions, with inference feasible on consumer-grade hardware.
SimpleFold demonstrates that protein folding does not require complex domain-specific architectures; general Transformer plus flow matching can achieve state-of-the-art results. This lowers the technical barrier for protein structure prediction, enabling more teams to participate and deploy, potentially accelerating drug discovery and synthetic biology. Additionally, its ensemble prediction capability outperforms deterministic models, which is significant for studying dynamic conformations.
SimpleFold uses standard Transformer blocks with adaptive layers, trained with a flow matching loss plus structural terms. The model has 3B parameters, trained on approximately 9 million distilled structures (from AlphaFold2, etc.) and experimental PDB data. On benchmarks like CASP, SimpleFold-3B performs comparably to AlphaFold2/3 but with faster inference, running on a single GPU. Its ensemble predictions (sampling multiple structures) are of higher quality than deterministic models, indicating that the flow matching framework is better suited for capturing multiple conformations. Ablation studies show that domain-specific modules are unnecessary; general architectures can learn folding rules through large-scale data and flow matching objectives.
Protein structure prediction is a core application of AI for Science. SimpleFold lowers the technical barrier in this field, enabling more small and medium-sized teams and companies to deploy high-performance folding models. This will accelerate applications such as target discovery, antibody design, and enzyme engineering, and may push protein design from the lab to industrial pipelines. Moreover, its efficient inference facilitates real-time structure prediction and virtual screening.
It is recommended that biotech and pharmaceutical companies evaluate SimpleFold as an alternative to AlphaFold for internal target structure prediction and virtual screening. Its low hardware requirements can reduce IT costs, and its ensemble prediction capability can improve the accuracy of candidate molecule screening. Consider collaborating with Apple or building customized folding services based on open-source code.
Attention should be paid to SimpleFold's performance on more complex multimers and membrane proteins; whether its flow matching framework can be extended to RNA, small molecules, etc.; the actual cost and speed of deployment on consumer-grade hardware; and whether the open-source community can reproduce and improve it. If the general architecture continues to be effective, it will trigger a paradigm shift in the protein folding field.