Jiarui Lu

So I understand that was unexpected for a lot of people, @Apple MLR has released a protein folding model! arxiv.org/abs/2509.18480. Here’s a summary of what SimpleFold is and what it represents: - What is SimpleFold? A generative model that essentially treats protein folding almost exactly as if it were a text-to-image or text-to-3D problem. - What are we sharing? A research paper and a codebase under an MIT license github.com/apple/ml-simplefo… (looking forward to people contributing to it!). We are also releasing pre-trained checkpoints of different sizes so that researchers can best tradeoff performance for efficiency. - Why protein folding? We are doing this work largely because protein folding is an excellent benchmark for structured data generation and multi-modality. Protein folding is a very interesting problem from a generative modeling perspective and we do research on generative modeling :) - Why is it interesting? IMO SimpleFold is interesting because I believe in finding recipes (architectures, training objectives, etc.) that generalize across the board to many different data modalities. Let’s say you are an ML expert in text-to-image or text-to-3D, now you can apply your latest and greatest architectural blocks or efficient samplers to protein folding with SimpleFold. I believe this is a net benefit for ML research and science in general. Now getting more into the technical details: - Our architecture is very simple (hence the name), just a stack of transformer blocks with time-step conditioning. This is important because it makes the model efficient at inference time. You can run SimpleFold directly on your Mac and get results quickly without data ever leaving your laptop. - SimpleFold is not necessarily a model that “rejects” inductive biases, it just doesn’t enforce them directly on the architecture. For example, we apply rotation augmentation to all the protein structures during training. This makes the model “softly” invariant to this symmetry. - There were some concerns online about data leakage from AFESM and that driving performance of SimpleFold or making it overfit. We filtered AFESM data so that the CASP14 sequences are not seen during training. As a matter of fact we distilled structures from AF2/ESMFold models, which have the same cutoff data as SimpleFold for PDB data. Both AF2 and ESMFold train on self-distilled datasets, we just train SimpleFold on a bigger set of distilled data. I want to thank my awesome team of collaborators, they are all rockstars. That’s all, for now :)