# BioTorch opens public beta with 116 PyTorch exercises

> Chris Hayduk’s side project starts with 17 model guides alongside its coding exercises. The workspace is free during beta and connects model components to AlphaFold2, ESM2 and RFdiffusion.

- Publication: Go Big News
- By Go Big News Staff (Staff News)
- Category: AI
- Published: 2026-09-07T12:40:19.898+00:00
- Canonical: https://www.gobignews.com/ai/biotorch-opens-public-beta-with-116-pytorch-exercises

---

[BioTorch](https://biotorch.chrishayduk.com) entered public beta on September 5, 2026, with 116 PyTorch exercises and 17 model guides. Chris Hayduk announced the launch in a [post on X](https://x.com/chrishayduk/status/2096371875009933817), describing BioTorch as a side project intended to give researchers a clearer route into biological AI.

The workspace lets users implement model operations with raw PyTorch tensors, test their code and see where the components appear in AlphaFold2, ESM2 and RFdiffusion. Hayduk said the goal is to help researchers understand influential biological AI papers and build on their work. BioTorch is free during the beta.

Hayduk said the path into biological AI is harder to assemble than the path into large language model research. He linked that friction to fewer researchers entering the field, fewer startups being built and less investment reaching promising ideas.

BioTorch groups its exercises under attention, protein geometry, diffusion, sequence models and genomics. Submitted code runs in an isolated environment, where it is checked for numerical values, shapes and gradients. The model guides show where each operation appears within biological AI architectures.

The named systems cover protein structure, language modeling and design. The 2021 AlphaFold2 paper reported accuracy competitive with experimental protein structures in most CASP14 cases. Meta described ESM-2 as a protein language model trained to predict masked amino acids, while RFdiffusion was produced by fine-tuning RoseTTAFold for protein-structure denoising.

BioTorch says its problems cover model components rather than complete reproductions. Its privacy page also says no affiliation with or endorsement by the named models is implied.

Accounts can be created with an email and password, Google or GitHub, with Auth0 handling sign-in. BioTorch also schedules FSRS reviews from recorded attempts and accepted submissions to estimate when a user should practice again. FSRS stands for Free Spaced Repetition Scheduler and models difficulty, stability and retrievability.

Hayduk is a Forward Deployed Engineer on OpenAI’s Life Sciences team, working with pharmaceutical and biotechnology organizations on drug discovery and clinical development. He previously worked as a machine learning engineer at Meta and as lead machine learning engineer for drug discovery at Deloitte.

---

## X profiles in this story

- @chrishayduk: https://www.gobignews.com/profile/chrishayduk