The archive · Bio & Materials · Product decision · 2024–2026
Ligo's open-source AlphaFold3 bet: Show HN 314 pts in 2024, repo at 1.1k stars
Biotech startup Ligo reimplemented AlphaFold3 from DeepMind's paper and open-sourced code plus training, single-chain first; 314 HN points, 1.1k stars.
Ligo Biosciences
What the business is
Ligo Biosciences is the team behind an open-source implementation of DeepMind's AlphaFold3: the full model and training code under Apache-2.0, with modular memory-efficient kernels, single-chain prediction shipped first and ligand/nucleic-acid beta predictions on a waitlist.
How it started
DeepMind's AlphaFold3 paper was published in Nature in May 2024, describing a model that predicts proteins, ligands, nucleic acids and their interactions — but no open code or training pipeline accompanied it. Ligo Biosciences built its own implementation from the published algorithms and supplementary pseudocode, reusing OpenFold modules for triangular attention and multiplicative updates, and launched the repository with a Show HN on 2024-09-04.
What happened
The first release shipped single-chain prediction with full training code; a demo video showed a model trained for 4,000 steps on 8 A100 GPUs in about 10 hours without templates. The README documented three discrepancies found against DeepMind's supplementary material — an MSA-module ordering issue, a loss-scaling factor that broke unit loss at initialization, and missing DiT residual connections — and asked the community to weigh in. Efficiency work focused on Triton kernels, including a custom MSA pair-averaging kernel that scales to thousands of tokens on one GPU, and a partnership with AdaptyvBio on a ProteinFlow-based data pipeline. Ligand, multimer and nucleic-acid prediction were listed as the next release, with beta testers recruited through a waitlist.
How it ended up
Still running as an active research project as of 2026-09-05: the crawl recorded 301 commits and 1.1k stars, single-chain prediction is released, and ligand, multimer and nucleic-acid prediction sit on a beta waitlist; no funding, revenue or customer numbers appear in the repo record.
Background
DeepMind's AlphaFold3 paper was published in Nature in May 2024, describing a model that predicts proteins, ligands, nucleic acids and their interactions. Ligo Biosciences built its own open-source implementation from the published algorithms and supplementary pseudocode, reusing battle-tested OpenFold modules and partnering with AdaptyvBio on a ProteinFlow-based data pipeline, and released it under Apache-2.0.
The bet had two parts: that a small team could reproduce AlphaFold3 faithfully enough to train, and that releasing code plus the training loop — not just weights or an API — would make Ligo the default open stack for biomolecular structure prediction. The first release shipped single-chain prediction with full training code; a demo video showed a model trained for 4,000 steps on 8 A100 GPUs in about 10 hours without templates. Ligand, multimer and nucleic-acid capabilities were promised for later releases, with beta testers recruited through a waitlist.
Ligo also published the discrepancies it found against DeepMind's supplementary information — an MSA-module ordering issue, a loss-scaling factor inconsistent with the Karras diffusion framework, and missing residual connections in the DiT blocks — and asked the community to help resolve the ambiguities. Efficiency work focused on Triton kernels, including a custom MSA pair-averaging kernel that scales to thousands of tokens on a single GPU.
The Show HN on 2024-09-04 drew 314 points and 37 comments, and the repository kept shipping: 301 commits by the 2026-09-04 crawl. Two years on, the project remains an early, explicitly not-production-ready research release; the record shows active development and clear adoption signals, but no funding, revenue or customer disclosures.
What has to be true
- Named-platform traction with a date: the Show HN on 2024-09-04 (item 41448439) drew 314 points and 37 comments.
- A concrete open bet: reimplement AlphaFold3 from the paper with training code, ship single-chain first, and train the harder ligand/nucleic-acid modalities behind a beta waitlist.
- Verifiable trajectory: the repo recorded 301 commits, 94 forks and 1.1k stars at the 2026-09-04 crawl, with single-chain prediction released.
- The discrepancies section is the tell: three ambiguities in DeepMind's supplementary material were published instead of silently resolved — a research-led adoption strategy, not a demo.
What can be applied
Open-sourcing the training loop, not just a demo, turns a reimplementation into community infrastructure — ship the slice compute can reach first and say plainly what is still missing.
Aftermath
As of 2026-09-05 Ligo's AlphaFold3 implementation is live as an active research project: Apache-2.0 code and the training loop for single-chain protein prediction are in the repository, and ligand, multimer and nucleic-acid prediction are listed as next-release work behind a beta-testing waitlist. The README still warns the tool is not production-ready. The 2026-09-04 crawl recorded 1.1k stars, 94 forks and 301 commits; the record contains no funding, revenue or team disclosures, so the company-building half of the bet remains unproven.
Sources
- Show HN: An open-source implementation of AlphaFold3 (314 points, 37 comments)
- Ligo-Biosciences/AlphaFold3 — open-source implementation (1.1k stars, 301 commits)
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card