---

About | Website

LiteFold is building the infrastructure for AI-driven drug discovery. Our work sits at the intersection of structural biology, machine learning, and reinforcement learning, with a focus on translating frontier ML methods into pipelines that pharma teams can actually use in production.

We work across the modality stack: small molecules, peptides (including non-canonical amino acids), and biologics. Our internal platform combines AI co-scientist workflows with rigorous computational chemistry, molecular dynamics, and structural prediction.

Why Open Science

The pace of progress in computational biology depends on shared benchmarks, open models, and reproducible pipelines. Drug discovery has historically been gated behind proprietary tooling and private datasets, and this has slowed down the entire field.

We do not believe everything should be open. We do believe far more should be open than currently is, and that frontier methods in biology benefit from the same culture of shared progress that has driven ML forward.

Our commitment on Hugging Face:

• Open models for protein language modeling, binding affinity prediction, and structural tasks
• Open benchmarks for evaluating language models and foundation models on real scientific reasoning
• Open datasets built around peptides, non-canonical amino acids, and curated structural data
• Open evals so the community can stress-test models honestly, including on their failure modes

What We Are Working On

Foundation models for biology

Protein language models trained with biology-aware tokenization, distilled structural models, and lightweight inference-friendly variants of larger systems.

Reinforcement learning environments

Verifiable RL environments for peptide design, small-molecule optimization, and structural reasoning. We want these to be useful both for frontier labs training scientific reasoners and for academic groups exploring new objectives in molecular design.

Benchmarks and evaluations

Tooling and benchmarks for measuring how language models behave on hard scientific questions, including sycophancy, false confidence, and silent failure modes in scientific tasks. If a model is going to do science, we should be able to measure when it is wrong.

Pipelines and tooling

End-to-end pipelines for virtual screening, MD-based stability, ADMET, proteolytic stability, and force field parameterization for non-canonical chemistry. Where we can open-source pieces of this, we will.

Get In Touch

If you are a researcher, lab, or company that wants to collaborate on open work, build on what we release, or contribute to the benchmarks we maintain, we would love to hear from you.

Reach out through litefold.ai.

---