OpenProtein.AI launches open-source protein design tools

OpenProtein.AI launched a suite of open-source protein-design models, training datasets and researcher-facing tools on April 17, 2026. The startup, founded by Tristan Bepler PhD ’20 and former MIT professor Tim Lu PhD ’07, published models, code and a browser interface so scientists can run design workflows locally or in cloud environments.

The release bundles model weights, training scripts, curated datasets and example pipelines on public repositories, alongside a web-based interface for interactive design and an API for programmatic access. The company positions the package to lower the technical barrier for labs that want to apply machine learning to enzyme engineering, binding design and other protein engineering tasks without building large internal ML stacks.

What OpenProtein.AI released and how researchers can use it

The core deliverables are: pre-trained protein-design models, curated sequence and structural datasets, training code and an online design app. The models are provided in formats compatible with common ML frameworks so they can be fine-tuned or run inference on standard GPUs. Documentation and example notebooks show end-to-end workflows from target specification to candidate sequence output.

The browser interface exposes a workflow for specifying targets, constraints and optimization objectives, and for visualizing predicted structures and confidence scores. Users can run quick in-browser queries for screening tasks, or export model inputs and outputs to run larger jobs on institutional clusters or commercial cloud instances.

OpenProtein.AI also published datasets intended to support training and evaluation, including labeled sequence-function pairs and structural templates aggregated from public sources. The release includes evaluation scripts so researchers can reproduce and extend baseline benchmarks and test new model variants.

Access options include cloning the public repositories, pulling model weights, using the HTTP API or running the browser UI. The documentation recommends hardware profiles for common tasks, and example pipelines show how to combine the released models with experimental feedback for iterative design.

The founders emphasize openness and reproducibility. Tristan Bepler brings expertise in ML for biological sequences, while Tim Lu contributes a background in synthetic biology and lab-scale implementation. The project aims to support both computational groups that will extend model architectures and experimental groups that will use models to generate candidates for laboratory testing.

Why it matters

Making pre-trained models, datasets and a user-facing app available reduces setup time for labs that lack large ML teams, potentially accelerating the path from computational design to experimental validation. Broader access also means more groups can test models on diverse protein classes, which will reveal strengths and failure modes sooner. Regulators, funders and labs that plan to deploy AI-generated sequences will need to consider validation, safety and reproducibility as usage spreads.