BoltzGen

BoltzGen extends the Boltz family from structure prediction into generative design, producing candidate binders against a specified target. It moves the open-source stack from answering “what shape is this” to “design me something that sticks to it” — the step that actually matters for therapeutic and reagent development.

What it does

Where a prediction model evaluates a given sequence, BoltzGen proposes new ones, generating binder candidates that can then be filtered, ranked and passed to synthesis. Combined with Boltz for structural validation, it forms an open design-and-check loop that previously required commercial platforms or an in-house model team.

Strengths

• Open source, so binder design is no longer gated behind a platform contract.
• Pairs directly with Boltz for a generate-then-validate loop.
• Fast enough to produce candidate sets rather than single designs.
• Removes a major cost barrier for small groups and academic labs.

Limitations

Generated binders are hypotheses, not molecules — the wet-lab hit rate is the only number that matters and it remains modest across every method in this class. Expect to synthesise and screen. GPU requirements are significant, and the field is moving fast enough that benchmarks age within months.