Wet-lab validationPredicted structures must be synthesized and tested experimentally; computational speed does not reduce the time and cost of physical validation.
Protein manufacturabilityDesigned proteins must be producible at scale; thermostability, expression yield, and purification are not yet reliably predicted from sequence alone.
Physics limits
Protein folding is deterministic but complexPredicting how a protein folds requires accounting for interactions across hundreds of thousands of atoms; neural networks approach but do not exceed experimental methods in all cases.
Binding affinity depends on contextBinding strength varies with pH, salt concentration, temperature, and cellular environment; models trained on one context may not generalize to others.
How it works
4 parts
Input biomolecules
Protein and ligand sequences
Scientists input protein sequences, small-molecule structures, and other biomolecular data they want to study.
Structure prediction
Predict 3D arrangement
Neural networks predict how proteins fold and bind molecules.
Binding affinity
Estimate binding strength
Boltz-2 additionally predicts how tightly molecules bind, a key parameter for drug efficacy and selectivity.
Generative design
Design new molecules
BoltzGen uses diffusion models to sample novel protein and small-molecule designs that bind specified targets, enabling rational therapeutic engineering.
Open biomolecular AI, structure prediction, protein design
Boltz develops open-source AI models for biomolecular design. In January 2026, Boltz raised $28M to expand partnerships with major pharmaceutical companies.
Partnerships with Pfizer, Takeda, GSK, and other leading pharma companies validate the platform's utility.
Boltz integrates open-source structure prediction and binding affinity models with generative AI workflows for drug discovery.
The platform is used by academic and commercial researchers for protein design and molecular screening.
