Protein Structure Prediction and De Novo Protein Design

Google DeepMind and Nvidia are major corporate players; startup names include Generate Biomedicines Inc, A Alpha Bio Inc., and Profluent Bio Inc. Google DeepMind is associated with deep-learning structure prediction, while Generate Biomedicines and Profluent develop generative protein-design methods. A Alpha Bio launched Atlas, a platform for protein-interaction data.

Generate Biomedicines Inc, A Alpha Bio Inc., and Profluent Bio Inc. are identified startups in this field. Generate Biomedicines has patented a diffusion-model system for de novo protein generation. A Alpha Bio launched Atlas in July 2026 to provide protein-interaction data, while Profluent develops AI methods for protein and antibody sequence design.

The University of Washington, University of California, and Massachusetts Institute of Technology are prominent research organizations in the field. Recent work also comes from Shanghai JiaoTong University and Zhejiang University. Their reported directions include generative design, sequence design, protein binders, inverse folding, and joint optimization of protein structure and sequence.

Recorded grant funding since 2020 totals $173,599,290 across 210 grants. This figure describes grant records, not a complete measure of investment in the field. The available funding-round data also includes entries that may not represent protein-design-specific financing, so it should not be treated as an equivalent total.

Genentech leads the listed patent holders with 10 patents, followed by Hoffmann-La Roche with 6. Iowa State University Research Foundation holds 4. The University of Washington, Fujitsu Limited, Nvidia, Novozymes AS, and The Regents of The University Of California each have 3 listed patents.

Patent activity is down: 142 patents were recorded in the 24-month window, with growth of -52.1%. This contrasts with publication activity, which reached 4,048 publications in the last year and grew 46.6%. The difference suggests research output is expanding faster than patenting in the measured period.

Structure prediction is an established research capability, while designing proteins that work reliably remains an active technical challenge. The field recorded 4,048 publications in the last year, up 46.6%, alongside ongoing work on function, dynamics and experimental validation. Recent design studies target practical outputs such as binders, enzymes and imaging proteins.

Recent developments include HighMorph, reported by Zhejiang University of Technology in September 2026, for interaction-guided cyclic-peptide sequence design. Shanghai JiaoTong University researchers reported a de novo binder for [NiFe]-hydrogenase that same month. In September 2026, NVIDIA also announced integration of its BioNeMo Agent Toolkit with Anthropic’s Claude Science workbench.

The main approaches are Template-based modeling, Physics-based folding, Deep-learning structure prediction, Generative design, and Sequence-based design. Template-based methods use related structures, while physics-based methods score conformations using energy models. Deep learning predicts structure from learned patterns; generative methods create candidates, and sequence-based design seeks sequences compatible with a target structure.

Documented application areas include therapeutics and drug discovery (2,093 documents), antibodies and molecular binders (1,189), industrial enzymes (754), synthetic biology and agriculture (765), and structural biology (967). These methods can help develop or study molecules with targeted properties, but proposed designs need experimental testing to establish their activity and performance.