- OpenFold3, released in October 2025, is an open-source co-folding model predicting protein interactions with small molecules, antibodies, and nucleic acids.
- Generative diffusion models now produce novel protein structures with validated binding properties, shifting the field from prediction to design.
- The LIGAND-AI partnership, announced January 2026, aims to build open datasets for AI drug candidate proposals by 2035.
Protein structure prediction has quietly outgrown its own origin story. AlphaFold2 solved the single-protein folding problem; what’s happening now is harder and more consequential, AI models predicting how proteins grab onto drugs, assemble into complexes and interact with DNA in real time. The October 2025 release of OpenFold 3 and a wave of generative design tools are pushing the question from “what does this protein look like?” toward “what protein should we build?”
Modeling the Messy Reality of Biology
Proteins almost never act alone. They form multi-protein assemblies, dock to small molecules that trigger or suppress their activity, and interact with nucleic acids to regulate gene expression. AlphaFold2’s landmark accuracy applied mainly to single proteins in isolation, a genuine breakthrough, but one that left the more biologically realistic problem largely unsolved.
DeepMind‘s AlphaFold3, announced in May 2024, extended the scope to proteins, DNA, RNA and small molecules together, and according to the company’s benchmarks outperformed prior techniques for protein interaction modeling by at least 50%, though independent validation of that figure is ongoing. The Baker lab’s RoseTTAFold All-Atom (RFAA) and RFdiffusion All-Atom take a similar multi-molecule approach, adding metals and covalent modifications to the modelling scope. OpenFold3, released by the OpenFold Consortium in October 2025, brings this capability fully into the open: trained on more than 300,000 publicly available structures and a curated synthetic database of over 13 million entries, it predicts 3D structures of proteins interacting with small-molecule drugs, antibodies, other proteins and nucleic acids, with open code, open weights and open workflows.
Designing Proteins That Don’t Exist Yet
Prediction and design are different problems. Predicting a structure means finding the shape a known sequence folds into. Design means starting from a desired function and engineering a sequence, and a shape, that delivers it. That second problem, historically intractable, is where diffusion models are making real headway.
RFdiffusion, from the Baker lab at the University of Washington, marries the structure-prediction backbone of RoseTTAFold with the generative mechanics of diffusion modelling. The result can produce high-affinity protein binders and symmetric protein assemblies that have been validated in vitromeaning they work in the lab, not just on paper. EigenFold takes a complementary angle: rather than producing a single predicted structure, it samples a distribution of conformations from a given sequence, capturing the flexibility that matters enormously for understanding how a protein actually behaves in a cell. As our earlier coverage of OpenFold3 noted, the line between prediction and design is blurring faster than most researchers anticipated.
RoseTTAFold-3 (RF3), introduced in February 2026, refines this further. Trained using AtomWorks infrastructure, RF3 builds chirality signals, the geometric handedness of molecules that determines whether a drug works or causes harm, directly into its diffusion denoising process. It also adds atom-level distance conditioning, giving researchers more direct control over the output structure. The stated goal is to match AlphaFold3’s performance across diverse biomolecular complexes while remaining openly accessible.
Drug Discovery Gets a Faster Lane
Knowing a protein’s structure in complex with a drug candidate is the foundation of structure-based drug discovery. The better that prediction is, the faster a team can narrow from thousands of candidate molecules to a handful worth testing in the lab. AI is now improving nearly every step in that pipeline: binding site identification, pose estimation, scoring and virtual screening.
AI-Bind 2.0, presented in June 2025, uses pre-trained protein language models including ProtTrans to predict protein-ligand binding with higher accuracy than its predecessors, validated against curated data from DrugBank and BindingDB. The tool was applied in COVID-19 research to identify promising therapeutic candidates.
LIGAND-AI aims to build open, high-quality datasets of protein-ligand interactions large enough to train models that can propose drug-like binders for thousands of human proteins.
There’s also PIONEER, a tool developed by Cleveland Clinic and Cornell University that predicts how single DNA mutations alter protein-protein interaction networks. In lab validation, PIONEER’s predictions helped identify candidate drug targets across dozens of cancers and complex diseases, a different angle on the same underlying problem of connecting molecular structure to therapeutic opportunity.
Open Source and the Hardware Ceiling
Access matters as much as capability. OpenFold3’s fully open release, code, weights and workflows, means research groups without DeepMind-scale compute can run co-folding predictions on their own infrastructure, and adapt the model for their specific needs. That openness is not incidental to the science; it’s what allows independent validation and community-driven improvement.
Benchmarks reported in early 2026 suggest that AMD Instinct MI355X GPUs delivered up to 45.5% faster inference for OpenFold3 than NVIDIA B200 at a sequence length of 5,000, and could handle sequences up to 6,000 tokens long. The source and methodology of these benchmarks have not been independently verified, so the figures should be treated as directional rather than definitive.
