When DeepMind released AlphaFold2 in 2020, it solved a 50-year-old problem in biology: predicting how a protein folds from its amino acid sequence alone. The achievement won Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. But solving structure prediction turned out to be the beginning, not the end, of a much larger revolution.
The question that matters for drug discovery was never just "what shape is this protein?" It was "how tightly does this drug bind to it?" And that question — binding affinity prediction — remained stubbornly out of reach until mid-2025.
The AlphaFold3 Moment
AlphaFold3, released by Google DeepMind in May 2024, represented a fundamental architectural shift. Where AlphaFold2 was a protein-only model, AF3 extended predictions to include DNA, RNA, small molecules, ions, and post-translational modifications. It used a diffusion-based architecture rather than the structure module of its predecessor, allowing it to model the full complexity of biomolecular assemblies.
The model's performance on protein-ligand complexes was impressive but imperfect. Independent benchmarks showed that AF3 could accurately predict binding poses for many drug-like molecules, but its confidence scores didn't reliably correlate with binding strength. You could predict where a drug sits in a protein's binding pocket, but not how tightly it grips.
More importantly, Google initially restricted access. The AlphaFold Server allowed limited predictions via a web interface, but the full model weights and training code were not released. For an academic field built on open science, this created tension. For commercial drug discovery, it created an opportunity.
The Open-Source Response
That opportunity was seized by a wave of open-source alternatives. Chai-1, from Chai Discovery, was among the first to match AF3's performance on protein-ligand structure prediction while releasing full model weights. Boltz-1, from a team led by researchers at MIT, followed with a fully open implementation that could be run locally on standard GPU hardware.
But the real breakthrough came with Boltz-2, released in June 2025. Boltz-2 doesn't just predict structure. It simultaneously predicts binding affinity — outputting both a 3D protein-ligand complex and a quantitative estimate of how strongly the molecule binds, all in approximately 20 seconds on a single GPU.
The technical innovation behind Boltz-2 is a unified architecture that treats structure prediction and affinity estimation as coupled tasks. The model learns from both crystallographic structures (which provide spatial information) and binding assay data (which provides thermodynamic information). A set of "physicality steering potentials" (in the variant called Boltz-2x) further improves predictions by enforcing physical constraints during inference — ensuring that predicted complexes obey basic chemistry.
Benchmark Reality
Rigorous benchmarking is essential in this field because overfitting to known structures can produce misleadingly optimistic results. The Boltz-2 paper evaluated performance on complexes deposited in the Protein Data Bank during 2024 and 2025 that were significantly different from any structure in the training set.
On structure prediction, Boltz-2 matched or exceeded AlphaFold3 and Chai-1 across protein-protein, protein-nucleic acid, and protein-ligand complexes. On binding affinity, Boltz-2 achieved correlations with experimental measurements approaching those of physics-based free energy perturbation (FEP) calculations — methods that typically require hours or days of GPU computation per compound rather than seconds.
This is the transformation. FEP calculations have been the gold standard for binding affinity prediction in pharmaceutical research, but their computational cost limits them to evaluating dozens of compounds at a time. Boltz-2-class models can screen thousands of compounds per hour at near-FEP accuracy. The throughput improvement is not incremental; it is three to four orders of magnitude.
The Pfizer Deal and Industrial Adoption
In January 2026, Pfizer signed a major platform deal with the Boltz team, implementing the model across their discovery pipeline. The deal was one of three major AI partnerships announced in the first weeks of the year — alongside Eli Lilly's deal with Chai Discovery and GSK's agreement with Noetik — but Pfizer's focus on Boltz specifically highlighted the commercial value of combined structure-affinity prediction.
Pharmaceutical companies had already been using AlphaFold2 and its successors for target identification and hit finding. But the addition of reliable affinity prediction changes the workflow fundamentally. Instead of predicting structure, then running separate (expensive) affinity calculations, then synthesising and testing the top candidates, teams can now computationally rank thousands of molecules by predicted binding strength before synthesising anything.
What Remains Unsolved
Despite the progress, significant gaps remain. Protein dynamics — how proteins move and flex over time — are poorly captured by any current model. AlphaFold and its successors predict a single static structure (or a small ensemble), but many drug targets exist in multiple conformational states. A drug that binds tightly to one conformation may be useless against another.
Membrane proteins, which represent roughly 60% of drug targets, remain more challenging than soluble proteins. Intrinsically disordered regions — stretches of protein that don't adopt a stable fold — are essentially invisible to current methods. And while Boltz-2's affinity predictions correlate well with experiment on average, the error bars on individual predictions are still too wide to replace experimental measurement entirely.
The field is also grappling with a data bottleneck. All current models are trained primarily on structures deposited in the Protein Data Bank, which contains roughly 220,000 experimental structures. This sounds like a lot, but it represents a tiny fraction of the protein universe, and the distribution is heavily biased toward proteins that crystallise easily — which are not necessarily the most important drug targets.
The Next Frontier
The trajectory is clear. Within two to three years, structure-affinity prediction models will likely incorporate molecular dynamics simulations, enabling predictions of how binding changes as proteins flex. Integration with generative chemistry models — which design novel molecules — will close the loop from target structure to optimised drug candidate in a single computational pipeline.
The most consequential implication may be for neglected diseases. Structure-based drug design has historically been dominated by well-funded therapeutic areas (oncology, immunology, metabolic disease) because the computational costs limited its application. If Boltz-2-class models reduce the cost of structure-affinity prediction to near zero, there is no computational reason why the same approaches cannot be applied to malaria, tuberculosis, or any other target with a known sequence.
AlphaFold solved the protein folding problem. The generation of models that followed — Boltz-2 chief among them — may solve the drug design problem. The question is no longer whether AI can predict how drugs bind to their targets. It is how quickly the pharmaceutical industry reorganises itself around the fact that it can.