🔬Beyond AlphaFold: How Boltz is Open-Sourcing the Future of Drug Discovery

Summary

• AlphaFold 2’s landmark result was powerful but narrower than “protein folding solved”: it became remarkably good at predicting single-chain protein structures when evolutionary clues are available. It still does not explain the folding pathway, intermediate states, protein motion, or state probabilities; Jeremy Wohlwend’s careful distinction is that models learned to go “straight to the answer,” not to reconstruct how biology gets there. AlphaFold also shifted Jeremy from theoretical and methodological ML toward applied ML.
• Structural-biology models follow a different scaling regime from LLMs, with fewer than 1 billion parameters but unusually high inference cost. AlphaFold 2 was only around 70 million to 100 million parameters, yet pairwise representations create cubic operations, specialized architecture remains “vastly superior” to a simple transformer, and recycling the answer three or four times—sometimes tens of times—resembles an early form of reasoning.
• AlphaFold 3’s closed release created the opening for Boltz-1 and ultimately the company. DeepMind had open-sourced AlphaFold 2, reportedly used by more than 1 million scientists, but withheld AlphaFold 3 as Isomorphic Labs pursued pharmaceutical ambitions. Boltz responded with a fully open model approaching AlphaFold 3 accuracy, trained once in a chaotic run that underwent live “surgery” and “by some miracle…worked out.”
• Boltz’s commercial thesis is that open weights do not commoditize the full product: value accumulates in scientific workflows, accelerated infrastructure, interfaces, and validation. Jeremy Wohlwend says “putting a model on GitHub is definitely not enough” for working chemists and biologists; Boltz intends to keep base models open while selling a substantially easier, faster service and learning from community feedback.
• The technical engine is increasingly generate, score, and search rather than trust one prediction. BoltzGen jointly designs a protein’s atomic structure and amino-acid identity, while structure-consistency, confidence, and increasingly direct affinity prediction filter candidates. The founders’ reduction is commercially consequential: if enough samples contain a good answer, “then it really just becomes a ranking problem.”
• BoltzGen’s strongest evidence is deliberately broad wet-lab validation, including generalization beyond known training interactions. Across nine PDB targets with no known protein interaction, the team tested 15 mini-proteins and 15 nanobodies per target and obtained nanomolar binders on two-thirds of the targets—roughly the binding strength they say may be needed therapeutically—within a program spanning about 25 external labs.
• Boltz Lab turns those models into large parallel design campaigns, but the founders explicitly reject the “AI makes drugs” shortcut. A campaign might screen 100,000 candidates, and the platform can marshal thousands of GPUs; its small-molecule screening pipeline is claimed to be 10× faster than the open-source version. Yet “these are not drugs”: developability, ADME, toxicity, cellular context, and lab iteration remain essential, and actual hits—not attractive predicted molecules—are what persuade skeptical scientists.

Deep dive

1. AlphaFold 2 redirected careers by cracking a bounded scientific problem

• Gabriele Corso frames structural biology as the attempt to see proteins and other cellular molecules at atomic resolution without always resolving crystals and shooting X-rays. AlphaFold 2 showed that applied machine learning could make dramatic progress on a long-standing scientific problem. Jeremy Wohlwend says it shifted him from theoretical and methodological ML toward applied ML.

• Gabriele says their group, already working on small molecules, shifted toward biologics after AlphaFold. It answered one question while opening many more: “Can we do this on other things than proteins?” That meant protein–small-molecule interactions, nucleic acids, multi-chain proteins, and increasingly complex molecular systems.

• CASP 14 supplied the decisive test: organizers assembled structures deliberately different from what was already known, and AlphaFold 2 “really blew everything out of the water.” CASP 15 and CASP 16 subsequently expanded toward other modalities, where performance remains materially weaker than on monomeric proteins.

2. Structure prediction advanced much further than folding or dynamics

• Jeremy resists saying the problem was “solved.” The sharpest claim is that models became very good at predicting single-chain protein structures, particularly when rich evolutionary information exists; without that co-evolutionary landscape, performance can decline substantially.

• Structure prediction and folding are different tasks. Today’s systems can often jump from sequence to a plausible final structure, but folding asks how a protein moves from disorder into that structure—including intermediate states—and Jeremy says, “I don’t think we’ve made that much progress on” the latter.

• Proteins are also not static objects: they move among shapes associated with different energy states. The unresolved problem is not merely enumerating those states, but estimating “at what frequency, what probability” each occurs.

• Gabriele explains why this matters for disease: misfolding contributes to some diseases, and without understanding the process “we don’t really know how to intervene.” A final structure is like seeing an assembled car; the folding pathway is how the car is made and where assembly can fail.

3. Evolutionary mutations guide models into the right structural valley

• Co-evolution supplies a powerful geometric clue. If one amino acid mutates and disrupts nearby three-dimensional neighbors, evolution may select compensating changes elsewhere; correlated mutations therefore suggest that two sequence positions sit close together in the folded structure.

• Jeremy’s analogy is a huge three-dimensional energy valley with an overwhelming search space. Evolutionary hints place the model “kind of close to the solution,” after which it may possess some learned physics to descend toward a low-energy state—even if it cannot navigate the whole landscape unaided.

• A multiple sequence alignment, or MSA, provides the initial contact hints. The pairwise architecture then propagates constraints—if two residues must be close and one is connected to another, further distances become constrained—before decoding the rough distance matrix into coordinates.

• This reframes an old theoretical puzzle: protein folding had been treated as a classic NP-hard combinatorial problem because possible conformations explode with sequence length. Machine learning revealed usable signal in sequences and evolution that humans could not explicitly extract.

4. AlphaFold 3 unified molecular interactions through generative modeling

• After AlphaFold 2, the obvious frontier was interaction: proteins derive function through other protein chains, small molecules, RNA, and DNA, while therapeutics intervene by changing those relationships. Earlier groups often modeled each interaction type separately; AlphaFold 3 trained one large system across them.

• Its critical conceptual shift was from regression toward generative modeling. A regression model uncertain between several structures tends to average them into one potentially bad answer; a generative model samples a posterior distribution of alternatives, allowing separate scoring models to rank them.

• That distributional view also creates a path toward dynamic systems with multiple valid structures. Gabriele nevertheless presents it as a beginning, not a complete treatment of molecular motion or uncertainty.

• The discussion contrasts a more transformer-like final coordinate-generating component with a specialized equivariant architecture, but Jeremy rejects a pure “bitter lesson” reading. The field’s working consensus is that specialized architectures remain far from simple transformers and deliver “vastly superior” performance.

5. Biology models spend far more computation per parameter than LLMs

• AlphaFold 2 was roughly 70 million to 100 million parameters, and AlphaFold 3 remains below 1 billion—tiny by contemporary LLM standards. Yet running these models can be more expensive because the architecture operates over pairwise representations, turning quadratic work into cubic interactions.

• Jeremy contrasts language models’ storage of enormous quantities of information with structure models’ job of “decoding this input.” Templates can act like database lookups, but much of the computation is devoted to resolving spatial constraints rather than memorizing facts.

• Recycling adds test-time depth: the model builds an initial account of pairwise interactions, feeds that result back into its input, and decodes it again. Standard runs may recycle three or four times; researchers have tried tens of iterations, which Gabriele calls a “very, very early version” of reasoning.

6. AlphaFold 3’s closed model catalyzed Boltz-1

• AlphaFold 2 was open source and reportedly reached more than 1 million scientists. AlphaFold 3 arrived as an exceptional paper without a released model, which the founders connect to DeepMind spinning out Isomorphic Labs and pursuing commercial pharmaceutical development.

• That removed the foundation on which academic and industry groups expected to build. Gabriele and Jeremy used the AlphaFold 3 manuscript, their prior diffusion work, and rebuilt data pipelines to produce Boltz-1—the first fully open-source model they say approached AlphaFold 3-level accuracy.

• Work began in late May and the model appeared in November. Compute, not ideas, was the main bottleneck: their Department of Energy cluster would train for two days, then leave the job queued for a week, until Genesis helped them finish the run.

• They trained the large model only once. Gabriele remembers stopping it, fixing bugs, and restarting without returning to the beginning; the irreproducible model followed such a strange curriculum that “it’s learned some weird stuff,” yet “somehow, by some miracle, it worked out.”

7. Benchmarking keeps impressive averages from hiding brittle generalization

• Boltz-1 represented a large jump over prior open models, though Jeremy concedes that AlphaFold 3 still has an edge in some settings, notably many antibody–antigen predictions. Different models can win on different examples even when one performs better in aggregate.

• Structure prediction permits relatively straightforward retrospective evaluation: train only on Protein Data Bank structures released before a cutoff, then test on newer structures deliberately different from the training set. The caveat is that comparisons require credible knowledge of each model’s training data.

• Gabriele’s earlier DiffDock experience established the operating loop. It beat traditional physics-based methods on then-standard benchmarks, yet biologists exposed failures on unfamiliar proteins; the response was to build a harder benchmark first, then redesign for generalization.

• Gabriele welcomes the field’s “healthy skepticism” because many failures remain obvious. His optimism rests instead on the pace of improvement: if anything like the recent rate continues, “things are going to look pretty cool at some point in the future.”

8. Open source builds both a research flywheel and a product opportunity

• Boltz is structured as a public-benefit company around democratizing access, but Jeremy says an academic GitHub repository cannot deliver models into real therapeutic programs. Scientists need prepared targets, domain-specific workflows, reliable compute, and interfaces that do not require configuring GPUs.

• The business split is explicit: continue releasing base models so researchers can improve them and reveal failure modes, while productizing the best scientific experience. Jeremy compares this with an ML expert still choosing ChatGPT or Claude Code over personally hosting an open LLM.

• A Slack community numbering in the thousands became largely self-sustaining, while researchers published extensions on top of Boltz. Jeremy attributes adoption partly to usability: code quality was “far from perfect,” but it was easier to run than many alternatives.

• Contributions ranged from a complex GPU kernel for an architectural component dating back to AlphaFold 2 to hacks for cyclic peptides. The suite also expanded beyond Boltz-1: Boltz-2 added affinity prediction, estimating interaction strength rather than only structure.

9. Inference-time search is turning prediction into a ranking problem

• Community member Tim O’Donnell exposed a useful failure mode in antibody–antigen prediction. He conditioned Boltz on possible binding residues at intervals—first, eleventh, twenty-first, and so on—then ranked the resulting structures by confidence; the crude scan materially improved results.

• The founders’ reaction captures the research value: “Why would the model be so dumb?” Yet the brute-force success suggested a better question—how to make that search systematic and efficient rather than expecting a single unconditioned prediction to find the right binding site.

• The founders generalize the mechanism: sample enough candidate structures, assume some fraction contains a good answer, and the bottleneck becomes identifying it. “It really just becomes a ranking problem,” making better scoring models a potential route to the next major gains.

10. BoltzGen jointly designs protein sequence and atomic structure

• BoltzGen treats structure predictors as field-specific foundation models. Users provide a target—protein, DNA, or RNA—and a high-level specification such as an antibody framework or peptide; blank tokens represent the protein to be designed.

• The diffusion model generates both the new protein’s structure and its amino-acid identity. Because amino acids have distinct atomic compositions, supervising atomic placement implicitly reveals which amino acid belongs at each position, avoiding separate discrete sequence and continuous structure objectives.

• Evaluation begins with consistency: Boltz-2 independently predicts how the generated sequence should fold, and the pipeline compares that result with BoltzGen’s intended structure. Agreement raises confidence that the sequence will realize the proposed design.

• Confidence alone, however, is “not a very good predictor of affinity.” Gabriele says subsequent work has improved hit rates by directly predicting binding affinity rather than treating structural confidence as a proxy, with new results expected to be announced.

11. Wet-lab hit rates—not model aesthetics—anchor the platform

• Because Boltz is neither a biolab nor a therapeutics company, Hannes Stärk coordinated roughly 25 academic and industry labs to test designs across peptides, disordered proteins, small-molecule-binding proteins, nanobodies, and therapeutically relevant targets. The BoltzGen paper already contained results from roughly eight to ten labs.

• The cleanest generalization test selected nine PDB targets with no known interaction in the database, limiting the model’s ability to imitate a familiar binder. Testing 15 mini-proteins and 15 nanobodies per target produced nanomolar binders for two-thirds of the targets; the founders describe nanomolar binding as approximately therapeutic-strength affinity.

• Boltz Lab packages target preparation, protein and small-molecule “agents,” infrastructure, API access, and a collaborative interface. A campaign may generate 100,000 candidates; parallelism can replace weeks of serial inference, and the hosted small-molecule screen is claimed to run 10× faster than the open-source pipeline.

• Access is intended for academics, startups, biotechs, and enterprises, with free credits for early users and secure deployments for large organizations. Yet Jeremy’s boundary is categorical: “These are not drugs.” ADME, developability, toxicity, cellular pathways, and iterative lab results still separate a binder from an approved therapy.

• The internal medicinal chemist initially brought strong skepticism and extensive requested changes, then became the team’s heaviest compute user—running parallel hypotheses across hundreds of GPUs and applying expert judgment rather than blindly selecting the top 15. Gabriele’s adoption test is empirical: “For people to be convinced, you have to show them something that they didn’t think was possible.”

• The ultimate persuasion happens after wet-lab results return. Jeremy and Brandon emphasize that scientists are convinced by hits, not merely attractive predicted molecules. Jeremy also says broad, repeated CRO testing helps track month-to-month progress without overfitting to one therapeutic program.