Beyond AlphaFold: Boltz’s Path from Structure Prediction to Molecular Design

One training run, with repairs along the way

The conversation opens with enough compute for one large-model training run. Bugs kept appearing, so the team stopped training, patched the code and resumed without returning to the beginning. The model accumulated a changing training history, but the participant recounting the run says it worked.

Boltz builds models for predicting molecular structures and designing new proteins, along with tools that scientists can use in discovery programs. Its ambitions grew beyond an academic project into a public benefit company: making models accessible also required usable workflows, computing infrastructure and experiments that distinguish plausible predictions from working designs.

AlphaFold’s breakthrough—and the problems it left open

Structural biology studies the shapes and interactions of proteins and other molecules at atomic detail. Computational structure prediction aims to infer those shapes without determining each one experimentally through methods such as X-ray crystallography. AlphaFold’s impact drew one participant from theoretical and methodological machine learning into applied structural biology. Another recalls their research group expanding from small molecules toward proteins and more complex molecular interactions.

AlphaFold 2’s major advance was single-chain structure prediction: predicting the shape of a protein represented by one amino-acid sequence. A participant recalls its performance at CASP 14, a competition testing methods on challenging protein structures, as a large improvement over competing methods and earlier competitions. Later competitions expanded to small molecules and nucleic acids, where substantial difficulties remained.

Three different problems sit behind claims about predicting proteins: - Structure prediction: infer a resulting three-dimensional arrangement. Performance is particularly strong when evolutionary information is available, and weaker without it. - Folding: explain how a disordered chain reaches a structured state. The participants judge progress on this process much less complete. - Changing states: identify the different shapes a protein occupies and their frequencies or probabilities. One predicted structure does not provide that account.

Shape helps explain function much as an assembled car reveals more than a parts list. Proteins perform many cellular tasks through interactions with other molecules, and functional protein machines often contain multiple chains. Designing a disease intervention or a biosensor therefore means designing something that interacts with a chosen target in a particular way. Diseases involving misfolding also make the path to a structure consequential.

Evolution constrains a vast structural search

The number of possible protein shapes grows combinatorially with sequence length. Evolution supplies clues that constrain this search. Across related proteins in different organisms, some sequence positions tend to mutate together. A change at one position alters its three-dimensional surroundings; a compensating change elsewhere may preserve structure and function. Aggregated across many sequences, correlated mutations can indicate spatial proximity.

Multiple sequence alignment (MSA) brings related sequences into comparison to supply these evolutionary hints. A partial explanation of AlphaFold 3 follows the information through its architecture: hints about nearby amino acids inform pairwise contacts; the model propagates those constraints into a rough pairwise distance representation, then decodes a potential structure.

One participant proposes that evolutionary clues place the model near a plausible region of the structural energy landscape, where learned local physical regularities can guide it toward a lower-energy configuration. This is an interpretation of model behavior, with an important limit: the model may not have learned enough physics to navigate the entire space unaided. Without co-evolutionary information, prediction becomes harder.

Sample possible interactions, then evaluate them

After single-chain prediction, the next priority was modeling interactions among proteins, small molecules and nucleic acids. In the team’s account, AlphaFold 3 unified previously separate molecular modeling tasks and advanced performance across them. A key architectural change, also explored by other teams, was generative structure prediction: sampling possible structures from a distribution instead of regressing toward one answer.

When several answers are plausible, regression can average competing answers. A generative model can instead produce distinct candidates for a separate evaluator to assess. This supports two complementary tasks: - Generate alternatives: sample distinct possible structures, including different conformations. - Select candidates: evaluate those structures and choose promising answers. Representing alternatives creates room to model dynamic systems, but does not itself establish the physical probabilities of their states.

The module that turns pairwise representations into structures became simpler and more like a conventional transformer, while the broader architecture remained specialized. A participant argues that specialized architectures still substantially outperform simple transformers in this field, while acknowledging debate.

Parameter count understates the computational work. The participants describe AlphaFold 2 and 3 as having fewer than a billion parameters but a high computation-to-parameter ratio. Pairwise representations and cubic interactions make the models expensive to run despite relatively small weight counts.

Recycling adds repeated computation: the model feeds its intermediate understanding of pairwise interactions back into its input for another decoding pass. Initial inputs can include evolutionary information and templates retrieved from similar known structures. A participant describes three or four passes as customary and recalls trying tens of passes, without reporting a measured benefit from those larger counts.

Boltz-1: an open foundation under compute constraints

AlphaFold 3’s paper arrived without the accompanying model release that Boltz’s builders wanted as a foundation for further research. Boltz-1 pursued an accessible model with similar accuracy, drawing heavily on the manuscript. A participant recalls starting in late May and releasing in November. Prior experience with related models helped, although the paper exposed substantial gaps in their data pipelines.

Boltz-1 was the setting for the opening account of training through bug fixes. Access scheduling compounded the compute shortage: much of the training ran on a shared Department of Energy cluster, with roughly two days of execution followed by about a week in the queue. Genesis supplied compute to help finish the model, which a participant estimates saved a couple of weeks.

The team describes Boltz-1 and contemporary open models as a substantial advance over earlier open releases, approaching AlphaFold 3. Performance still depended on the input: one participant judged AlphaFold 3 stronger in aggregate, especially at the time, with an apparent edge in many antibody–antigen predictions.

For structure prediction, a temporal split tests generalization to later discoveries. Models train on structures published in the Protein Data Bank (PDB), a shared structure database, before a cutoff. Evaluation then uses newer structures selected to differ substantially from earlier ones. Fair comparisons require knowing the models’ training sets and aligning their cutoffs; evaluating entirely new molecular designs is more complicated.

User failures can expose gaps in those tests. An earlier protein–small-molecule interaction model reportedly beat some traditional physics-based methods on its benchmarks, yet struggled when biologists applied it to proteins unlike its training data. The team collaborated with a Harvard group to build a benchmark targeting that failure before pursuing architectural improvements. The sequence was observe a failure, build a test that exposes it, then improve the model.

Open releases turn users into model contributors

Publishing a model does not make it usable in a therapeutic program. Chemists and biologists need workflows that accept their data, answer practical questions and run without GPU provisioning. Boltz intends to keep releasing some base models openly while selling infrastructure and a more usable experience. The team reports better results from additional inference, making the cost of running open models a central product concern.

The team credits attention to code usability with helping adoption, while acknowledging room for improvement. It reports a Slack community with thousands of members answering one another’s questions, papers extending Boltz, a contributed GPU kernel and modifications for cyclic peptides. Continued releases expanded the suite: Boltz-2 added affinity prediction, estimating the strength of molecular interactions, followed by a protein-design model.

A community idea credited to Tim O’Donnell began with an antibody–antigen failure: the model repeatedly placed the antibody against the wrong part of its antigen target. The workaround converted that error into inference-time search: 1. Choose candidate binding positions across the antigen—for example, amino-acid positions 1, 11 and 21, continuing the scan. 2. Condition separate predictions on binding at each position. 3. Rank the resulting structures by model confidence and select the top predictions.

The participants report that this crude search helped antibody–antigen prediction substantially, without quantifying the improvement. It prompted work on smarter alternatives to brute force. The broader opportunity depends on both halves of the system: sampling helps when it produces good candidates that the scorer can recognize.

BoltzGen encodes sequence inside atomic structure

The team treats structure predictors as foundation models for molecular work: learning how molecules interact provides a starting point for other tasks. Boltz-2 reuses that learning to predict affinity. BoltzGen fine-tunes a structure predictor to design new proteins, replacing the sequence to be designed with blank tokens and training the model to predict both structure and amino-acid identities.

A design request provides a binding target, such as a protein or nucleic acid, and a specification for the new protein—for example, an antibody with a particular framework or a peptide. The specification becomes tokens and conditioning. A diffusion model then generates the new structure and sequence.

The distinctive representation merges two prediction tasks. Amino acids have different atomic compositions, so atom placement encodes both geometry and amino-acid identity. This lets the model use structural supervision rather than separate discrete sequence and continuous structure signals. The team acknowledges precedents in earlier papers and credits Hannes Stark with taking the approach to large scale.

Selecting designs requires keeping three questions distinct: - Structural consistency: does another prediction of the generated sequence’s structure, using a model such as Boltz-2, agree with BoltzGen’s intended structure? Agreement supports the possibility that the sequence will realize the design. - Structural confidence: how confident is the prediction model in that structure? This provides another filter in the released BoltzGen pipeline. - Binding affinity: how strong is the interaction? Structural confidence is a poor predictor of affinity. A participant describes better hit rates from direct affinity prediction as forthcoming results, without numerical improvements or evaluation conditions.

Experimental validation tests binding and generalization

Experimental validation required outside collaborators because the team was neither a biology lab nor a therapeutics company. It sought varied tasks to avoid concentrating on one convenient system. A participant credits Stark with coordinating approximately 25 academic and industry labs willing to test designs, and recalls results from roughly eight to ten labs appearing in the paper, with further testing ongoing. Tasks included peptide design, ordered and disordered protein targets, nanobodies and proteins designed to bind small molecules for potential biosensing applications.

In one nanobody evaluation, the team generated tens of thousands of candidates, ranked them and selected 15 designs per target across 14 targets. Two metrics measured different outcomes: - Target coverage: how many targets yielded at least one binder? - Design hit rate: what fraction of tested designs proved to be good binders? The participants describe the setup without giving its numerical success rates.

A separate experiment tested design beyond imitation of known interactions. The team selected nine PDB targets, excluding cases where the target or a similar protein had a known protein-binding interaction in the database. It reports testing 15 miniproteins and 15 nanobodies per target and obtaining nanomolar binders for two-thirds of the targets. This measures target coverage; separate hit rates for the two design classes were not specified.

The speaker characterizes nanomolar affinity as roughly the binding strength sought for a therapeutic. It establishes a binding result, not therapeutic efficacy. Smaller protein designs may be easier to manufacture, but a participant identifies potentially reduced selectivity as a tradeoff. Proposed nanobody therapeutics still need humanization and work to establish suitable safety characteristics.

Boltz Lab combines workflows, compute and collaboration

Boltz Lab turns individual predictions into molecular design workflows. The platform addresses three needs: - Design and search: prepare the target, improve its predicted structure, generate candidates and rank them. For small molecules, a generative model selects appropriate building blocks to keep the search within a space described as synthesizable. - Parallel computation: distribute large searches across a GPU fleet. The illustrative workload is 100,000 candidates, with small-molecule designs taking a few seconds each and proteins potentially longer. - Integration and collaboration: provide an API for existing company workflows and an interface for launching jobs and analyzing results. Medicinal chemists can supply individual rankings and build consensus about what to test.

The protein and small-molecule workflows are called agents: automated recipes that coordinate multiple models and pipeline stages. Their role is to carry out the complex sequence of preparation, generation and evaluation needed for molecular design.

At recording time, anyone could request access while usage was being ramped up. The team described credits for academics, introductory calls and credits for startups and biotechs, and custom deployments for larger companies. Serving scientists beyond large enterprises was a stated product principle.

A participant claims that Boltz Lab’s small-molecule screening pipeline runs ten times faster than the open-source version, with lower costs for large screens. No hardware specifications, workload details or numerical prices accompany the comparison, so the claimed gain cannot be applied to a particular deployment from this account alone.

Close the loop with experiments and expert judgment

The full design system must be compared experimentally on both hit rate and binder strength. The team aims to test across tens of targets, choosing problems difficult enough to expose limitations while still allowing progress to be measured. Academic collaborations continue, with increasing use of external experimental providers for repeatability and speed. Repeated tests against the same targets help track changes from month to month.

Boltz intends to remain a tools provider. Academic collaborators had been allowed to keep and use their experimental results, while results from contracted experiments had so far been released. A designed binder still needs a therapeutic hypothesis and substantial downstream development before it can become a drug.

The team is exploring whether properties relevant to drug development can be incorporated into designs from the start, including properties that require understanding cellular behavior. One proposed feedback loop begins with uncertainty about how to attack a viral target: design binders against different sites, test them experimentally and feed the results back into the models. The aim is to learn how molecular interactions behave in their cellular setting. The team explicitly stops short of describing this as a virtual cell.

A medicinal chemist identified as Jeffrey helped reshape the workflows and interface, then became an intensive user. The team reports that he ran several screens in parallel across hundreds of GPUs to explore different targeting hypotheses and molecular spaces. He reviewed the results rather than simply taking the top fifteen, combining model assessments with his chemical judgment. Expert interpretation remains part of design selection, and the collaborative interface is built to support it.

For skeptical scientists, a participant argues, confidence ultimately comes from experimental hits. An attractive interface and plausible-looking predicted molecules are insufficient; the tool has to produce a useful result the scientist did not think possible.

The conversation closes with a recruiting invitation across software, machine learning, science and infrastructure. At recording time, Boltz welcomed applicants regardless of their biology or chemistry experience, with the ambition of helping shape those fields over the next five to ten years.