What it is
Boltz-1 was the first fully open, commercially-usable model to match AlphaFold 3 on biomolecular complex prediction. Boltz-2 (2025) added fast, accurate binding-affinity prediction — approaching physics-based free-energy methods at a fraction of the cost — a key missing piece for virtual screening.
Evidence trail
BioAtlas keeps the path from source to decision visible. A connection records provenance; it does not imply that evidence is sufficient for every context.
Model passport
How Boltz-1 / Boltz-2 represents biology
Category is navigation. These fields describe the model-specific computational transformation and deliberately override broad category defaults.
Biological scale
Modalities & tasks
Registry, claims and frontier intelligence
Boltz-2
2 version records · latest curated year 2025. Model-family identity remains separate from capability and access changes.
Explore version lineage →1 normalized claim
Biomolecular complex prediction · PoseBusters and affinity benchmarks
Open claim intelligence →6 connected frontiers
Open DDEs · Recent preprint
Inspect research horizon →Connected research frontiers
These records describe active research directions, not guaranteed capabilities of this model. Evidence stages and unresolved questions are preserved separately.
Open reproductions of frontier drug-design engines
Aureka AI OpenDDE project · 2026-07-04Can the community reproduce and extend proprietary all-atom drug-design engines with open training code, checkpoints and benchmarks?
Evidence boundary and unresolved questions
OpenDDE is a very recent July 2026 preprint. Its claimed parity has not yet received broad independent evaluation.
- Can external teams reproduce the reported training and benchmark results?
- What data provenance and leakage controls are documented?
- How do open checkpoints perform in prospective discovery projects?
open science · co-folding · reproducibility · scaling lawsOpen frontier record →Joint structure and binding-affinity reasoning
Isomorphic Labs · MIT · Open community · 2026-03-06Can one model predict both the bound geometry and the energetic ranking needed to prioritize compounds?
Evidence boundary and unresolved questions
Recent independent evaluations report that strong global correlations may not preserve ranking among top compounds, where lead-selection decisions occur.
- Can models rank close analogues within a chemical series?
- How well do they extrapolate across assays, protonation states and target families?
- Can uncertainty identify when physics or experiment should take over?
binding affinity · ranking · virtual screening · free energy · uncertaintyOpen frontier record →Finding cryptic pockets from sequence
Isomorphic Labs · 2026-02-10Can a model reveal ligandable pockets that are hidden in the unbound protein and only open after a ligand or allosteric change?
Evidence boundary and unresolved questions
The public evidence is currently a company technical report and benchmark narrative. Prospective medicinal-chemistry validation and independent replication remain essential.
- How often are predicted pockets experimentally ligandable rather than geometrically plausible?
- How robust is pocket discovery across membrane proteins, intrinsically disordered regions and low-data target families?
- Can calibrated confidence distinguish genuine induced pockets from model hallucinations?
cryptic pockets · allostery · induced fit · ligandability · dockingOpen frontier record →Induced-fit co-folding beyond familiar targets
Isomorphic Labs · 2026-02-10Can structure models represent large ligand-driven protein rearrangements when the target, pocket or conformational transition is far from training examples?
Evidence boundary and unresolved questions
Out-of-distribution claims depend strongly on benchmark construction, training-set leakage controls and exact success thresholds.
- How are unseen chemotypes and target families isolated from training data?
- Does structural accuracy translate into enrichment or medicinal-chemistry decisions?
- How stable are alternative conformational ensembles?
co-folding · induced fit · OOD generalization · protein flexibilityOpen frontier record →High-fidelity antibody and biologic interfaces
Isomorphic Labs · 2026-02-10Can general co-folding models accurately resolve antibody–antigen and other biologic interfaces with low sequence homology?
Evidence boundary and unresolved questions
Benchmark composition, success thresholds and independent reproduction will determine how broadly the reported advantage generalizes.
- How does performance vary across CDR loops, nanobodies and multispecific formats?
- Can interface prediction improve prospective affinity maturation?
- How are glycosylation and conformational heterogeneity handled?
antibodies · biologics · interfaces · low homologyOpen frontier record →Fast, uncertainty-aware affinity screening
Terray Therapeutics research team · 2026-02-08Can virtual screening retain useful structural and affinity accuracy without expensive all-atom diffusion for every compound?
Evidence boundary and unresolved questions
The results are preprint claims; proprietary assay details and cross-lab prospective replication remain limited.
- Does coarse representation preserve water, ion, metal and covalent chemistry?
- How well does uncertainty calibrate under target and chemistry shift?
- Can active learning improve real design-make-test cycles?
virtual screening · uncertainty · active learning · coarse representationOpen frontier record →Inputs and outputs
Inputs
Biomolecular complex specificationOutputs
Complex structuresBinding-affinity predictionsScientific and technical profile
Scientific principles
Technology
Scientific lineage
These are transparent concept matches—not claims that one scientist alone caused this model. Each connection is based on the model’s recorded domain, scientific principles, technical terms or an explicit lineage link.
Anfinsen’s dogma—the thermodynamic hypothesis
Christian B. AnfinsenProtein structure prediction, inverse folding and generative protein design all assume that sequence strongly constrains structure and function.
Denoising diffusion generative models
Jascha Sohl-Dickstein, Jonathan Ho and collaboratorsModern protein-backbone, molecular-pose and biomolecular-complex generators use diffusion to sample valid three-dimensional structures and designs.
First atomic structures of globular proteins
John Kendrew & Max PerutzProtein structure prediction and structure-based design became meaningful because experimental crystallography established the target reality to predict against.
Levinthal’s paradox and efficient folding pathways
Cyrus LevinthalModern folding algorithms, energy landscapes, learned priors and diffusion models solve a constrained search problem rather than brute-force conformational enumeration.
Multiscale modelling of chemical systems
Martin Karplus, Michael Levitt & Arieh WarshelQM/MM, molecular dynamics, free-energy methods and physics–ML hybrid platforms descend directly from this multiscale strategy.
Computational protein structure prediction and de novo design
David BakerModern binder design, inverse folding and diffusion-based protein generation build on this computational-design lineage.
Evaluation evidence
Task-specific evidence only; not comparable as a universal leaderboard score.
PoseBusters and affinity benchmarks
Boltz-2 · PoseBusters and affinity evaluation setsOpen evaluation reports structure prediction and later affinity capabilities.
Claim caveats
- Boltz-1 structure claims and Boltz-2 affinity claims should be separated by version.
- Affinity performance depends strongly on target family and split design.
Known limitations
- Performance depends on the evaluation dataset and operating conditions.
- Task-specific benchmark results should not be compared across unlike domains.
- Outputs require task-specific scientific and experimental validation.
Milestones
BoltzGen extends the line toward universal binder design.
Widely used as an open AF3 alternative.