BoltzMol-1

BoltzMol-1 is Boltz's first small-molecule hit-discovery pipeline, announced on June 16, 2026 alongside BoltzProt-1 and the commercial Boltz API. It addresses one of the central bottlenecks in early-stage drug discovery: finding chemical matter that genuinely binds a therapeutic target without the cost and timeline of traditional high-throughput screening (HTS). Boltz describes it as the most extensive experimental validation of AI-driven small-molecule hit discovery, and of AI co-folding for hit discovery, reported to date.

At its core, BoltzMol-1 is a version of Boltz-2 — the open structure and binding-affinity co-folding model — optimized for prospective screening. Where Boltz-2 established that a single model can jointly predict protein-ligand structure and affinity, BoltzMol-1 packages that affinity prediction into an end-to-end workflow that goes from a target to validated wet-lab hits in weeks rather than months. It operates in two modes: ranking compounds that are already buyable from in-stock catalogs, or generating and searching an ultra-large make-on-demand chemical space of more than 74 billion compounds.

The pipeline reflects a broader trend toward generative, affinity-aware screening that replaces brute-force physical assays with targeted, model-guided selection of a few dozen compounds per target — making hit discovery accessible to academic labs and small teams, not just large pharmaceutical organizations.

#Key Features

• Two screening modes: Rank compounds from in-stock catalogs, or generate and search a make-on-demand chemical space of more than 74 billion compounds for novel hits.
• Boltz-2 affinity backbone: Builds directly on the proven structure and binding-affinity co-folding of Boltz-2, optimized specifically for prospective virtual screening.
• Property-steered generation: Compound generation can be steered toward a desired property space, with a new set of ADME models included to bias toward developable molecules.
• Validated hit rate: Across ten challenging, therapeutically-relevant targets (most disparate from training data), it delivered confirmed hits on 6 of 10 while testing only 28–51 compounds per target.
• Fast, low-cost workflow: Reported to go from target to validated lab hits in 3–8 weeks on a combined compute and experimental budget of $10–15k, versus months and hundreds of thousands of dollars for traditional HTS.

#Technical Details

BoltzMol-1 inherits the co-folding architecture of Boltz-2, a diffusion-based model that jointly predicts protein-ligand complex structure and binding affinity, and tunes it for prospective screening. Inference is accelerated with NVIDIA cuEquivariance kernels. In prospective validation across ten targets, confirmed results included binders to the pseudokinase ROR1 across three orthogonal biophysical assays; new agonists and antagonists for the GPCR MRGPRX2 and functional small molecules for the GLP-2 receptor; small-molecule binders for the STAT6 transcription factor; functionally active compounds against PknB, an essential tuberculosis kinase (a collaboration with Anthony Gitter and Nathan Wlodar); and functionally active binders for the autophagy proteins LC3B and GABARAP (a collaboration with the Kritzer Lab at Tufts University). A technical report is published rather than a peer-reviewed paper or preprint, and the hit-discovery pipeline itself is delivered through the commercial Boltz API rather than as open weights or code.

#Applications

BoltzMol-1 targets early-stage drug discovery, supplying confirmed small-molecule hits for difficult or undrugged targets including pseudokinases, GPCRs, transcription factors, and antimicrobial targets. It is served through the Boltz API (predictions from roughly $0.025 each) with Python and JavaScript SDKs, first-party Claude Code integration, and Codex and Gemini CLI support, and is available through partner platforms such as Benchling, Phylo, Amazon Bio Discovery, Rowan, Tamarind, Kiin Bio, Pauling.ai, Mirror Physics, and Cultivarium. Academic labs, biotech startups, and pharmaceutical teams benefit from a screening workflow that compresses cost and timeline by orders of magnitude.

#Impact

By demonstrating confirmed hits on 6 of 10 challenging targets while testing only tens of compounds each, BoltzMol-1 offers one of the strongest prospective validations of AI co-folding for hit discovery to date, extending the open Boltz-2 affinity model into a practical discovery engine. Its reported 3–8 week, $10–15k path from target to validated hits — against months and hundreds of thousands of dollars for conventional HTS — signals a shift toward model-guided, low-throughput screening accessible to smaller teams. A key limitation is openness: unlike the open-source Boltz-1, Boltz-2, and BoltzGen, the BoltzMol-1 pipeline is API-only, with no peer-reviewed paper, preprint, or released weights, so independent reproduction currently depends on the published technical report and the commercial API.