In a blind test, Deep Origin’s model placed AstraZeneca’s oral PCSK9 inhibitor at an unexpected pocket months before a crystal structure that confirmed the pose to ~1.2 Å was published
Prospective screen across 80 billion virtual compounds yielded a 30.6% hit rate on CD73, ~100x higher than a prior machine-learning screen
By pairing machine learning with physics, DODock maintains over 50% pose accuracy on novel targets where leading co-folding models drop below 25%
Research published on bioRxiv discovered novel chemical hits across four targets of escalating difficulty up to an allosteric protein-protein interface
SOUTH SAN FRANCISCO, Calif., Aug. 05, 2026 (GLOBE NEWSWIRE) — Deep Origin today announced a comprehensive preprint on bioRxiv describing a major advance in virtual screening and validating the methodology prospectively in physical wet-lab assays, discovering chemically novel small molecules against four challenging therapeutic targets. Deep Origin’s novel framework – DODock and DOScore – maintains predictive accuracy on novel biological targets where conventional machine-learning models collapse.
AI-driven drug discovery has promised to unlock vast chemical space and find hits on novel targets far beyond the reach of traditional screening. In practice, however, the accuracy of current models collapses when confronted with genuinely novel biology and chemistry. Leading co-folding models drop in accuracy from mid-80s to below 25%, and their predictions frequently fail in the lab. Virtual screening is credited as the primary hit-finding strategy in approximately 1% of drug discovery campaigns, as legacy methods routinely deliver false positives that fail to translate into active physical hits.
“Virtual screening has been widely used for more than 40 years, yet it has rarely delivered an actual drug,” said Garegin Papoian, Ph.D., co-founder and chief scientific officer of Deep Origin. “The field learned to live with a low accuracy ceiling. We set out to change that by pairing physics with machine learning, ensuring our method holds its accuracy on novel targets where training data alone runs out.”
Docking: Unmasking the “memorization trap”
One challenge in solving this problem is the leakage of both similar and identical test data into model training datasets. Standard time-based benchmarks leave look-alike proteins and twin chemical skeletons in training sets, allowing AI models to score high simply by memorizing historical ligand-protein pairs. This leads to a “memorization trap” where models appear more accurate than they are.
The real-world impact of this is evident on the independent Runs N’ Poses benchmark. AI co-folding models (AlphaFold 3, Boltz-1, Chai-1, Protenix) achieve 75% to 88% pose accuracy on familiar targets but drop below 25% when tested across unfamiliar targets and novel chemical space. DODock achieves 89% accuracy over familiar complexes, while maintaining over 50% pose accuracy on the most novel complexes in this set.
To prevent data leakage across its own docking and scoring benchmarks, Deep Origin uses a strict double-similarity filter, removing any protein sharing 30% or more sequence identity and any ligand exceeding 0.4 Tanimoto similarity to test sets. Under these un-leaked conditions, DODock reaches 80% pose accuracy on OpenBind, a benchmark built specifically on novel target biology, while co-folding models achieve just 4% to 28% pose accuracy (and DiffDock falls to 2%). DOScore maintains high hit-enrichment across standard virtual screening sets.
Detailed benchmark analysis in the paper suggests that while co-folding architectures successfully reconstruct protein pockets, their accuracy collapses specifically during ligand placement and docking mechanics. No prior analysis has described this distinction.
How DODock and DOScore Work
Deep Origin’s methodology overcomes this trade-off by grounding AI in physics, not only in patterns learned from data:
- DODock (binding pose prediction): A diffusion model generates broad 3D candidate pose proposals. An expanded 80-parameter physical energy engine (DOFast) refines those poses by calculating fundamental molecular forces rather than relying on memorized data patterns. Finally, a powerful AI model ranks the refined poses by evaluating atomic contact points across the binding interface. On the independent PoseBusters benchmark, DODock consistently produces poses that are both accurate to within 2 Å and free of steric clashes and distorted geometries 84% of the time on the most familiar targets and 80% on the least, the top score in every similarity tier. Glide SP, a leading commercial program, falls from 80% to 47% across the same range. Of ten competing methods, none exceeds 72% on the least-similar targets.
- DOScore (binding affinity ranking): To resolve the issue of data scarcity in affinity modeling, Deep Origin used DODock to generate high-quality 3D structures for hundreds of thousands of assay-labeled compounds. DOScore evaluates atomic environment interactions alongside the broader 3D geometry of the binding pocket to rank true binders above decoys under strict zero-leakage splits.
“What matters is how a screen behaves on a target no one has solved yet, so that is where we tested our model with strict splits and cases built to make it fail,” said Michael Antonov, co-founder and chief executive officer of Deep Origin. “How our model performs on those rigorous tests, not on inflated benchmarks, is what dictates how many molecules brought to the bench are likely to be real.”
Summary table of benchmark results
| Task / Benchmark | Standard or Competitor | DODock / DOScore | What it measures | |
| Redocking pose accuracy (CASF-2016, strict split) | 40–55% (Glide, AutoDock Vina, Gold) | 81.1% | Heavy-atom pose prediction within 2 Å of the crystallographic pose | |
| Novel target pose accuracy (OpenBind) | 4–28% (AlphaFold 3, Boltz-1, Chai-1; DiffDock: 2%) | 80.0% | Pose accuracy on targets structurally distinct from training data | |
| Novel target generalization (Runs N’ Poses, 0–20% bin) | <25% (All leading co-folding models) | >50.0% | Combined pose accuracy (<2.0 Å) and interaction quality (LDDT-PLI > 0.8) | |
| Accurate and physically valid poses (PoseBusters benchmark, most-similar→ least-similar targets) | Glide SP, 80% → 47%; Vina, 64% → 50%; GNINA, 74% → 72%; DiffDock, 23% → 2% | 84% → 80% | Share of predictions within 2 Å of the crystallographic pose and passing all physical validity checks | |
| Early hit enrichment, EF@1% (DUD-E, DEKOIS 2.0, strict split) | Legacy ML models drop significantly under strict splits | >10-fold enrichment across the majority of targets | Rate of placing true binders in the top 1% vs. random selection | |
| CD73 prospective hit rate | ~0.3% (AtomNet 318-target screen; 0 hits <100 µM) | ~30% (56 of 183 <500 µM; 54 < 100 µM) | Confirmed inhibitors from a real lab screen | |
Results of Prospective Virtual Screening Across a Deliberate Difficulty Gradient
Deep Origin validated its workflow in prospective wet-lab assays against a ~80-billion synthetic compound library. The prospective campaigns evaluated four targets representing a deliberate gradient of structural difficulty:
- IRAK4 (kinase): yielded 19 active compounds among 127 tested (15.0% hit rate), including a high-potency 158 nM inhibitor engaging the ATP-competitive back pocket.
- CD73 (nucleotidase): yielded 56 active compounds among 183 tested (54 active under 100 µM; 30.6% hit rate), with a representative non-nucleotide hit reaching 570 nM cellular potency. A prior large-scale AI screen against CD73 confirmed a single weak inhibitor at 176 µM among 335 compounds (~0.3% hit rate), making Deep Origin’s result a ~100-fold improvement in screening efficiency.
- Factor XIa (protease): yielded 13 active compounds among 299 tested (4.3% hit rate), within a solvent-exposed, peptide-adapted cleft.
- IL-17A (protein-protein (PPI) target): yielded 6 active compounds among 194 tested (3.1% hit rate), successfully discovering small molecules that allosterically disrupt a distal protein-protein interface.
Across all four targets, validated hits demonstrated high scaffold novelty (ECFP4 Tanimoto similarity of 0.22–0.27 relative to known binders). The screens discovered entirely new chemical matter rather than minor variations on existing drugs.
In a fully blind prospective test on PCSK9, DODock predicted the binding pose for AstraZeneca’s Phase 3 candidate, laroprovstat (AZD0780), at an atypical C-terminal domain site. Deep Origin subsequently solved the 2.07 Å crystal structure experimentally, confirming the blind prediction to 1.2 Å heavy-atom RMSD.
Setting a standard for model performance reporting
The manuscript published today includes more than 80 pages of detailed supplementary documentation covering every algorithm, hyper-parameter and data split. By releasing the full methodology, Deep Origin invites the scientific community to audit, test and replicate these results.
“We published the whole method — every algorithm, the data strategies, the full supplement — and left it open for others to test it on molecules we have not seen,” said Papoian. “For science to advance, model performance reporting has to adhere to strict data rigor and transparency. My hope is that this represents a genuine step forward for docking and scoring, helping move the field forward alongside our own drug discovery programs.”
While the manuscript establishes a public baseline, Deep Origin’s internal production platform features next-generation models already operating above these published metrics.
About the Research
View the preprint, “Overcoming the accuracy-generalization tradeoff in docking and scoring for prospective virtual screening,” including supplemental information, on bioRxiv here.
This research was co-led by Head of AI Garik Petrosyan, Ph.D., and Dr. Papoian. The experimental validation was carried out in the company’s South San Francisco laboratory and with partner contract research organizations.
About DODock and DOScore
DODock and DOScore are the core modeling and screening engines of Deep Origin’s computational discovery suite. DODock predicts how a small molecule binds to a protein by combining machine-learning pose generation with a compact, interpretable physics model; DOScore estimates thermodynamic binding affinity across ultra-large chemical spaces. For more information on DODock, visit https://deeporigin.com/docking/.
About Deep Origin
Deep Origin is building computational discovery systems that model life, to close the gap between preclinical predictions and clinical outcomes. Co-founded by Michael Antonov, co-founder of Oculus, and Garegin Papoian, Ph.D., former Monroe Martin Professor of biochemistry at the University of Maryland, Deep Origin runs drug discovery programs — its own and its partners’ — through hybrid AI-mechanistic models that span biological scales from quantum to cellular to human body scale. The company engages through discovery partnerships and SaaS platform access. Deep Origin is backed by more than $50 million in capital and more than $32 million in non-dilutive funding, including an ARPA-H CATALYST award to build in silico models that can reduce reliance on animal testing in preclinical drug development. For more information, visit https://deeporigin.com/.
Media Contact:
Steve Cooper, Orangefiery
[email protected]
Photos accompanying this announcement are available at
https://www.globenewswire.com/NewsRoom/AttachmentNg/76c2b5c4-c48f-4f12-bd11-c1f38aa5bea9
https://www.globenewswire.com/NewsRoom/AttachmentNg/7ed22068-4b67-4a2d-a480-1cdb5696050f
https://www.globenewswire.com/NewsRoom/AttachmentNg/d016084e-8036-4808-b25e-5fbb1cdb30e0
https://www.globenewswire.com/NewsRoom/AttachmentNg/fe82ef89-0821-4753-b3e8-e66371a7d459 https://www.globenewswire.com/NewsRoom/AttachmentNg/f2081657-119b-409c-b408-5a83bcb5891c