DOScore

Virtual screening promises access to tens of billions of purchasable compounds, yet it is rarely used as a primary hit-discovery strategy because the underlying scoring functions underperform in practice. Classical docking scores generalize across targets but are capped in accuracy by simple, largely additive functional forms, while recent machine-learning scoring functions are far more expressive but fail to transfer to novel molecules and pockets — their headline accuracy inflated by train-test leakage between closely related complexes.

DOScore, from Deep Origin, is a scoring model built to keep the expressiveness of machine learning while degrading gracefully as candidates move away from the training distribution. It is the scoring half of a paired system introduced in a single 2026 preprint; its companion docking model DODock generates poses, and DOScore ranks them by predicted binding affinity. DOScore represents each protein-ligand complex as an atomic graph whose edges are defined by spatial proximity rather than covalent bonds, so non-bonded and long-range interactions become first-class features rather than after-the-fact corrections.

Because DOScore is trained on complexes that DODock itself docked at scale, the two form a self-reinforcing pipeline: docking produces the structural data that scoring learns from, and scoring selects among the poses that docking proposes. The result is a fixed checkpoint that the authors apply prospectively to fresh targets without any target-specific retraining.

#Key Features

• Proximity-graph representation: Each complex is an atomic graph with edges set by spatial proximity, and local environments are encoded with atomic environment vectors, letting the model capture non-covalent and long-range contacts directly.
• Dual GNN branches: Two parallel 4-layer message-passing branches — a geometric PPFConv branch using pairwise distance and orientation descriptors and a GENConv branch using softmax neighborhood aggregation — each emit 256-dimensional node embeddings that are concatenated.
• Over-smoothing controls: Both branches use manifold-constrained hyper-connections (4 streams, 10 Sinkhorn iterations) to preserve discriminative node features across depth.
• Structure-biased transformer readout: A transformer encoder injects pairwise distance bias into its attention scores, then ligand-focused attention pooling reads out a single scalar affinity from the ligand atoms alone.
• Two task-specific variants: A hit-identification variant trained with an asymmetric loss that penalizes only false-positive decoys, and a fine-grained variant using intra-assay mean-centered regression with a pairwise ranking penalty to absorb assay-to-assay bias.

#Technical Details

The graph module's two 256-dimensional branches concatenate to 512-dimensional node embeddings before the distance-biased transformer aggregates global context. Scoring training data was built from PocketAffDB-derived assay records with receptors reprocessed from the RCSB PDB: assay ligands were docked by DODock into their assay-specific pockets, the top pose retained, and property-matched but topologically dissimilar decoys (ECFP4 Tanimoto below 0.28) added for the hit-identification task. Models train for up to 10,000 steps at an effective batch of 256 complexes, with EMA weight averaging across five checkpoints and a mix of AdamW and Muon optimizers. On the DUD-E and DEKOIS 2.0 screening benchmarks under a stringent split that removes every training protein above 0.3 sequence identity and every ligand above 0.4 Tanimoto to the test set, DOScore reaches EF@1% of 22.33 and 16.74; under the protein-only 30% split used by prior work it rises to 32.11 and 21.14, ahead of LigUnity (26.67 / 17.78) and DrugCLIP (18.5 / 11.5). On affinity ranking it posts R² of 0.47 (JACS) and 0.37 (Merck), and a Spearman correlation of 0.47 on OpenBind, comparable to specialized methods and ahead of Boltz-2 at 0.40.

#Applications

DOScore is aimed at structure-based hit discovery, where the question is which of billions of library compounds to prioritize for synthesis and assay. The authors ran prospective campaigns against four targets spanning distinct protein classes — the ectoenzyme CD73, the kinase IRAK4, the extended-substrate protease Factor XIa, and a protein-protein interface on IL-17A — each yielding chemically novel, biochemically and cellularly active inhibitors. On CD73, a target historically resistant to virtual screening, the DODock/DOScore workflow returned roughly a 30% hit rate against a prior large-scale machine-learning screen's 0.3%, a hundredfold improvement.

#Impact

DOScore is evidence that the long-standing accuracy ceiling in virtual screening reflects the limits of specific scoring formulations rather than a fundamental barrier, and that a single scoring checkpoint can hold its enrichment under simultaneous protein- and ligand-similarity filtering, where the baselines it is compared against lose most of theirs. The caveats are concrete: DOScore is a preprint that has not completed peer review; it scores a pose rather than generating one, so it depends on an upstream docking step such as DODock to place the ligand; its retrospective gains are tied to how strictly similarity is controlled; and no code or weights have been released — the system is exposed only through Deep Origin's commercial products.