{
  "id": 117100,
  "title": "Overcoming the accuracy-generalization tradeoff in docking and scoring for prospective virtual screening",
  "url": "https://urgent.news/2026/08/03/overcoming-the-accuracy-generalization-tradeoff-in-docking-and",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-03T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.08.03.742480v1?rss=1"
  },
  "original_language": "en",
  "account": "The article discusses the limitations of classical docking and scoring methods in virtual screening, which typically yield only a limited number of hits. These methods are generalizable but lack accuracy due to simplistic functional forms and insufficient parameterization. Conversely, recent machine learning approaches are highly expressive but struggle to generalize to new molecules and pockets, with reported accuracy often inflated by training-test leakage. To overcome these challenges, the authors introduce DODock and DOScore, hybrid frameworks that combine machine learning with physics-based methods. These hybrid frameworks maintain expressive power while demonstrating superior generalization performance when applied to novel, out-of-distribution molecules and protein pockets. The effectiveness of DODock and DOScore was verified through prospective testing, including blind prediction of a drug candidate's binding pose to PCSK9 with an RMSD of only 1.2 angstroms. Additionally, the authors employed DODock and DOScore in virtual screening campaigns targeting four diverse therapeutic targets, including an ectoenzyme (CD73), a kinase (IRAK4), an extended-substrate protease (FXI), and an allosteric protein-protein interface (IL17). These screening efforts led to the discovery of numerous chemically novel, biochemically active, and cell-based inhibitors. Particularly noteworthy was the CD73 target, which had historically yielded poor results in virtual screening attempts. The use of DODock and DOScore in CD73 screening resulted in a hundredfold improvement in hit rate compared to a recent machine-learning screening effort. The authors conclude that the long-standing plateau in virtual screening accuracy may not be an inherent limitation, but rather a result of inadequate methodologies. They posit that structure-based exploration of ultralarge chemical spaces may ultimately prove to be a viable primary approach for discovering novel chemical entities, thereby revolutionizing drug discovery strategies.",
  "summary": "Virtual screening promises access to tens of billions of synthetically accessible, diverse compounds, yet it is rarely used as a primary hit-discovery strategy in contemporary drug-discovery campaigns. We argue that this gap reflects the real-world underperformance of the underlying docking and scoring methods: classical docking is generalizable but limited in accuracy by simple functional forms…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}