{
  "id": 6097781,
  "title": "OmniSyn unifies target-aware molecular generation and optimization within a synthesis-native LLM framework across the human proteome",
  "url": "https://urgent.news/2026/09/06/omnisyn-unifies-target-aware-molecular-generation-and-optimization",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-06T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.09.02.748775v1?rss=1"
  },
  "original_language": "en",
  "account": "OmniSyn, a language model developed by researchers, enables the simultaneous creation and optimization of bioactive molecules targeting the human proteome. This innovation addresses the challenge of designing target-specific molecules with feasible synthesis routes. OmniSyn utilizes a protein-sequence-conditioned Mixture-of-Experts (MoE) architecture, incorporating a task-conditioned interaction module and a synthesis-action decoder. This design allows for the generation of molecules with explicit synthesis traces, merging de novo ligand generation, synthesizability projection, and hit-to-lead (H2L) optimization within a chemically traceable space.\n\nThe model is pre-trained through self-distillation and subsequently refined using task-specific reinforcement learning (RL) to refine expert routing and optimize molecular attributes across various design modes. When tested on unseen protein targets, OmniSyn demonstrated superior performance in de novo design and H2L optimization compared to existing benchmarks. Notably, it achieved retrosynthetic success rates of 68.47% for de novo generation and 71.92% for H2L optimization, outperforming the leading alternatives by 61.3% and 184.2%, respectively. This model also excels in synthesizability projection, enhancing the feasibility of generated molecules while maintaining their similarity to original structures.\n\nApplying OmniSyn on a scale relevant to the human proteome, which includes over 21,000 targets, resulted in the construction of the largest virtual library of target-specific molecules. This library, totaling 2.7 billion members, each tag along with synthesis traces and prioritization scores. This achievement opens new avenues for exploring therapeutic targets, particularly those without experimentally resolved structures, thereby expanding the potential for discovering novel and effective drugs.",
  "summary": "Designing target-specific bioactive molecules with actionable synthesis routes for the human proteome holds enormous potential for expanding therapeutic discovery, but remains a challenge. Existing target-aware generative models often depend on protein structures and generate molecules before assessing synthetic feasibility. Here we present OmniSyn, a protein-sequence-conditioned…",
  "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."
}