{
  "id": 14150260,
  "title": "Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?",
  "url": "https://urgent.news/2026/10/13/do-llm-proposers-improve-budgeted-search-for-ionizable-lipids",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-13T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.10.06.756984v1?rss=1"
  },
  "original_language": "en",
  "account": "Efficient design of lipids with constrained experimental budgets remains a key challenge in computational chemistry and drug discovery. Introducing LEAF (Lipid Evolutionary Adaptive Framework), a matched-budget protocol, researchers sought to determine if language-model proposers can enhance combinatorial molecular search compared to classical baselines. Employing a 30,044-candidate ionizable-lipid space, a frozen Gaussian-process oracle with potency labels, and an active-learning surrogate, four language models evaluated across eight seeds and three splits showed no significant improvement over genetic-algorithm or random-search baselines. For the Llama 8B model, 79.1% of acquisition-selected queries duplicated previously evaluated archive members. Removing duplicates lessens the performance gap on certain splits but raises proposal costs and risks exhausting the budget. Larger models decrease duplication without proving superiority or equivalence to others. Implementing exploratory scored prompts and stricter novelty controls resulted in positive median differences against the genetic algorithm, although wide intervals and dependence on shared seeds limit stronger conclusions. A validated single-assay reanalysis maintains the chemotype-cold deficit at a smaller numerical scale, and exploratory expected hypervolume improvement does not negate the result. Subsequent checks explored whether proposers truly utilize injected evidence, employing zero-shot leakage probes, rank-only and chain-of-thought rationale prompts, score-shuffle perturbations, evidence-conflict probes, and reasoning-mode proposers. These checks revealed measurable engagement with supplied information without corresponding performance gains. LEAF isolates proposal quality as the experimental variable, shedding light on a specific failure mode of zero-shot LLM proposers under controlled budgets.",
  "summary": "Efficient lipid design under limited experimental budgets is a central challenge in computational chemistry and drug discovery. We introduce LEAF (Lipid Evolutionary Adaptive Framework), a matched-budget protocol for testing whether language-model proposers improve combinatorial molecular search relative to classical baselines. The study uses a 30,044-candidate ionizable-lipid space, a frozen…",
  "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."
}