Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?
- 1Department of Computer Science and Engineering, Shiv Nadar University, Delhi-NCR, India
Abstract
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 Gaussian-process oracle fitted to experimental potency labels, and a separate active-learning surrogate. Four language models evaluated over eight seeds and three splits show no significant improvement over genetic-algorithm or random-search baselines. For Llama 8B, 79.1% of acquisition-selected queries repeat previously evaluated archive members. Removing duplicates reduces the performance gap on some splits but increases proposal cost and can prevent budget exhaustion. Larger models reduce duplication without establishing superiority or equivalence. Exploratory scored prompts and stricter novelty controls yield positive median differences against the genetic algorithm, yet wide intervals and shared-seed dependence preclude stronger conclusions. A validated single-assay reanalysis preserves the chemotype-cold deficit at a smaller numerical scale, and exploratory expected hypervolume improvement does not reverse the result. Follow-up checks address whether proposers genuinely use injected evidence. These checks include a zero-shot leakage probe, rank-only and chain-of-thought-rationale prompts, a score-shuffle perturbation, an evidence-conflict probe, and a genuine reasoning-mode proposer. They show measurable engagement with supplied information without a corresponding performance gain. LEAF isolates proposal quality as the experimental variable and documents a concrete failure mode of zero-shot LLM proposers under controlled budgets.
Related
- Cyclic-thiourea ionizable lipids program multilamellar mRNA-LNPs for sustained expression, frozen-storage stability and potent vaccine immunotherapy
- CADS: Measuring the Incremental Contribution of LLM Proposers in Sequential Biological Screens
- A cytotoxicity dataset for data-driven design of ionizable lipid nanoparticles
- A new class of lipid transfer proteins is required for the recycling of lipids from the P. falciparum digestive vacuole
- Bioengineered algal lipids enriched in structured medium- and long-chain triacylglycerols, linoleate, and sn-2 palmitate for human milk fat substitutes
BibTeX
@misc{singh2026proposers,
title = {{Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?}},
author = {H. Singh and R. K. SINGH and C. S. Krishna and S. P. Srivastava and R. Gorantla},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.06.756984},
url = {https://www.biorxiv.org/content/10.64898/2026.10.06.756984v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?
AU - H. Singh
AU - R. K. SINGH
AU - C. S. Krishna
AU - S. P. Srivastava
AU - R. Gorantla
PY - 2026
DA - 2026/10/06
PB - bioRxiv
DO - 10.64898/2026.10.06.756984
UR - https://www.biorxiv.org/content/10.64898/2026.10.06.756984v1
AB - 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 Gaussian-process oracle fitted to experimental potency labels, and a separate active-learning surrogate. Four language models evaluated over eight seeds and three splits show no significant improvement over genetic-algorithm or random-search baselines. For Llama 8B, 79.1% of acquisition-selected queries repeat previously evaluated archive members. Removing duplicates reduces the performance gap on some splits but increases proposal cost and can prevent budget exhaustion. Larger models reduce duplication without establishing superiority or equivalence. Exploratory scored prompts and stricter novelty controls yield positive median differences against the genetic algorithm, yet wide intervals and shared-seed dependence preclude stronger conclusions. A validated single-assay reanalysis preserves the chemotype-cold deficit at a smaller numerical scale, and exploratory expected hypervolume improvement does not reverse the result. Follow-up checks address whether proposers genuinely use injected evidence. These checks include a zero-shot leakage probe, rank-only and chain-of-thought-rationale prompts, a score-shuffle perturbation, an evidence-conflict probe, and a genuine reasoning-mode proposer. They show measurable engagement with supplied information without a corresponding performance gain. LEAF isolates proposal quality as the experimental variable and documents a concrete failure mode of zero-shot LLM proposers under controlled budgets.
ER -