Research map / Large language models / Agents, code and math

Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback

AAAI 2024 · Large language models › Agents, code and math · approach: LLMs for scientific discovery · 15 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LMLF, an iterative procedure that separates drug-design domain constraints (logical form) from a simple text query and refines the constraints via logical feedback to steer GPT-3 and PaLM toward generating high-binding-affinity lead molecules for JAK2 and DRD2.

Problem

LLM-based lead generation for drug targets depends heavily on manual prompt engineering, making results subjective and hard to reproduce.

Approach

Keep the query fixed and move domain knowledge into formal logical constraints that are automatically generalised across iterations (MIMIC-inspired), with newly generated molecules checked against the constraints to label feedback that updates the prompt's background knowledge.

Key results

• GPTLF++ achieved mean JAK2 binding affinity 7.74 (median 7.71) versus 6.50 for unconstrained GPTLF and 6.53 for VAE-GNN baseline.
• On DRD2, GPTLF++ achieved mean affinity 7.66 vs 6.48 for GPTLF without constraints.
• Chemists found 15% of JAK2 and 40% of DRD2 generated molecules contained target-selective functional groups; 10 of 11 such molecules had Tanimoto similarity below 0.75 to known inhibitors.

Measured effects reported in the paper

Surprising findings

• PaLM's expert-assessed results were comparable to or slightly better than GPT-3 on novel/effective inhibitors despite lower quantitative affinity, though differences were not statistically significant.

Limitations

• Qualitative assessment used only 20 molecules judged by chemists, with differences between GPT-3 and PaLM not statistically significant.
• Comparison with VAE-GNN baseline is not like-for-like: LLMs have access to more information and the baseline lacked binding-affinity constraints.
• Relies on closed API LLMs (text-davinci-003, text-bison-001) with temperature 0.7, limiting reproducibility.

Experimental setup

Datasets: ChEMBL JAK2 labelled molecules ChEMBL DRD2 labelled molecules known JAK2 and DRD2 inhibitors

Benchmarks: JAK2 binding affinity (GNINA docking) DRD2 binding affinity (GNINA docking)

Metrics: mean/median docking binding affinity Mann-Whitney U test on affinity histograms substructure match of selective functional groups Tanimoto similarity novelty expert chemist assessment

Baselines: Known inhibitors VAE-GNN GPTLF (no constraints) GPTLF+ (target-agnostic constraints)

Builds on: MIMIC Chemical VAE-GNN conditional generation (Dash et al. 2021)

Closest papers by meaning

• LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.87)
• GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · similarity 0.86)
• MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning (ICML 2025 · similarity 0.86)
• Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.86)
• MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 2025 · similarity 0.85)
• Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.85)