henhaobio
Sign in

Journal of Chemical Information and ModelingPharmacology & ToxicologyPeer-reviewed journal article

Large Language Model Agent for Modular Task Execution in Drug Discovery

Janghoon OckiD, Radheesh Sharma MedaiD, Srivathsan BadrinarayananiD, Neha S. Aluru, Achuth ChandrasekhariD, Amir Barati FarimaniiD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

We present a modular framework powered by large language models (LLMs) that automates and streamlines key tasks across the early-stage computational drug discovery pipeline. By combining LLM reasoning with domain-specific tools, the framework performs biomedical data retrieval, domain-specific question answering, molecular generation, property prediction, property-aware molecular refinement, and 3D protein-ligand structure generation. In a case study targeting BCL-2 in lymphocytic leukemia, the agent autonomously retrieved relevant biomolecular information--including FASTA sequences, SMILES representations, and literature--and answered mechanistic questions with improved contextual accuracy over standard LLMs. It then generated chemically diverse seed molecules and predicted 67 ADMET-related properties, which guided iterative molecular refinement. Across two refinement rounds, the number of molecules with QED > 0.6 increased from 34 to 55, and those passing at least four out of five empirical drug-likeness rules rose from 29 to 52, within a pool of 194 molecules. The framework also employed Boltz-2 to generate 3D protein-ligand complexes and provide rapid binding affinity estimates for candidate compounds. These results demonstrate that the approach effectively supports molecular screening, prioritization, and structure evaluation. Its modular design enables flexible integration of evolving tools and models, providing a scalable foundation for AI-assisted therapeutic discovery.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (5 Jul 2025) · CC BY 4.0

AgentD, a GPT-4o agent wired to retrieval, generative, ADMET and Boltz-2 tools, chains early drug-discovery steps for BCL-2 and raised molecules with QED > 0.6 from 34 to 55.

The paper says most computational drug-discovery methods are task-specific and need manual orchestration by experts. AgentD instead uses an LLM to coordinate retrieval, molecule generation, property prediction, refinement and 3D complex generation in one modular workflow.

Key results

  1. ShownTwo rounds of LLM-driven SMILES refinement raised overall drug-likeness of the BCL-2 molecule pool, measured by QED and by the number of empirical drug-likeness rules passed.

    Molecules with QED > 0.6: 34 originally, 49 after round 1, 55 after round 2. Molecules passing at least four of five rules: 29, then 44, then 52. Single target, no comparison with another refinement method. · Property-Aware Molecular Refinement; Figure legends; Abstract

  2. ShownRefinement helped permeability more than toxicity, and individual edits often failed: unintended SMILES changes, or correct edits that did not improve the property.

    Round 1 targeted 57 of the initial 100 molecules for low permeability: about 44% improved, 26% declined, 30% unchanged; round 2: 52%, 17%, 31%. Toxicity improved in 24% (round 1) and 20% (round 2) of high-toxicity molecules. · Property-Aware Molecular Refinement

  3. SuggestedRetrieval-augmented answers from AgentD were judged more mechanistically specific than plain GPT-4o answers on questions built from one venetoclax paper (Weller et al.).

    Qualitative side-by-side (Figure 4): the RAG answer names ISR activation, ATF3/ATF4-driven NOXA transcription and eIF2α phosphorylation, which GPT-4o omitted. No quantitative score, one source paper, questions written by the authors. · Domain-Specific Question Answering; Discussion

  4. SuggestedThe agent was run on three target-disease queries (BCL-2/lymphocytic leukemia, JAK2/myelofibrosis, KIT/renal cell carcinoma). The authors report it pulled the needed sequences and drug data, for example naming venetoclax as a BCL-2 inhibitor.

    Three queries shown in Figure 3. Sources used: UniProt, Serper web search, ChEMBL, Semantic Scholar. No accuracy rate reported. · Data Extraction; Data Extraction

  5. ShownThe two generators explored chemical space differently, and a shortlisted molecule went on to a Boltz-2 3D complex whose affinity estimates were only moderate.

    t-SNE of Mordred descriptors: REINVENT molecules widely spread, Mol2Mol molecules clustered near venetoclax. Example candidate: 4 of 5 rules passed, QED 0.68, predicted pKd 6.18, with predicted IC50 and inhibitor probability read as moderate binding. · Seed Molecule Generation; 3D Structure Generation; Figure legends

Methods, data and reproducibility

A LangChain agent using GPT-4o as its reasoning engine calls external tools for data retrieval, RAG question answering, molecule generation, property prediction and structure generation. SMILES refinement is done by the LLM alone, with no optimization tool. It is demonstrated in a BCL-2 case study.

  • BCL-2 / lymphocytic leukemia case study: 100 initial seed molecules from REINVENT and Mol2Mol, plus refinement-round additions · human (UniProt queries restricted to Homo sapiens) · n = 100 initial; 99 and 95 additional valid SMILES in rounds 1 and 2
  • Extra retrieval test queries: JAK2 for myelofibrosis and KIT for renal cell carcinoma · human (UniProt queries restricted to Homo sapiens) · n = 2 queries
  • Reference paper by Weller et al. on venetoclax, ISR and NOXA, used to build QA questions · n = 1 paper
  • Randomly sampled ChEMBL compounds used as background in the chemical-space plot · n = 5,000
Methods & tools
GPT-4o, LangChain, UniProt, ChEMBL, Semantic Scholar, Serper API, REINVENT, Mol2Mol, Deep-PK, BAPULM, Boltz-2, RDKit
Reproducibility gaps
No versions, requirements or hardware are stated in the text. No dataset deposit is described. The full ADMET property list and risk-property list are in supplementary tables (S1, S2) not included here. The drug-likeness rule thresholds are given in the Methods text.
Limitations (5)
  • Boltz-2 IC50 and inhibitor-probability outputs are described as not highly accurate; they serve only as rough guides for initial prioritization. · stated by the authors
  • LLM refinement is unreliable per molecule: edits can diverge from the stated intent, or succeed without improving the property. Improving one property can worsen others (permeability gains lowered logP and logS), so multi-objective optimization is needed. · stated by the authors
  • RAG quality depends on the retrieved literature, which is limited to open-access papers; the shortlisting filters (rule counts, QED > 0.55, pKd > 6.0) are illustrative rather than target-specific. · stated by the authors
  • Evidence rests on one target (BCL-2) with predicted properties only; there is no experimental validation, no baseline refinement method and no quantitative QA metric. · noted in this summary
  • The abstract's pool of 194 molecules is not reconciled with 100 initial molecules plus 99 and 95 added valid SMILES reported in the results text. · noted in this summary

Read this if you want a concrete example of an LLM agent chaining retrieval, generation, ADMET prediction and Boltz-2 structure steps, and want its refinement success rates and failure modes.

Citations
5
FWCI
18.03
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISVichentijev '25Liu '26Vangala '26Igashov '26Liang '24Smbatyan '26
75% semantic matchBioinformatics
Ivana Vichentijevikj, Kostadin Mishev, Monika Simjanoska·Bioinformatics Advances·Dec 20, 2025· 6 citations