LIDDIA: Language-based Intelligent Drug Discovery Agent

1 Introduction

• Drug discovery is complex and nonlinear because satisfying one molecular requirement can violate another, making chemists’ iterative analysis and optimization time-consuming.Successful drugs must bind targets and exhibit suitable physicochemical, pharmacodynamic, and pharmacokinetic properties.Source: §1 · p. 1
• LIDDIA combines four interconnected components to navigate pre-clinical drug discovery in silico: REASONER, EXECUTOR, EVALUATOR, and MEMORY.The agent orchestrates computational tools including docking simulation, property prediction, and molecule optimization.Source: §1 · p. 1
• LIDDIA integrates generative AI for molecular design to explore chemical space beyond conventional libraries and uses a modular architecture designed for extension.The authors describe it as, to their knowledge, the first effort toward low-cost, high-efficiency autonomous drug discovery.Source: §1 · p. 2
• More than 70% of 30 major therapeutic targets had promising candidates meeting key pharmaceutical properties with LIDDIA.The paper also reports that LIDDIA balances exploration and exploitation and identifies a promising candidate for AR/NR3C4.Source: §1 · p. 2

2 Related Work

• Existing biomedical LLM agents address tasks such as multi-omics analysis, de novo protein design, and materials-science laboratory automation.The cited examples include AutoBA, PROTAGENT, and A-LAB.Source: §2 · p. 2
• COSCIENTIST plans and controls organic chemistry experiments using web search, programming, and physical and computational tools, but lacks domain-specific grounding tools.Its grounding relies on the LLM’s chemistry knowledge, web search, and physical-experiment results.Source: §2 · p. 2
• CHEMCROW, CACTUS, and DRUGAGENT support chemistry tasks including molecule generation, property prediction, retrosynthesis planning, and drug repurposing.The authors state that none of these agents is grounded with established computational tools for novel structure-based drug discovery.Source: §2 · p. 2

3 LIDDIA Framework

• The framework combines computational tools and reasoning through REASONER, EXECUTOR, EVALUATOR, and MEMORY to navigate drug discovery.Its stated goal is to return diverse, high-quality molecules meeting a target’s property specifications.Source: §3 · p. 2, Figure 1 · p. 3

3.1 REASONER

• REASONER uses MEMORY’s molecule and property information to choose GENERATE, OPTIMIZE, or SCREEN actions during iterative molecular design.These actions correspond to hit identification and lead optimization steps.Source: §3.1 · p. 3

3.2 EXECUTOR

• EXECUTOR implements REASONER’s actions with structure-based generation, molecular-refinement models, and a processor for screening and managing molecules.Generative models support both hit identification and lead optimization, enabling de novo molecular generation beyond known molecules.Source: §3.2 · p. 3

3.3 EVALUATOR

• EVALUATOR assesses binding affinity and molecular properties, stores results in MEMORY, and signals termination when user requirements are met.Its plug-and-play design allows additional or updated tools for new property requirements.Source: §3.3 · pp. 3–4

3.4 MEMORY

• MEMORY retains user inputs, action trajectories, generated molecules, and their properties, then supplies this evolving information to REASONER for planning.The repository is continuously updated as the drug-discovery process progresses.Source: §3.4 · p. 4

4.1 Evaluation Metrics

• High-quality molecules must meet target-average QED, LRF, SAS, and VNA thresholds and NVT ≥0.8; generated sets count as diverse at DVS ≥0.8.Source: §4.1 · p. 4
• TSR is the percentage of targets for which a method generates a diverse set containing at least five high-quality molecules.Source: §4.1 · p. 4

4.2 Protein Target Dataset

• The manually curated benchmark contains 30 disease-associated protein targets with resolved PDB structures, binding-pocket ligands, and known drugs for comparison.Source: §4.2 · pp. 4–5

4.3 Implementation Details

• Claude 3.5 Sonnet powers REASONER and EVALUATOR, with at most 10 actions; Pocket2Mol generates molecules and GraphGA optimizes QED, SAS, or VNA.Source: §4.3 · p. 5

4.4 Baselines

• Baselines include Pocket2Mol and DiffSMol molecule generators, plus GPT-4o, o1, o1-mini, and Claude 3.5 Sonnet general-purpose LLMs.Source: §4.4 · p. 5

5.1 Main Results

• 73.3% TSR versus Pocket2Mol’s 23.3% and 85% HQ versus GPT-4o’s 35% show LIDDIA’s advantage in target success and multi-property molecule quality.Table 2 compares baseline methods with LIDDIA.Source: §5.1 · p. 5, Table 2 · p. 6
• 97.2% of molecules per target exceeded average QED and 95.8% exceeded average VNA; 97.8% were novel, second only to DiffSMol.The reported comparisons are relative to target averages for QED and VNA, and to other methods for novelty.Source: §5.1 · p. 6

5.2.1 LIDDIA action patterns

• LIDDIA typically follows GENERATE with OPTIMIZE and SCREEN; about 90% of generated molecules proceed to optimization before screening.Optimization is used when generated molecules do not meet all property requirements, while SCREEN selects candidates.Source: §5.2.1 · p. 7
• Output molecules roughly double initial-pool QED and VNA, but top candidates often have limited structural diversity; later steps favor OPTIMIZE or SCREEN.Figure 3 shows molecule quality in its left panel and actions in its right panel.Source: Figure 3 · p. 6, §5.2.1 · p. 7

5.2.2 Analysis on LIDDIA action trajectories

• Across target trajectories, LIDDIA adapts actions to evaluator feedback, iteratively refining and screening toward candidates, though some targets remain unsuccessful after 10 iterations.CDK5 illustrates iterative recovery; PIK3CA, MET, and ADRB2 are examples of suboptimal outcomes.Source: §5.2.2 · p. 8, Figure 4 · p. 8

5.3 Ablation Study

• Reasoning-enabled LIDDIA exceeded the no-reasoning loop’s target success rate by more than 40 percentage points; Claude 3.5 and DeepSeek-R1 each reached 73% TSR, with different HQ and diversity rates.DeepSeek-R1’s HQ rate was >99% versus Claude’s 84%, while its diversity rate was 77% versus Claude’s 90%.Source: §5.3 · p. 9

5.4 Case Study on AR/NR3C4

• For AR/NR3C4, NL-1 had better QED, VNA, and SAS than the ligand and at least one approved drug; its -8.81 kcal/mol VNA and binding interactions support its promise as a candidate.NL-1 passed several computational filters; its carboxylic acid forms hydrogen bonds in the pocket.Source: Figure 6 · p. 8, Figure 7 · p. 8, §5.4 · p. 9

• LIDDIA is presented as an agent for autonomous drug discovery, with its capabilities examined across therapeutic targets.Source: §6 · p. 9
• The examination reveals key insights into LIDDIA’s success and its performance across major therapeutic targets.Source: §6 · p. 9
• On EGFR, the paper investigates generated molecules and reports their potential as drug candidates.Source: §6 · p. 9

• The study evaluates LIDDIA only in silico; wet-lab validation is planned for follow-up research.Source: §8 · p. 10
• Evaluation covers a few key pharmaceutical properties, while the authors plan to extend the agent to additional metrics.Source: §8 · p. 10
• Experiments use a limited number of API calls because of budget constraints.Source: §8 · p. 10
• Benchmarking covers a small set of clinically relevant targets, reflecting the lack of large, well-annotated benchmarks for these tasks.The authors describe this as a trade-off between scope and substance and plan to expand the dataset and benchmarking.Source: §8 · p. 10

Ethics Statement

• LIDDIA functions only in silico and does not currently generate synthesis plans for its molecules.Source: Ethics Statement · p. 10
• The EVALUATOR and example prompts focus on priority human-benefit metrics such as QED and LRF.Source: Ethics Statement · p. 10
• The authors cannot guarantee that LIDDIA will avoid harmful or incorrect content.Source: Ethics Statement · p. 10

A Dataset

• The dataset table lists protein targets and indicates disease categories including autoimmune disease, cancer, cardiovascular disease, diabetes, infectious disease, and neurological conditions.Source: A Dataset · p. 14
• The table headings identify gene name, PDB identifiers, pocket and ligand structures, and disease category.Source: A Dataset · p. 14

A.1 On the Necessity of Dataset Curation

• Existing benchmarks include few human proteins and may lack associated drugs needed to reference multi-property requirements; the authors curate 30 human proteins linked to major diseases and known drugs.Source: §A.1 · pp. 19–20

B.1 REASONER Prompt

• The REASONER prompt supplies molecules, pockets, available actions, requirements, action history, and prior evaluation results, and asks it to select one action and its input given remaining resources.Source: §B.1 · p. 20
• The prompt requests stepwise reasoning and formats the response with an action name and input identifier; CODE actions also require a descriptive explanation of at most three sentences.Source: §B.1 · p. 20

B.2 EVALUATOR Prompt

• The EVALUATOR prompt asks whether every molecule in a pool satisfies the specified requirements and requests a brief reason with a YES-or-NO answer.Source: §B.2 · p. 20

B.3 SCREEN Prompt

• The SCREEN prompt specifies a Python function receiving dataframes with molecule and property columns and returning a dataframe with the same columns.The listed columns include SMILES, QED, SAScore, Lipinski, Novelty, and Vina Score; the task is supplied as an input description.Source: §B.3 · pp. 20–21
• The SCREEN prompt also requires an output format, import and code sections, and necessary library imports; the toxicity table caption says lower values indicate a better profile.Source: §B.3 · p. 21, Table A2 · p. 21

C.1 Additional Baselines

• Both LIDDIA variants significantly outperform TargetDiff and DecompDiff; unlike those methods, LIDDIA generates new binding molecules better than existing drugs.Table A3 compares task-specific molecule-generation methods, general-purpose LLMs, and LIDDIA variants.Source: §C.1 · p. 21

C.2 Toxicity Predictions

• Toxicity predictions used ADMET-AI and selected reference properties from Chen et al. (2025) that were available in that tool.The selected properties were compared with drugs in the dataset.Source: §C.2 · p. 21
• Generated molecules are better than or comparable to known drugs on toxicity properties, although LIDDIA is designed for high-quality—not specifically safe—molecules.The authors describe the resulting safety profiles as promising and identify room for improvement.Source: §C.2 · p. 21

C.3 Failure Analysis

• Some targets have fewer than five high-quality molecules or insufficient diversity, despite LIDDIA generating at least one high-quality molecule for every target.The authors hypothesize that pocket structure can restrict the scaffolds that bind, making diverse high-quality molecules difficult to generate.Source: §C.3 · pp. 21–22
• Figure A1 compares EGFR molecules with three drugs and a binding ligand on VNA, SAS, and QED; red squares mark metrics where a LIDDIA molecule performs better.Figure A2 identifies NL-2's two enol groups and NL-3's fulvene as problematic candidate features.Source: Figure A1 · p. 22, Figure A2 · p. 22

C.4 Computational Costs

• LIDDIA uses about 9K input tokens and 5K output tokens per target, with an estimated generation cost of around $0.03 using Claude pricing.The estimate is based on the number of tokens used in the experiments.Source: §C.4 · p. 22

C.5 Case Study on EGFR

• LIDDIA's EGFR molecules surpass the native ligand on VNA and QED while showing profiles comparable to approved drugs.Some molecules outperform Olmutinib and Masoprocol on VNA, QED, and SAS.Source: §C.5 · pp. 22–23
• NL-2 uses hydrogen bonds on opposite sides of the pocket, whereas NL-3 relies on hydrophobic packing and shape complementarity.NL-3's fluorine contacts hydrophobic residues near the pocket entrance, while its diarylketone is buried in the pocket.Source: §C.5 · p. 23, Figure A3 · p. 23
• NL-2's enol groups and NL-3's fulvene raise drug-candidate concerns; the diarylketone is also known to be phototoxic.The authors note that favorable binding can conflict with other drug-relevant properties.Source: §C.5 · p. 23
• No standalone in-silico evaluation tool detects all the identified issues; Lilly rules identify NL-2's enols but do not flag NL-3.PAINS, BRENK, and NIH filters also fail to capture the problematic features in NL-2.Source: §C.5 · p. 23
• The authors identify human-in-the-loop validation, improved in-silico tools, and wet-lab validation as priorities for practical deployment.They state that in-vitro and in-vivo laboratory work will be necessary to assess real-world impact.Source: §C.5 · pp. 23–24