MolMem: Memory-Augmented Agentic Reinforcement Learning for Sample-Efficient Molecular Optimization

Dimensions overall score 8.0

• Sample-Efficient Molecular Optimization Ziqing Wang1 Yibo Wen1 Abhishek Pandey2 Han Liu1 Kaize Ding1* 1Northwestern University 2AbbVie {ziqingwang2029, yibowen2024}@u.northwestern.edu abhishek.pandey@abbvie.com {hanliu, kaize.ding}@northwestern.edu Abstract In drug discoveryclaim 82%
  Evidencepartial

  Sample-Efficient Molecular Optimization Ziqing Wang1 Yibo Wen1 Abhishek Pandey2 Han Liu1 Kaize Ding1* 1Northwestern University 2AbbVie {ziqingwang2029, yibowen2024}@u.northwestern.edu abhishek.pandey@abbvie.com {hanliu

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• et al., 2024), ChemLLM (Zhang et al., 2024), PEIT-claim 82%
  Evidencepartial

  et al., 2024), ChemLLM (Zhang et al., 2024), PEIT-

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• 2025). Direct Retrieval and SFT-only also serve as ablations of MolMem. Notably, MolMem uses a compactQwen2.5-1.5Bbackbone, while most task- specific LLM baselines use 7–8B parameters. De- tails of these baselines are in Appendix B. Tasks and Constraints.We evaluate on five single-property tasks (QED, plogP,claim 82%
  Evidencepartial

  2025). Direct Retrieval and SFT-only also serve as ablations of MolMem. Notably, MolMem uses a compactQwen2.5-1.5Bbackbone, while most task- specific LLM baselines use 7–8B parameters

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• molecules randomly sampled from ZINC-250k (Ir- win and Shoichet, 2005). We enforce two practical constraints: (1) a Tanimoto similarity threshold of γ= 0.4 to preserve similarity to the lead, and (2) an oracle-call budget of B= 500 per lead molecule. Evaluation Metrics.Following prior work (Dey et al., 2025)claim 82%
  Evidencepartial

  molecules randomly sampled from ZINC-250k (Ir- win and Shoichet, 2005). We enforce two practical constraints: (1) a Tanimoto similarity threshold of γ= 0.4 to preserve similarity to the lead

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• widens on bioactivity targets: 50.5% to 96.0% onclaim 82%
  Evidencepartial

  widens on bioactivity targets: 50.5% to 96.0% on

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• require precise navigation of structure-activity rela- tionships, suggesting that MolMem is more effective on such complex landscapes. Smaller backbone, larger gains.Despite using a compact Qwen2.5-1.5B backbone, MolMem out- performs much larger baselines (ChemLLM 7B, GeLLM3O 8B, PEIT-LLM 8B)claim 82%
  Evidencepartial

  require precise navigation of structure-activity rela- tionships, suggesting that MolMem is more effective on such complex landscapes. Smaller backbone, larger gains.Despite using a compact Qwen2.5-1.5B backbone

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• Transformer) to maximize a scoring function un- der reward shaping. Apprentice baselines. • Direct Retrievaluses the same static retriever as MolMem to retrieve structurally similar molecules from the external database and selects the best candidate under the task objective, subject to the similarity constraintclaim 82%
  Evidencepartial

  Transformer) to maximize a scoring function un- der reward shaping. Apprentice baselines

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• task-specific LLM baselines, we use a shared prompting template (Appendix D) and a common parsing and validation pipeline. Each model pro- duces candidate SMILES strings, which are canon- icalized and checked for validity. Valid candidates are then evaluated by the oracle and counted to- ward the budgetclaim 82%
  Evidencepartial

  task-specific LLM baselines, we use a shared prompting template (Appendix D) and a common parsing and validation pipeline. Each model pro- duces candidate SMILES strings

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

Competitive landscape

A memory-augmented reinforcement learning agent for sample-efficient molecular optimization in drug discovery.