Research map / Reinforcement learning / Visual and representation RL
Beam Enumeration: Probabilistic Explainability For Sample Efficient Self-conditioned Molecular Design
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 Beam Enumeration, which exhaustively enumerates top-k token sub-sequences from language-based molecular generators to extract substructures that then self-condition RL-based generation, improving sample efficiency of Augmented Memory and REINVENT on PMO and docking tasks.
Problem
Sample efficiency and explainability in generative molecular design: existing language-based generators need many oracle calls, and their learned knowledge is not easily interpretable or actionable during a design run.
Approach
Enumerate the top-k most probable token sub-sequences of a SMILES generator's policy, extract frequent substructures (Bemis-Murcko scaffolds or any valid substructure) from them, and filter future sampled batches to contain these substructures so only those molecules are sent to the oracle.
Key results
- Augmented Memory with Beam Enumeration (Scaffold 15) yields 425 vs 85 molecules above 0.8 reward (illustrative MPO task), with Oracle Burden(100) at 0.8 reduced from 4827 to 2704 calls.
- In drug discovery case studies, Yield above 0.8 improves more than 4x for Augmented Memory across targets, with MK2 improving 29x.
- Beam Enumeration reaches 100 molecules above 0.8 reward for MK2 in about 2,189 oracle calls where baseline Augmented Memory fails.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Beam Enumeration with self-conditioned substructure filtering | Augmented Memory · Illustrative MPO, reward >0.8 | Generative Yield: 85 $\pm$ 56 -> 819 $\pm$ 291 | Helped |
| Add Beam Enumeration with Structure extraction and minimum size 15 | Augmented Memory · DRD2, reward >0.8 | Generative Yield: 363 $\pm$ 195 -> 1780 $\pm$ 439 | Helped |
| Add Beam Enumeration with Structure extraction and minimum size 15 | Augmented Memory · MK2, reward >0.8 | Generative Yield: 34 $\pm$ 13 -> 987 $\pm$ 211 | Helped |
| Add Beam Enumeration with Structure extraction and minimum size 15 | Augmented Memory · AChE, reward >0.8 | Generative Yield: 556 $\pm$ 47 -> 2059 $\pm$ 327 | Helped |
| Add Beam Enumeration with Structure extraction and minimum size 15 | Augmented Memory · MK2, reward >0.8, 100 molecules | Oracle Burden: Failed -> 2189 $\pm$ 181 | Helped |
| Enforce a minimum extracted Structure size of 15 (ablation) | Augmented Memory + Beam Enumeration · Illustrative MPO, reward >0.8 | Generative Yield: 199 $\pm$ 122 -> 700 $\pm$ 389 | Helped |
| Enforce a minimum extracted Structure size of 15 (ablation) | Augmented Memory + Beam Enumeration · Illustrative MPO, reward >0.8, 100 molecules | Oracle Burden: 4415 $\pm$ 437 (20) -> 2943 $\pm$ 811 (6) | Helped |
| Use Structure extraction instead of Scaffold extraction (ablation) | Augmented Memory + Beam Enumeration, minimum size 15 · DRD2, reward >0.8 | Generative Yield: 1607 $\pm$ 379 -> 1780 $\pm$ 439 | Helped |
| Use Structure extraction instead of Scaffold extraction (ablation) | Augmented Memory + Beam Enumeration, minimum size 15 · MK2, reward >0.8 | Generative Yield: 523 $\pm$ 438 -> 987 $\pm$ 211 | Helped |
| Use Structure extraction instead of Scaffold extraction (ablation) | Augmented Memory + Beam Enumeration, minimum size 15 · AChE, reward >0.8 | Generative Yield: 2124 $\pm$ 326 -> 2059 $\pm$ 327 | No clear effect |
| Use sample token sampling instead of topk sampling (ablation) | Beam Enumeration · Illustrative MPO | Generative Yield | Mixed |
| Use larger extracted substructures via Structure Minimum Size (ablation) | Beam Enumeration · Illustrative MPO hyperparameter search | Generative Yield | Helped |
Surprising findings
- Enforcing larger minimum substructure size improves performance across all hyperparameter combinations, suggesting strong bias toward extracted substructures does not hurt sample efficiency.
- Sampling tokens instead of taking top-k sometimes performs as well or better, though with higher variance.
Limitations
- Exhaustive enumeration grows exponentially with beam steps, limiting scalability.
- Reduced diversity (IntDiv1) as a trade-off for sample efficiency.
- Extracted substructures explain the oracle rather than the true target objective.
Experimental setup
Datasets: ChEMBL
Benchmarks: Practical Molecular Optimization (PMO)
Metrics: Generative Yield Oracle Burden IntDiv1 Vina docking score QED SA score
Baselines: Augmented Memory REINVENT
Builds on: Augmented Memory REINVENT Beam Search Bemis-Murcko scaffold
Closest papers by meaning
- Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.86)
- Procedural Synthesis of Synthesizable Molecules (ICLR 2025 · similarity 0.86)
- Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.86)
- Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.86)
- Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.85)
- GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.