Research map / Graph learning / Graph algorithms and clustering

Amortized Molecular Optimization via Group Relative Policy Optimization

NeurIPS 2026 · Graph learning › Graph algorithms and clustering · approach: Neural combinatorial optimization on graphs · 0 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 AMORTIX, a Graph Transformer policy trained with group-relative RL (GRPO/Dr. GRPO) that generates structurally constrained molecular optimizations in a single forward pass with zero inference-time oracle calls.

Problem

Structurally constrained molecular optimization is dominated by instance optimizers that restart costly oracle-driven search for each starting structure; existing amortized policies fail to generalize to diverse structural constraints and struggle with heterogeneous difficulty across starting struct

Approach

Adapts GRPO to amortized molecular design by sampling a group of completions per starting structure and normalizing rewards within that group, so advantages are compared only against peers sharing the same input structure; the policy is a Graph Transformer with valence-based action masking and hierarchical atom/bond actions.

Key results

• On PMO, AMORTIX reaches a cumulative top-10 AUC sum of 16.433, first among amortized methods and second overall behind GenMol (17.540).
• On scaffold decoration, AMORTIX scores 0.775 GSK3B, 0.796 JNK3 and 0.619 Kinase MPO versus 0.410/0.260/0.455 for LibINVENT and 0.735/0.659/0.461 for Mol GA.
• Amortized training (~50,000 oracle calls) is recouped after about 100 test scaffolds against Mol GA on Kinase MPO; 52 held-out prodrug parents yield 16 pure prodrugs (31%).

Surprising findings

• REINFORCE with a global baseline matches AMORTIX on single-objective tasks but is severely unstable on Kinase MPO (0.475 ± 0.120).
• GenMol and Mol GA, despite strong single-objective scores, fall well behind on Kinase MPO by exploiting easy QED/SA components.

Limitations

• Operates only on 2D topologies without stereochemistry or 3D conformations.
• Assumes a single connected starting scaffold and only additive actions (no deletions or substitutions).
• Training still requires tens of thousands of oracle calls.

Experimental setup

Datasets: ChEMBL v35 ZINC-250k TDC GSK3B/JNK3 oracles TDC BBB_Martins

Benchmarks: PMO Kinase MPO scaffold decoration (GSK3B, JNK3, Kinase MPO) Prodrug transfer case study

Metrics: top-10 AUC objective score oracle call break-even advantage standard deviation

Baselines: LibINVENT GraphXForm DrugEx v3 Mol GA GenMol f-RAG Genetic GFN REINVENT SynGBO GP BO AMORTIX-REINFORCE AMORTIX-PPO

Builds on: GRPO Dr. GRPO GraphXForm REINVENT PMO benchmark

Closest papers by meaning

• Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.87)
• Amortized Active Causal Induction with Deep Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
• Amortized Sampling with Transferable Normalizing Flows (NeurIPS 2025 · similarity 0.86)
• DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.85)
• On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization (NeurIPS 2026 · similarity 0.85)
• ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition (NeurIPS 2025 · similarity 0.85)