Research map / Probabilistic and scientific modeling / Molecules and proteins
InversionGNN: A Dual Path Network for Multi-Property Molecular Optimization
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 InversionGNN, a dual-path GNN whose frozen property predictor is inverted via gradient-based Pareto search over differentiable scaffolding trees to generate multi-objective molecules sample-efficiently.
Problem
Multi-objective molecular optimization where properties conflict or correlate, and existing methods underuse learned chemical knowledge, rely on linear scalarization, or are costly (GA, Bayesian optimization).
Approach
A dual-path GNN: a direct property-prediction path trained on Oracle labels, and an inversion path that back-propagates gradients to edit a differentiable scaffolding tree using a non-dominated (QP-based MGDA-style) descent direction relaxed to discrete chemical space, with weight-conditioned search across weight vectors to cover the Pareto front.
Key results
- On the synthetic task, InversionGNN reaches HV 0.328 versus 0.202 for the best baseline (GAN-GA) and 0.071 for I-LS.
- On GSK3β+JNK3 drug discovery, InversionGNN achieves APS 0.841 (best) and HV 0.763 versus 0.592 for HN-GFN.
- On the weight-conditioned task, InversionGNN reports APS 0.648 and NU 0.026, with diversity 0.435 lower than HN-GFN's 0.571.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InversionGNN instead of GAN-GA | Synthetic multi-objective optimization · Synthetic task | HV: 0.202_{\pm 0.017} -> 0.328_{\pm 0.001} | Helped |
| Replace linear scalarization with gradient-based Pareto inversion (ablation) | InversionGNN · Weight-conditioned molecular optimization | APS: 0.529_{\pm 0.006} -> 0.648_{\pm 0.012} | Helped |
| Replace linear scalarization with gradient-based Pareto inversion (ablation) | InversionGNN · Weight-conditioned molecular optimization | NU: 0.049_{\pm 0.002} -> 0.026_{\pm 0.001} | Helped |
| Replace linear scalarization with gradient-based Pareto inversion (ablation) | InversionGNN · Weight-conditioned molecular optimization | Div: 0.541_{\pm 0.007} -> 0.435_{\pm 0.009} | Hurt |
| Use InversionGNN instead of HN-GFN | Multi-objective drug discovery · GSK3β + JNK3 | HV: 0.592_{\pm 0.042} -> 0.763_{\pm 0.031} | Helped |
| Use InversionGNN instead of DST | Multi-objective drug discovery · GSK3β + JNK3 | APS: 0.827 -> 0.841 | Helped |
| Use InversionGNN instead of HN-GFN | Multi-objective drug discovery · GSK3β + JNK3 | Div: 0.784 -> 0.768 | Hurt |
| Use InversionGNN instead of MOGFN-AL | Multi-objective drug discovery · GSK3β + JNK3 + QED + SA | HV: 0.377_{\pm 0.046} -> 0.519_{\pm 0.038} | Helped |
| Use InversionGNN instead of DST | Multi-objective drug discovery · GSK3β + JNK3 + QED + SA | APS: 0.752 -> 0.773 | Helped |
Surprising findings
- InversionGNN has lower molecular diversity (0.435) than HN-GFN (0.571) and MOGFN-PC, which the authors attribute to concentration near the desired weight.
Limitations
- Diversity is lower than some baselines (HN-GFN, MOGFN-AL) on the weight-conditioned task.
- The convergence theorem relies on strong assumptions (submodularity and curvature) that are motivated by observation rather than proven.
- Evaluation is restricted to small molecules with bounded size N, as stated in the assumptions.
Experimental setup
Datasets: ZINC 250K
Benchmarks: synthetic two-objective Pareto task (n=20) GSK3β+JNK3 drug discovery GSK3β+JNK3+QED+SA drug discovery MOGFN synthetic n-gram sequence task
Metrics: hypervolume novelty top-K diversity top-K average property score non-uniformity oracle calls
Baselines: I-LS GAN-GA HN-GFN MOGFN-AL MOGFN-PC MOEA/D NSGA-III LigGPT GCPN MolDQN GA+D MARS
Builds on: differentiable scaffolding tree (DST) MGDA EPO Pareto MTL
Closest papers by meaning
- MARS: Markov Molecular Sampling for Multi-objective Drug Discovery (ICLR 2021 · similarity 0.88)
- Differentiable Scaffolding Tree for Molecule Optimization (ICLR 2022 · similarity 0.86)
- Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.86)
- Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.85)
- MIPT: Multilevel Informed Prompt Tuning for Robust Molecular Property Prediction (ICML 2025 · similarity 0.84)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.84)
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.