Research map / Label-efficient and robust learning / OOD and anomaly detection
Relational Invariant Learning for Robust Solvation Free Energy Prediction
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 RILOOD, a relational invariant learning framework for out-of-distribution solvation free energy prediction combining mixup-based conditional variational modeling, multi-granularity context-aware refinement, and an invariant learning objective.
Problem
Solvation free energy prediction with graph neural networks degrades under out-of-distribution shifts across solvents and scaffolds, since existing molecular relational learning methods assume IID data and rely on substructure-based invariance.
Approach
Combines a mixup-enhanced conditional VAE that models solute representations across solvent environments without explicit environment labels, a coarse/fine-grained self-attention and local-interaction refinement (MCAR) producing context-aware invariant features, and a mutual-information contrastive objective, all trained jointly on top of the CIGIN-style solute-solvent message-passing encoder.
Key results
- On the MNSolv solvent split, RILOOD reaches RMSE 0.6784 versus 0.7135 for the best baseline (MixUp).
- On the MNSolv scaffold split, RILOOD reaches RMSE 1.0780 versus 1.3478 for ERM, and on CompSolv solvent split 0.3660 versus 0.3777 for CMRL.
- Ablation on CompSolv and MNSolv shows that removing MCAR raises scaffold RMSE on MNSolv from 1.0780 to 1.3929.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RILOOD instead of GAT | MNSolv · solvent split | RMSE: 0.7022±0.012 -> 0.6784±0.007 | Helped |
| Use RILOOD instead of GCN | MNSolv · scaffold split | RMSE: 1.2752±0.022 -> 1.0780±0.013 | Helped |
| Use RILOOD instead of CMRL | CompSolv · solvent split | RMSE: 0.3777±0.023 -> 0.3660±0.022 | Helped |
| Use RILOOD instead of CGIB | CompSolv · scaffold split | RMSE: 0.5476±0.026 -> 0.5209±0.014 | Helped |
| Use RILOOD instead of MixUp | QM9Solv · solvent split | RMSE: 0.7227±0.003 -> 0.7001±0.001 | Helped |
| Use RILOOD instead of ERM | QM9Solv · scaffold split | RMSE: 0.7261±0.005 -> 0.6991±0.003 | Helped |
| Add MCAR to the baseline (ablation) | CompSolv solvent split · CompSolv solvent split | RMSE: 0.5215±0.007 -> 0.4914±0.004 | Helped |
| Add MCAR to the baseline (ablation) | CompSolv scaffold split · CompSolv scaffold split | RMSE: 0.6383±0.011 -> 0.5842±0.003 | Helped |
| Add MCAR to the baseline (ablation) | MNSolv solvent split · MNSolv solvent split | RMSE: 0.7662±0.016 -> 0.7115±0.003 | Helped |
| Add MCAR to the baseline (ablation) | MNSolv scaffold split · MNSolv scaffold split | RMSE: 1.2648±0.018 -> 1.2191±0.012 | Helped |
| Remove MCAR from the full loss configuration (ablation) | RILOOD · CompSolv scaffold split | RMSE: 0.5209±0.014 -> 0.6641±0.013 | Hurt |
| Remove MCAR from the full loss configuration (ablation) | RILOOD · MNSolv scaffold split | RMSE: 1.0780±0.013 -> 1.3929±0.002 | Hurt |
Surprising findings
- Several baselines (CGIB, CMRL) that are strong on some splits degrade sharply on others, e.g. MNSolv scaffold RMSE above 2.1 versus about 1.3 for ERM.
Limitations
- Environment modeling uses solvent categories as a proxy and the authors avoid a learnable soft-environment assignment because of limited solvent diversity.
- Evaluated on a small number of solvent-based datasets; mixup and CVAE modules add complexity relative to the base encoder.
- Synthetic spurious-correlation experiments rely on hand-designed injection rather than natural shifts.
Experimental setup
Datasets: MNSolv QM9Solv CompSolv CombiSolv MolMerger Abraham DDI datasets
Benchmarks: MNSolv CompSolv QM9Solv DDI datasets
Metrics: RMSE AUROC
Baselines: GCN GIN GAT ERM GroupDRO MixUp MolMerger CIGIN CGIB CMRL
Builds on: CIGIN ZIN CMRL CGIB MixUp
Closest papers by meaning
- PRISM: Prior Relational Information for Self-supervised Modeling to Enhance Solubility OOD Generalization (NeurIPS 2026 · similarity 0.89)
- Chemically Interpretable Graph Interaction Network for Prediction of Pharmacokinetic Properties of Drug-Like Molecules (AAAI 2020 · similarity 0.87)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.85)
- BOOM: Benchmarking Out-Of-distribution Molecular Property Predictions of Machine Learning Models (NeurIPS 2025 · similarity 0.84)
- Colour me shocked: Exact Molecular Hessians from MLIPs in O(N) time using sparse differentiation! (NeurIPS 2026 · similarity 0.84)
- MIPT: Multilevel Informed Prompt Tuning for Robust Molecular Property Prediction (ICML 2025 · similarity 0.83)
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.