Research map / 3D vision / 3D generation
3D Interaction Geometric Pre-training for Molecular Relational Learning
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 3DMRL, a pre-training method that builds a virtual 3D interaction environment (a large molecule surrounded by randomly placed small molecules) and pre-trains 2D molecular relational learning encoders with global contrastive alignment and local relative-direction prediction.
Problem
Molecular relational learning (MRL) models use only 2D topology because obtaining 3D interaction geometry of molecule pairs is prohibitively expensive, and MRL pre-training has been underexplored compared with single-molecule 3D pre-training.
Approach
Construct a cheap virtual interaction geometry by placing randomly rotated and translated small molecules around atoms of a larger molecule, then pre-train 2D MRL encoders with an SE(3)-invariant global contrastive loss (vs. SchNet embedding of the environment) plus an SE(3)-equivariant local loss predicting relative atom directions via local frames.
Key results
- 3DMRL improves CIGIN RMSE on 40 tasks across five architectures, with up to 24.93% relative reduction (CompSol) versus training from scratch.
- On out-of-distribution DDI, CIGIN+3DMRL reaches 74.85 AUROC on ZhangDDI scaffold split vs 70.96 without pre-training.
- MPP pre-training methods (3D Infomax, GraphMVP, MoleculeSDE) give mixed results and sometimes negative transfer on MRL tasks; removing the global loss drops performance below MPP baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply 3DMRL pre-training to CIGIN | CIGIN · Molecular interaction tasks (all nine datasets/properties) | RMSE: 18.71 -> 18.00; 24.59 -> 24.21; 0.790 -> 0.729; 0.585 -> 0.528; 0.873 -> 0.839; 0.321 -> 0.277; 0.426 -> 0.371; 0.464 -> | Helped |
| Apply 3DMRL pre-training to CIGIN | CIGIN · ZhangDDI and ChChMiner molecule split | AUROC: 71.75 -> 74.00; 76.21 -> 78.93 | Helped |
| Apply 3DMRL pre-training to CIGIN | CIGIN · ZhangDDI and ChChMiner scaffold split | AUROC: 70.96 -> 74.85; 75.81 -> 78.56 | Helped |
| Add global geometry learning to local-only 3DMRL (ablation) | 3DMRL · Molecular interaction tasks | performance | Helped |
| Add local relative geometry learning to global-only 3DMRL (ablation) | 3DMRL · Molecular interaction tasks | performance | Helped |
| Do not account for molecule radius when placing molecules (ablation) | 3DMRL virtual interaction environment · Chromophore Absorption, Emission, and Lifetime | RMSE: 18.00 -> 18.68; 24.21 -> 25.37; 0.729 -> 0.745 | Hurt |
| Fix the placement direction toward the target atom (ablation) | 3DMRL virtual interaction environment · Chromophore Absorption, Emission, and Lifetime | RMSE: 18.00 -> 18.26; 24.21 -> 24.24; 0.729 -> 0.734 | Mixed |
| Double the molecule radius during placement (ablation) | 3DMRL virtual interaction environment · Chromophore Absorption, Emission, and Lifetime | RMSE: 18.00 -> 18.23; 24.21 -> 24.25; 0.729 -> 0.730 | Mixed |
| Regenerate virtual environment after every collision (ablation) | 3DMRL virtual interaction environment · Chromophore Absorption, Emission, and Lifetime | RMSE: 18.00 -> 18.20; 24.21 -> 23.86; 0.729 -> 0.727 | Mixed |
| Use only local geometry loss instead of both geometry losses (ablation) | 3DMRL · Molecular interaction tasks | performance | Hurt |
| Use only global geometry loss instead of both geometry losses (ablation) | 3DMRL · Molecular interaction tasks | performance | Hurt |
| Increase the number of small molecules beyond five (ablation) | 3DMRL · Molecular interaction tasks | performance | Hurt |
Surprising findings
- Single-molecule 3D pre-training (3D Infomax, GraphMVP, MoleculeSDE) often underperforms or causes negative transfer on MRL downstream tasks.
- Scaling MPP pre-training to 50K GEOM molecules does not consistently help MRL performance.
Limitations
- Molecule collisions (atomic overlap about 25%) occur in the randomly generated virtual environment.
- Virtual interaction geometry is purely random and not physically validated against MD or quantum calculations.
- Evaluated mainly on solvation, chromophore and DDI tasks; applicability to protein-ligand binding is untested.
Experimental setup
Datasets: Chromophore CombiSolv DDI (ZhangDDI, ChChMiner, DeepDDI) MNSol FreeSolv CompSol Abraham GEOM
Benchmarks: Chromophore (Absorption, Emission, Lifetime) MNSol FreeSolv CompSol Abraham CombiSolv ZhangDDI ChChMiner
Metrics: RMSE AUROC
Baselines: MPNN AttentiveFP CIGIN CGIB CGIB_Cont 3D Infomax GraphMVP MoleculeSDE
Builds on: CIGIN SchNet 3D Infomax GraphMVP MoleculeSDE
Closest papers by meaning
- Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.86)
- ModuLM: Enabling Modular and Multimodal Molecular Relational Learning with Large Language Models (NeurIPS 2025 · similarity 0.86)
- Generalist Equivariant Transformer Towards 3D Molecular Interaction Learning (ICML 2024 · similarity 0.86)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.86)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.86)
- FlexMol: A Flexible Toolkit for Benchmarking Molecular Relational Learning (NeurIPS 2024 · similarity 0.86)
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.