Learning the Neighborhood: Contrast-Free Multimodal Self-Supervised Molecular Graph Pretraining

High-quality molecular representations are essential for property prediction and molecular design, yet large labeled datasets remain scarce. While self-supervised pretraining on molecular graphs has shown promise, many existing approaches either depend on hand-crafted augmentations or complex generative objectives, and often rely solely on 2D topology, leaving valuable 3D structural information underutilized. To address this gap, we introduce C-FREE (Contrast-Free Representation learning on Ego-nets), a simple framework that integrates 2D graphs with ensembles of 3D conformers. C-FREE learns molecular representations by predicting subgraph embeddings from their complementary neighborhoods in the latent space, using fixed-radius ego-nets as modeling units across different conformers. This design allows us to integrate both geometric and topological information within a hybrid Graph Neural Network (GNN)-Transformer backbone, without negatives, positional encodings, or expensive pre-processing. Pretraining on the GEOM dataset, which provides rich 3D conformational diversity, C-FREE achieves state-of-the-art results on MoleculeNet, surpassing contrastive, generative, and other multimodal self-supervised methods. Fine-tuning across datasets with diverse sizes and molecule types further demonstrates that pretraining transfers effectively to new chemical domains, highlighting the importance of 3D-informed molecular representations.

Designing new medicines and materials relies heavily on AI models that can predict molecular properties, but training these models requires large amounts of labeled experimental data that is costly and slow to produce. Most existing approaches also only consider a molecule's 2D bonding structure, ignoring the 3D shape that largely determines its function. We developed C-FREE, a self-supervised framework that learns molecular representations from both 2D connectivity and ensembles of 3D shapes, without needing labeled data or complex preprocessing. It works by training a model to predict how different parts of a molecule relate to each other across multiple realistic 3D configurations. Tested on MoleculeNet, the standard benchmark for molecular property prediction, C-FREE outperforms all existing self-supervised methods and transfers well to new chemical domains, suggesting that 3D molecular geometry is a key ingredient for building better molecular AI.