Molecular representation learning: cross-domain foundations and future Frontiers

Abstract

Molecular representation learning has catalyzed a paradigm shift in computational chemistry and materials science—from reliance on manually engineered descriptors to the automated extraction of features using deep learning. This transition enables data-driven predictions of molecular properties, inverse design of compounds, and accelerated discovery of chemical and crystalline materials—including organic molecules, inorganic solids, and catalytic systems. This review provides a comprehensive and comparative evaluation of deep learning-based molecular representations, focusing on graph neural networks, autoencoders, diffusion models, generative adversarial networks, transformer architectures, and hybrid self-supervised learning (SSL) frameworks. Special attention is given to underexplored areas such as 3D-aware representations, physics-informed neural potentials, and cross-modal fusion strategies that integrate graphs, sequences, and quantum descriptors. While previous reviews have largely centered on GNNs and generative models, our synthesis addresses key gaps in the literature—particularly the limited exploration of geometric learning, chemically informed SSL, and multi-modal representation integration. We critically assess persistent challenges, including data scarcity, representational inconsistency, interpretability, and the high computational costs of existing methods. Emerging strategies such as contrastive learning, multi-modal adaptive fusion, and differentiable simulation pipelines are discussed in depth, revealing promising directions for improving generalization and real-world applicability. Notably, we highlight how equivariant models and learned potential energy surfaces offer physically consistent, geometry-aware embeddings that extend beyond static graphs. By integrating insights across domains, this review equips cheminformatics and materials science communities with a forward-looking synthesis of methodological innovations. Ultimately, advances in pretraining, hybrid representations, and differentiable modeling are poised to accelerate progress in drug discovery, materials design, and sustainable chemistry.