ALIGNing Learned Representations: Domain-General Models Triangulate Physicochemical Interactions in Analytical Separations
Year: 2026
Type: preprint
Abstract: Despite rapid progress, foundation models for chemistry have yet to consistently deliver broadly transferable, experiment-grounded representations for structure-to-measurement tasks in practice. This reflects both the scarcity and fragmentation of experimental datasets and the weak alignment between common molecular representations (e.g., SMILES and graphs) and measured chemical behaviour. Here, we formalize domain-general models: pretrained models constrained across multiple measurement contexts within a physically coherent domain to learn transferable chemical representations that capture how molecular structure gives rise to experimental observables. Using liquid chromatography as a testbed, we develop ALIGN and show that it outperforms conventional baselines, including foundation models and transfer learning approaches, across diverse chromatographic contexts while retaining performance in data-limited and chemically out-of-distribution settings. By training across multiple chromatographic methods, ALIGN triangulates an underlying physicochemical interaction space, with each method providing a complementary projection that constrains the learned representation. We show that ALIGN learns embeddings consistent with physicochemical properties and chromatographic principles and provide evidence for a mechanism underlying improved method-specific adaptation. These representations transfer to related analytical modalities, including GC, CCS, and DMS, suggesting that these measurements encode overlapping physicochemical relationships. Together, these results demonstrate the value of physically grounded representation learning and suggest that domain-general approaches may extend to other fragmented scientific domains.
Source: ChemRxiv
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7204876394