English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Self-supervised learning has recently gained growing interest in molecularmodeling for scientific tasks such as AI-assisted drug discovery. Currentstudies consider leveraging both 2D and 3D molecular structures forrepresentation learning. However, relying on straightforward alignmentstrategies that treat each modality separately, these methods fail to exploitthe intrinsic correlation between 2D and 3D representations that reflect theunderlying structural characteristics of molecules, and only performcoarse-grained molecule-level alignment. To derive fine-grained alignment andpromote structural molecule understanding, we introduce an atomic-relationlevel "blend-then-predict" self-supervised learning approach, MoleBLEND, whichfirst blends atom relations represented by different modalities into oneunified relation matrix for joint encoding, then recovers modality-specificinformation for 2D and 3D structures individually. By treating atomrelationships as anchors, MoleBLEND organically aligns and integrates visuallydissimilar 2D and 3D modalities of the same molecule at fine-grained atomiclevel, painting a more comprehensive depiction of each molecule. Extensiveexperiments show that MoleBLEND achieves state-of-the-art performance acrossmajor 2D/3D molecular benchmarks. We further provide theoretical insights fromthe perspective of mutual-information maximization, demonstrating that ourmethod unifies contrastive, generative (cross-modality prediction) andmask-then-predict (single-modality prediction) objectives into one singlecohesive framework.

Multimodal Molecular Pretraining via Modality Blending