ClipMol: A Molecular Representation Learning Framework for CCS Prediction via SMILES-InChI Dual-View Chemical...

Area of Science:

• Computational chemistry
• Machine learning
• Analytical chemistry

Background:

• Existing molecular pretraining models often require computationally expensive 3D conformer sampling or external corpora, impacting structural fidelity.
• There is a need for efficient and structurally accurate molecular representation learning frameworks.

Purpose of the Study:

• To introduce ClipMol, a novel molecular representation learning framework.
• To evaluate ClipMol's performance on general molecular tasks and specifically for collision cross-section (CCS) prediction in ion mobility-mass spectrometry (IM-MS).

Main Methods:

• ClipMol utilizes a SMILES-InChI dual-view chemical-language alignment approach.
• The framework jointly models local chemical microenvironments and global structural constraints without explicit 3D conformers or external corpora.
• ClipMol and ClipMol-XL were benchmarked on classification, regression, and CCS prediction tasks.

Main Results:

• ClipMol and ClipMol-XL demonstrated strong performance on classification and regression tasks.
• For CCS prediction on METLIN-CCS and ALLCCS datasets, ClipMol showed stable and competitive results, outperforming state-of-the-art models.
• ClipMol exhibited robustness across different adduct compositions and chemical categories in CCS prediction.

Conclusions:

• ClipMol offers a scalable and structurally faithful solution for molecular representation learning.
• The framework is particularly effective for IM-MS-related CCS prediction in analytical chemistry.
• ClipMol advances the field by providing an efficient alternative to existing methods requiring 3D conformers.