Large Language Model-Empowered Compound Collision Cross-Section Prediction.

Area of Science:

• Computational chemistry
• Analytical chemistry
• Biinformatics

Background:

• Collision cross section (CCS) is vital for precise compound identification in ion mobility-mass spectrometry (IM-MS).
• Current computational CCS prediction methods struggle with limited data and inadequate multimodal feature handling, leading to suboptimal performance.
• Accurate CCS prediction is essential for building large-scale compound databases for IM-MS applications.

Purpose of the Study:

• To develop a novel computational framework, HyperCCS, for accurate collision cross section (CCS) prediction.
• To leverage chemical large language models (CLLMs) for capturing complex molecular information.
• To integrate multimodal features effectively for improved predictive performance in IM-MS.

Main Methods:

• Fine-tuned a chemical large language model (CLLM) pre-trained on extensive SMILES sequences.
• Developed a cross-modal feature fusion module to integrate CLLM-derived features with other heterogeneous data.
• Evaluated HyperCCS on benchmark datasets (METLIN-CCS, AllCCS2) and in-house experimental data.

Main Results:

• HyperCCS demonstrated robust CCS prediction across various molecular masses, adduct types, and ion modes, outperforming existing methods.
• The framework accurately resolved isomers and extrapolated predictions to high-mass analytes on experimental data.
• SHAP analysis and ablation studies confirmed the significant contribution of CLLM features and the fusion mechanism.

Conclusions:

• HyperCCS offers a significant advancement in computational CCS prediction for IM-MS.
• The integration of CLLMs and cross-modal fusion effectively addresses limitations of previous prediction models.
• HyperCCS provides a high-throughput, adaptable computational tool for metabolomics and structural biology research.