IALA-LNN: Deep Learning for Peptide Retention Time Prediction Based on Improved Artificial Lemming...

Area of Science:

• Proteomics
• Computational Biology
• Analytical Chemistry

Background:

• Accurate peptide retention time prediction is vital for reliable identification in liquid chromatography-tandem mass spectrometry (LC-MS/MS) proteomics.
• Existing methods using static neural networks struggle with complex peptide sequence dependencies.
• Current hyperparameter optimization is computationally inefficient and yields suboptimal results.

Purpose of the Study:

• To introduce IALA-LNN, a framework utilizing liquid neural networks (LNNs) governed by ordinary differential equations (ODEs) for enhanced peptide retention time prediction.
• To improve the modeling of sequential dependencies in peptide structures for more accurate retention time prediction.
• To develop an efficient hyperparameter optimization strategy for complex machine learning models in proteomics.

Main Methods:

• Employed liquid neural networks (LNNs) with state evolution governed by ordinary differential equations (ODEs) to capture sequential information.
• Integrated dual encoding using ESM-2 and ProtT5 protein language models.
• Utilized an improved artificial lemming algorithm (IALA) with advanced optimization techniques for hyperparameter tuning.

Main Results:

• IALA-LNN achieved high accuracy across RP, SCX, and HILIC chromatography, with R² values of 0.994, 0.998, and 0.998, respectively.
• Mean Absolute Error (MAE) values were as low as 0.07 min, demonstrating exceptional precision.
• Outperformed established methods like DeepRT, DeepLC, and Prosit in retention time prediction.

Conclusions:

• Differential equation-based neural networks effectively model peptide retention patterns.
• IALA-LNN significantly enhances peptide identification reliability and reduces false discovery rates in LC-MS/MS proteomics.
• The framework supports precision proteomics applications, including biomarker discovery and targeted workflows.