Single-residue linear and conformational B cell epitopes prediction using random and ESM-2 based projections

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning in Immunology

Background:

• B cell epitopes are crucial for antibody recognition and vaccine design.
• Current prediction methods are broadly categorized into linear (sequence-based) and conformational (structure-based).
• Existing linear predictors often rely on machine learning applied to sequence-derived representations.

Purpose of the Study:

• To introduce CALIBER (Conformational And LInear B cell Epitopes pRediction), a novel prediction tool.
• To evaluate CALIBER's performance against state-of-the-art methods for both linear and conformational B cell epitopes.
• To investigate the impact of sequence information and structural data on epitope prediction accuracy.

Main Methods:

• Development of CALIBER, employing a bidirectional long short-term memory (BiLSTM) neural network with random projection.
• Integration of Evolutionary Scale Modeling-2 (ESM-2) projection for enhanced feature representation.
• Assessment of prediction accuracy using Area Under the Curve (AUC) on test datasets.

Main Results:

• CALIBER achieved superior prediction accuracy for linear epitopes (AUC=0.789) compared to existing methods.
• The BiLSTM model combined with ESM-2 projection improved state-of-the-art conformational epitope prediction (AUC=0.776).
• Long-range sequence information proved essential for high prediction accuracy, while 3D structural graphs did not enhance performance.

Conclusions:

• CALIBER demonstrates enhanced accuracy for both linear and conformational B cell epitope prediction.
• The BiLSTM architecture effectively leverages sequence information for improved immunoinformatics predictions.
• Separate model training is necessary for linear and conformational epitopes despite structural similarities in the model.