Semi-supervised prediction of SH2-peptide interactions from imbalanced high-throughput data

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Src homology 2 (SH2) domains are crucial peptide-recognition modules binding phosphotyrosine peptides.
• Understanding SH2-domain binding partners is key to deciphering cellular processes.
• Existing in-silico prediction methods for SH2-peptide interactions have limitations in coverage, modeling, and computational complexity.

Purpose of the Study:

• To develop a novel, effective machine learning approach for predicting human SH2-peptide interactions.
• To improve upon the performance of existing state-of-the-art prediction methods.
• To provide a valuable tool for the scientific community to predict SH2-peptide binding.

Main Methods:

• Utilized comprehensive data from micro-array and peptide-array experiments for 51 human SH2 domains.
• Employed a semi-supervised machine learning setting to address data imbalance and high signal-to-noise ratio.
• Incorporated information from non-interacting peptides (negative examples) and considered high-order correlations using regularization techniques.

Main Results:

• Achieved a high predictive performance with 0.83 AUC ROC and 0.93 AUC PR, surpassing the PSSM-based SMALI approach (0.71 AUC ROC, 0.87 AUC PR).
• Demonstrated the benefit of including negative examples and high-order correlations in modeling.
• Successfully tackled data imbalance using a semi-supervised strategy.

Conclusions:

• The developed machine learning approach offers a significant improvement for predicting SH2-peptide interactions.
• The study provides valuable insights into SH2-domain binding specificities and their biological relevance through genome-wide predictions.
• Models and genome-wide predictions are made publicly available to facilitate further research.