CancerPred

CancerPred predicts cancerlectins from non-cancer lectins using sequence-derived features and machine learning to identify lectins implicated in cancer metastasis, tumor progression, and oncogenesis.

Key Features:

• Dataset: Uses a curated non-redundant dataset of 178 cancerlectins and 226 non-cancerlectins with a sequence similarity threshold of 50%.
• Amino Acid Compositional Analysis: Identified amino acid preferences in cancerlectins (e.g., Leucine, Proline) versus non-cancerlectins (e.g., Aspartic acid, Asparagine).
• Domain Identification (PROSITE): PROSITE domain analysis showed the "Crystalline beta gamma" domain is abundant in cancerlectins while the "SUEL-type lectin domain" predominates in non-cancerlectins.
• Machine Learning Models: Multiple SVM models were trained: amino acid composition (MCC 0.32, accuracy 64.84%), dipeptide composition (MCC 0.30, accuracy 64.84%), split composition (2-part MCC 0.31, accuracy 65.10%; 4-part MCC 0.32, accuracy 66.09%), and a PSSM-based model (PSSM generated by PSI-BLAST) (MCC 0.36, accuracy 68.34%).
• Integrated Model: Integration of PROSITE domain information with PSSM achieved the highest reported performance (MCC 0.38, accuracy 69.09%).

Scientific Applications:

• Lectin classification in cancer research: Distinguishes cancerlectins from non-cancerlectins to support studies in cancer biology and lectin research by identifying sequence patterns and domain preferences relevant to metastasis, tumor progression, and oncogenesis.

Methodology:

Dataset curation (non-redundant, 178 cancerlectins, 226 non-cancerlectins, ≤50% sequence similarity), amino acid composition analysis, PROSITE domain analysis, PSSM generation by PSI-BLAST, SVM classifiers trained on amino acid composition, dipeptide composition, split composition, and PSSM, and a final model integrating PROSITE domains with PSSM.