Predicting potential residues associated with lung cancer using deep neural network.

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Genomics

Background:

• Lung cancer, particularly lung adenocarcinoma (LUAD) and lung squamous carcinoma (LUSC), causes high mortality globally.
• Somatic driver mutations in proteins are key in LUAD and LUSC development.
• Current mutation screening methods are often costly and time-consuming.

Purpose of the Study:

• To systematically analyze preferred amino acid residues, pairs, and motifs at disease-prone versus neutral sites in proteins.
• To develop a computational method for predicting disease-prone sites in lung cancer using sequence-based features.
• To provide a freely accessible web server for identifying these critical sites.

Main Methods:

• Comparative analysis of 4172 disease-prone sites against 4137 neutral sites across 195 proteins.
• Utilized deep neural networks incorporating physicochemical properties, conservation scores, secondary structure, and peptide motifs.
• Developed and validated the CanProSite web server using 10-fold cross-validation and independent test sets.

Main Results:

• Identified specific motifs (e.g., LG, QF, TST) and amino acids (e.g., Gly, Asp, Glu) enriched in disease-prone sites.
• The deep learning model achieved 81% accuracy (82% sensitivity, 78% specificity, 0.91 AUC) in cross-validation.
• Independent testing yielded 80% accuracy and 0.89 AUC, demonstrating robust predictive performance.

Conclusions:

• The developed method effectively distinguishes between disease-causing and neutral sites in lung cancer.
• CanProSite offers a valuable tool for researchers to identify potential driver mutations in lung cancer.
• This approach can streamline the identification of critical sites, potentially reducing costs and time in cancer research.