Accurate Prediction of Formylation PTM Site using Multiple Feature Fusion with LightGBM Resolving Data Imbalance Issue

Formylation is a lysine post-translational modification that is recently discovered in histone. It has significance in DNA binding and responsible for the epigenetics of chromatin activities. Identifying lysine formylation sites accurately is very important for understanding the implicit characteristics and functions which may cause a diversity of living beings as well as result in severe diseases. Mass spectrometry is the manual laboratory technique that requires human labor which is time demanding as well as costly. So it’s urgent to develop computation based effective predictor which can effectively predict the formylation PTM sites accurately. In our model, a multiple feature fusion technique is used which combines sequence based CKSAAP, physicochemical property based AAIndex and mutation based evolutionary information. The imbalanced dataset is handled by generating synthetic data using SMOTE. LightGBM, a decision tree based gradient boosting classifier, is chosen for our model. Our model achieved accuracy, sensitivity, specificity and MCC of 94.81 ± 1.2%, 90.21 ± 2.8%, 99.41 ± 1.08% and 0.90 for cross-validation and 91.5%, 83.87%, 99.14% and 0.84 for independent test set performance respectively which surpassed the previously developed tool CKSAAP_FormSite to large margin.