構造化配列特徴からの細胞浸透ペプチドの正確な同定のためのマルチモデル説明可能ディープラーニングフレームワーク：XCPP

The therapeutic candidate CPPs are short amino acid sequences that can deliver attractive therapeutic molecules across cellular membranes. Hence, CPPs can provide the backbone and mechanisms for drug tailoring and diagnostics.

A structured dataset consisting of 473 confirmed CPPs derived from the EnDM-CPP database was analyzed. Four sequence descriptors were computed: Position Relative Incidence Matrix (PRIM), Reverse PRIM (RPRIM), Accumulative Absolute Position Incidence Vector (AAPIV), and Reverse AAPIV. These feature vectors were used to train and test three deep learning architectures: Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks. Self-consistency testing, independent testing, and 10-fold crossvalidation were used for model evaluation. Additionally, SHAP values were employed to explain the XAI sequences and identify the most important sequence components.

The best score during cross-validation was achieved by the CNN model, with an accuracy of 99.05%, followed by the DNN. In contrast, the LSTM model achieved an accuracy of 95.24%, a substantial drop from the former, suggesting it may be suffering from overfitting. In general, all models were able to predict outputs with reasonable accuracy given the structured input features. Biologically relevant descriptors were highlighted by the SHAP analysis, which improved the transparency of the predictions.

The performance of the CNN indicates its ability to process structured sequence-based biological data effectively. The drop in accuracy with the LSTM model suggests insufficient training data or lack of proper regularization. The use of SHAP enhances explainability by linking features to biological properties. While in vitro or in vivo validation has not yet been performed, the current analysis suggests that in silico CPP prediction is feasible.

Deep learning in this case has created a framework that can accurately identify CPPs. The CNN model also has additional features that make it ideal for use in drug delivery, molecular diagnostics, and personalized medicine. SHAP-based XAI further adds explainability, increasing trust in the model's predictions.