Predicting Gene Mutations in Colon Cancer Using Long-Term Temporal Dependency Learning on a Directed Co-Occurrence Asymmetry Graph.

[MOTIVATION] Accurate prediction of mutational dependencies to model tumor evolution can improve our understanding of cancer progression and is crucial for early diagnosis and interventions. This is especially true in colorectal tumorigenesis which often follows stepwise mutational progression patterns. Here, we propose a graph-based approach to infer a longest path as well as global gene mutational order based on data-driven co-occurrence conditional probabilities. we train long-term deep learning models (Long-Short-Term Memory (LSTM) recurrent neural network and dilated Convolutional Neural Network (CNN) to predict, for each path and at each pseudo-temporal step, whether the next gene in the inferred path is mutated. We use a large gene network (23,858 genes) and a large curated cohort (2,344 samples) of publicly available human colon adenocarcinoma samples for path inference and training and validation tasks.

[RESULTS] From the proposed directed co-occurrence asymmetry graph we inferred a longest path and global graph-derived mutational orders. Both proposed long-term models achieve high prediction accuracies and dramatically improve precision and recall in mutation prediction over previous studies' results, with longest weighted path achieving best results overall.

[AVAILABILITY] https://github.com/moussa-lab/MutationPrediction.