AHDSN: an attention-enabled hybrid deep sequential network for cancer survivability prediction from multi-omics data

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Oncology

Background:

• Cancer survivability prediction is crucial for patient management and treatment strategies.
• Multi-omics data offers a comprehensive molecular profile for enhanced predictive power.
• Existing survival prediction models often focus on fixed time points, limiting their scope.

Purpose of the Study:

• To introduce a novel deep learning model, the Attention-Enabled Hybrid Deep Sequential Network (AHDSN), for accurate cancer survivability prediction.
• To predict overall survival across the entire follow-up period using comprehensive multi-omics data.
• To evaluate the performance of AHDSN against state-of-the-art methods.

Main Methods:

• Utilized Long Short-Term Memory, Bidirectional Gated Recurrent Unit, and attention mechanisms for feature extraction from multi-omics data.
• Employed Dense layers with softmax activation for classification.
• Applied Random Oversampling (ROS) and Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance.
• Evaluated performance on Glioblastoma, Colon, Breast, Kidney, and Lung cancer multi-omics datasets.

Main Results:

• The AHDSN method achieved superior accuracy, precision, recall, and F1-score compared to existing methods.
• Accuracies ranged from 80.00% to 98.33% depending on the dataset and oversampling technique used.
• Confidence Interval tests confirmed AHDSN's lower error rate and smaller error bound.
• SHapley Additive exPlanations and heatmaps provided insights into feature importance.

Conclusions:

• The proposed AHDSN method demonstrates significant advancements in cancer survivability prediction using multi-omics data.
• The hybrid deep sequential architecture effectively extracts latent features for accurate overall survival prediction.
• The model's interpretability features offer valuable insights into the contribution of individual omics features.