Adversary-Aware Multimodal Neural Networks for Cancer Susceptibility Prediction From Multiomics Data

Artificial intelligence (AI) systems are increasingly used in health and personalized care. However, the adoption of data-driven approaches in many clinical settings has been hampered due to their inability to perform in a reliable and safe manner to leverage accurate and trustworthy diagnoses. A critical and challenging usage scenario for AI is aiding the treatment of cancerous conditions. Providing accurate diagnosis for cancer is a challenging problem in precision oncology. Although machine learning (ML)-based approaches are very effective at cancer susceptibility prediction and subsequent treatment recommendations, ML models can be vulnerable to adversarial attacks. Since adversarially weak models can lead to wrong clinical recommendations, such vulnerabilities is critical - especially when AI-guided systems are used to aid medical doctors. Therefore, it is indispensable that healthcare professionals employ trustworthy AI tools for predicting and assessing disease risks and progression. In this paper, we propose an adversary-aware multimodal convolutional autoencoder (MCAE) model for cancer susceptibility prediction from multi-omics data consisting of copy number variations (CNVs), miRNA expression, and gene expression (GE). Based on different representational learning techniques, the MCAE model learns multimodal feature representations from multi-omics data, followed by classifying the patient cohorts into different cancer types on multimodal embedding space that exhibit similar characteristics in end-to-end setting. To make the MCAE model robust to adversaries and to provide consistent diagnosis, we formulate robustness as a property, such that predictions remain stable with regard to small variations in the input. We study different adversarial attacks scenarios and take both proactive and reactive measures (e.g., adversarial retraining and identification of adversarial inputs). Experiment results show that the MCAE model based on latent representation concatenation (LRC) exhibits high confidence at predicting cancer types, giving an average precision and Matthews correlation coefficient (MCC) scores of 0.9625 and 0.8453, respectively and shows higher robustness when compared with state-of-the-art approaches against different attack scenarios w.r.t. ERM and CLEVER scores. Overall, our study suggests that a well-fitted and adversarially robust model can provide consistent and reliable diagnosis for cancer.