Abstract: In oncology, machine learning (ML) is implemented for personalized medicine to predict patient response to treatment and select optimal treatment accordingly. ML models are built from high-dimensional omics data, and therefore require a dimension reduction step to avoid “n<<p” ML issue. For now, dimension reduction algorithms are either unsupervised or not always dedicated to keep important information for following-step classification1.
In that purpose, we developed the Reductive Discriminating Network (ReDiN). This new deep learning algorithm reduces dimension with a focus on binary classification. As in Generative Adversarial Networks2, two neural networks learn simultaneously: one reduces data, the other assigns a score to each reduced sample to evaluate its class. These networks are optimized so that the Wasserstein distance between distributions of the two reduced data classes is maximized.
Random forests (RF) were used afterward to evaluate ReDiN on final prediction performance. Several synthetic datasets mimicking biological data were used to identify dataset key characteristics leading to performance variations. With uncorrelated-variable datasets, ReDiN improved RF prediction from 30% initial error to 10%. Increasing correlated variable proportion in datasets led to error classification augmentation from 10% to 45%. Accordingly, we used network inference and graph neural networks to further improve ReDiN performances.

Keywords: dimension reduction; deep learning; binary classification; Wasserstein distance

1. Review of dimension reduction methods. S. Nanga, A.T. Bawah, B.A. Acquaye, M.-I. Billa, F.D. Baeta, N.A. Odai, S.K. Obeng, A.D. Nsiah. Journal of data analysis and information processing, 2021
2. Generative adversarial nets. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio. Advances in neural information processing systems, 2014