Understanding Tumor-microenvironments

By investigating the relationship between cancer cell populations and their microenvironment, such as the immune system, we can model and predict response to treatment, side effects, and the acquisition of resistance on an individual basis (Figure 1-1). Recently, through the analysis of the whole genome sequence of liver cancer cells from 300 patients, we found a cluster with mutations within a novel cancer-related gene [1]. Patients in this cluster have a good prognosis with a lower chance of recurrence after surgery. We also discovered a new cluster in a subsequent analysis of colorectal cancer [2]. Further detailed analysis revealed the population characteristics of cancer cells and their relationship with the microenvironment, such as immunity, vary greatly among clusters. Cancer cells are abnormal versions of our own cells which can be recognized as non-self, and therefore be eliminated by the immune system. However, selective pressures can drive cancer cells to change and evade elimination. The individuality of the cancer cells and the microenvironment, as well as the treatment implemented, will affect the future development of the tumor. For most cancers, the higher the immune activity, the better the prognosis. In surprising contrast, for cancers in immune-privileged sites, such as uveal melanoma and low-grade gliomas, we found that high immune activity is associated with worse overall survival (Figure 1-2), and we elucidated the mechanism [3]. We aim to clarify the mechanisms that cause these differences through the analysis of omics and pathological images, and to create a model for predicting the effects of treatment using mathematical simulations, with the aim of providing precision medicine that allows treatment to be optimized for each patient.

Methodology for original analysis with deep learning

We research the ability of convolutional neural networks to process not only images, but also non-image data, particularly omics data. We developed a unique technique “DeepInsight” for transforming omics data to look like an image and applying convolutional neural networks to it, so that we can distinguish cancer types from omic data (Figure 2-1) [4]. We also developed the technique “DeepFeature” which analyzes the inner workings of neural network to identify how it discriminates between cancers (Figure 2-2) [5]. Through these techniques, we discovered a new signaling system that indicates the individuality of cancer types. In other words, we were able to show that deep learning can be used to make new scientific discoveries.

Prediction of anticancer drug response by multi-omics and deep learning

We further advanced our DeepInsight method and proposed DeepInsight-3D, a new method using deep learning to predict patient-specific responses to anticancer drugs from multi-omics data (Figure 3-1) [6]. Along with significantly better prediction performance than conventional methods, we discovered gene patterns (Figure 3-2) and new pathways that differentiate responses to anticancer drugs.

Deep learning to identify cell types for single-cell RNA-seq data

Based on our DeepInsight methodology, we proposed scDeepInsight (Figure 4-1), a new method using deep learning to identify the original cell type from single-cell RNA-seq (scRNA-seq) data [7]. Cell type identification is key to studying cell population heterogeneity through the analysis of scRNA-seq data. This method utilizes convolutional neural networks (CNNs), which have high image classification and feature extraction capabilities, by transforming scRNA-seq non-image data into images using our originally proposed DeepInsight method. As a result, scDeepInsight was able to identify cell types with much higher accuracy than other state-of-the-art methods (Figure 4-2). The proposed method is expected to make a broad contribution to future research related to the elucidation of the mechanisms of cell regulation in vivo and in disease using scRNA-seq data.