Research

[Keywords] Genomic medicine, Precision medicine, Tumor microenvironment, Deep learning, Quantum computing, Mathematical simulation

Understand Life Phenomena and Diseases with Advanced Mathematical Science and Molecular Observation

The goal of our research is to overcome diseases such as cancer by analyzing large-scale data on biomolecules from numerous clinical samples using deep learning and quantum computing, thereby elucidating the role of biological phenomena—such as the immune response—in these diseases. This is expected to enable the delivery of appropriate types and doses of treatment personalized for each patient, as well as the prevention of disease onset. To achieve this, we must devise new mathematical and scientific methodologies based on deep learning and quantum computing, and analyze big data in human life sciences—including spatiotemporal omics molecular data, image data, and clinical information. By discovering new causes of disease in this way and reconstructing the relationships between these factors as quantitative networks, we can understand disease mechanisms as holistic systems. As one example, by unraveling and quantitatively reconstructing the relationships between cancer cells and the tumor microenvironment—including the immune system—and incorporating their underlying structures, we can predict the dynamics of treatment efficacy, side effects, and the development of resistance on an individual basis.

Understanding Tumor-microenvironments

By unraveling the relationship between cancer cell populations and their microenvironment—including the immune system—we can predict treatment response, side effects, and the development of resistance on an individual basis (Figure 1-1). For example, we analyzed the whole-genome sequences of cancer cells from 300 liver cancer patients and identified a cluster harboring novel mutations [1]. Patients in this cluster are less likely to experience cancer recurrence after surgery and tend to have a favorable prognosis. We also identified a new cluster in colorectal cancer [2]. Further detailed analysis revealed that the collective characteristics of cancer cells and their relationship with the microenvironment, including the immune system, vary significantly between clusters. Differences in a cancer’s unique characteristics, its compatibility with the microenvironment, and the specifics of treatment lead to variations in subsequent behavior. Recently, we conducted multi-omic tumor analyses of metastatic urothelial carcinoma, focusing on the fact that the response to immune checkpoint inhibitors varies between tumor sites (Figures 1-2) [3] (a joint study with Keio University School of Medicine). The results revealed that non-responsive malignant subclones acquire unique genetic mutations during treatment, which may contribute to disease progression. Furthermore, spatial transcriptomics and single-cell RNA-seq analyses demonstrated that these malignant subclones create site-specific immunosuppressive environments. We aim to elucidate the mechanisms underlying these differences through omics and pathological image analysis, while also developing a treatment efficacy prediction model using mathematical simulations, with the goal of achieving precision medicine tailored to individual patients and tumor sites.

Figure 1-1. Unravel and predict diseases with patients’ omics, clinical, and pathological image data.

Figure 1-2. The response to immune checkpoint inhibitors in metastatic urothelial carcinoma varies depending on the tumor site. This is due to different subclones forming an immunosuppressive environment at each site.

Original analysis methodology with deep learning

We are conducting research to maximize the ability of deep learning to analyze not only images but also genomic and omics data. As one example, we have devised a unique method called DeepInsight, which transforms omics data to make it appear as an image and then applies deep learning to it (Figure 2-1) [4] (We have also published a review article covering this and the application examples below [5]). Furthermore, we proposed the DeepFeature method, a technique that analyzes the inner workings of deep learning to identify what the model observes and how it makes its classifications (Fig. 2-2) [6]. Through this, we discovered a new signaling pathway that characterizes cancer. In other words, we demonstrated that deep learning can lead to new scientific discoveries.

Figure 2-1. Our unique method that converts omics data into images and feeds them into deep learning models identifies the distinct characteristics of cancer.

Figure 2-2. Extracting features important for differentiating cancer types.

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

Building on our DeepInsight method described above, we have proposed DeepInsight-3D, a new deep learning-based approach that predicts patient-specific responses to anticancer drugs using multi-omics data (Figure 3-1) [7]. In addition to achieving predictive performance that significantly surpasses that of conventional methods, we have identified gene patterns (Figure 3-2) and novel pathways that underlie differences in anticancer drug response.

Figure 3-1. Prediction of anticancer drug response by multi-omics and deep learning.

Figure 3-2. Differential gene patterns involved in non-response (left) and response (right) to the anticancer drug paclitaxel.

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

Based on our DeepInsight method described above, we have developed a new deep learning-based approach called scDeepInsight [8], and scHDeepInsight, an additional novel deep learning-based method that considers the hierarchical relationships in cell classification—commonly observed in immune cells—to accurately identify original cell types and subtypes from single-cell RNA-seq data (Figure 4-1) [9]. Accurately identifying cell types and their subtypes is key to studying the heterogeneity of cell populations through the analysis of scRNA-seq data. In this method, by converting non-image scRNA-seq data into images using the DeepInsight method we have independently proposed, we can leverage convolutional neural networks (CNNs), which excel at image classification and feature extraction. As a result, scHDeepInsight was able to identify cell types with significantly higher accuracy than other state-of-the-art methods (Fig. 4-2). It is expected that this proposed method will contribute broadly to future research aimed at elucidating the mechanisms of cell regulation in vivo and in disease using scRNA-seq data.

Figure 4-1. scHDeepInsight Pipeline.

Figure 4-2. Color-coded results for each cell type predicted by scHDeepInsight.

Predicting drug efficacy using quantum computing and a new method for analyzing gut microbiome using variational Bayes

As an application of quantum computing, we constructed a quantum-classical hybrid machine learning model that combines quantum machine learning with deep learning to predict the efficacy of anticancer drugs. We achieved high accuracy and proposed improvements to the interface between the quantum circuit and the deep learning network [10]. We also developed a new method called Variational Bayesian Microbiome Multi-omics (VBayesMM) [11]. This method enables the high-precision prediction of human metabolite levels at specific sites based on data from the microbiome within and on the surface of the human body, making it possible to elucidate the complex interaction mechanisms between microorganisms and host (human) metabolites in both healthy and diseased states. This will, for example, lead to dramatic progress in understanding the biochemical functions of the gut microbiota, which has traditionally been difficult. In this way, we are actively proposing new methods and approaches to address challenges in biomedical science.

Figure 5-1. Quantum-classical hybrid machine learning model.

Figure 5-2. Overview of the VBayesMM method.

Various people gather and research

Our laboratory is located at Hongo campus. In addition to the research topics mentioned above, we are developing methods for omics analysis using quantum computers. We are also collaborating with the Uemura Laboratory in the Department of Biological Sciences at the Graduate School of Science, The University of Tokyo, to develop methods for identifying molecules and determining their sequences by analyzing signals from custom-built nanopores. Furthermore, we are conducting numerous collaborative research projects with hospitals and other medical institutions. Our lab includes bioinformatics researchers, clinical doctors, researchers specializing in genetics, sequence analysis, network analysis, and mathematics, as well as many international researchers. We engage in daily research by exchanging diverse perspectives with each another. Our laboratory promotes medical science with an emphasis more on analytical methodologies than experimentation.

References:

[1] Fujimoto A, Tsunoda T, et al. Whole genome mutational landscape and characterization of non-coding and structural mutations in liver cancer. Nature Genetics, 48, 500-509 (2016).
[2] Sugawara T, Miya F, Ishikawa T, Lysenko A, Nishino J, Kamatani T, Takemoto A, Boroevich KA, Kakimi K, Kinugasa Y, Tanabe M, Tsunoda T. Immune subtypes and neoantigen-related immune evasion in advanced colorectal cancer. iScience, 25, 1003740 (2022).
[3] Kamatani T, Umeda K, Iwasawa T, Miya F, Matsumoto K, Mikami S, Hara K, Shimoda M, Suzuki Y, Nishino J, Kato M, Kakimi K, Tanaka N, Oya M, Tsunoda T. Clonal diversity shapes the tumour microenvironment leading to distinct immunotherapy responses in metastatic urothelial carcinoma. Nature Communications, 16, 7995 (2025).
[4] Sharma A, Vans E, Shigemizu D, Boroevich KA, Tsunoda T. DeepInsight: A methodology to transform a non-image data to an image for convolution neural network architecture. Scientific Reports, 9, 11399 (2019).
[5] Sharma A, Lysenko A, Jia S, Boroevich KA, Tsunoda T. Advances in AI and machine learning for predictive medicine. Journal of Human Genetics, 69, 487-497 (2024).
[6] Sharma A, Lysenko A, Boroevich KA, Vans E, Tsunoda T. DeepFeature: feature selection in nonimage data using convolutional neural network. Briefings in Bioinformatics, 22, bbab297 (2021).
[7] Sharma A, Lysenko A, Boroevich KA, Tsunoda T. DeepInsight-3D architecture for anti-cancer drug response prediction with deep-learning on multi-omics. Scientific Reports, 13, 2483 (2023).
[8] Jia S, Lysenko A, Boroevich KA, Sharma A, Tsunoda T. scDeepInsight: a supervised cell-type identification method for scRNA-seq data with deep learning. Briefings in Bioinformatics, 24, bbad266 (2023).
[9] Jia S, Lysenko A, Boroevich KA, Sharma A, Tsunoda T. scHDeepInsight: a hierarchical deep learning framework for precise immune cell annotation in single-cell RNA-seq data. Briefings in Bioinformatics, 26, bbaf523 (2025).
[10] Ito T, Lysenko A, Tsunoda T. Optimal normalization in quantum-classical hybrid models for anti-cancer drug response prediction. arXiv:2505.10037 (2025).
[11] Dang T, Lysenko A, Boroevich KA, Tsunoda T. VBayesMM: variational Bayesian neural network to prioritize important relationships of high-dimensional microbiome multiomics data. Briefings in Bioinformatics, 26, bbaf300 (2025).