Research Projects

🧬 Morpho-VAE: Morphological Difference Representation Learning

Overview

Morphology is a phenotype acquired over a long period of time. There are as many morphologies as there are organisms in the world, and even the same organism has slightly different morphologies. Morpho-VAE is an innovative model that represents these differences using deep learning.

Key Features

• Landmark-free: Unlike conventional methods, no specific point designation is required
• Robust to missing data: Analysis possible even with incomplete data
• Application to mandible shape: Effectiveness demonstrated in primate mandible morphology analysis

Technical Details

Using a deep learning architecture based on Variational Autoencoder (VAE), high-dimensional morphological data is efficiently mapped to low-dimensional latent space. Successfully extracts features that distinguish species/families.

Applications

Wide-ranging applications expected in evolutionary biology, morphology, medical image analysis, archaeology, and other fields.

Related Links:

• 📄 Paper
• 💻 Code

🔬 KANADE: Batch Effect Removal Technology

Background

In single-cell transcriptomics, batch effects due to differences in experimental conditions are a major obstacle to analysis. Conventional methods have the problem of simultaneously removing biological signals.

KANADE’s Innovation

KANADE is a new method that cleverly distinguishes between biological signals and batch effects by applying Morpho-VAE. Using deep learning, it removes only technical variation while preserving the essential biological patterns in the data.

Key Features

• Variational Autoencoder: Efficient representation learning through deep learning
• Selective batch effect removal: Preserves biological signals, removes only technical variation
• Scalability: Handles large-scale datasets
• Versatility: Applicable to various single-cell analyses

Practicality

Demonstrates power in integrative analysis, meta-analysis, and database construction from multiple experimental batches.

Related Links:

• 📄 Paper
• 💻 Code

🧠 mindGAP: Quantifying the Population–Individual Gap

Overview

mindGAP (Measuring INDividual–population GAPs in psychiatric energy landscapes) is a hierarchical variational Bayesian framework that jointly estimates population-level and individual-level symptom dynamics from longitudinal questionnaire data and quantifies how they diverge. It formalizes mismatches between collective assessment and personal symptom organization—such as appearing stable while struggling internally, or the reverse—as energy landscapes.

Key Features

• Hierarchical variational Bayes: Joint estimation of population parameters and participant-specific deviations
• Energy landscape analysis: Extracts attractor states and transitions from multivariate response patterns (e.g., PHQ-9)
• Individual–population landscape divergence: Quantified via Jensen–Shannon divergence between landscapes
• Links to external measures: Associated not only with depressive severity but also with modern-type depression traits (TACS-22) and interpersonal sensitivity (IPS-22)

Study (Longitudinal survey during the COVID-19 pandemic)

We applied mindGAP to longitudinal PHQ-9 data from 248 participants across five time points collected in Japan during the COVID-19 pandemic. The population landscape revealed three major stable states (resilient, intermediate, severe), whereas individualized landscapes often diverged substantially from this shared structure. The same symptom trajectories could show different attractor-transition patterns at the individual versus population level (e.g., steady–unsteady, unsteady–steady).

Significance and Outlook

As a way to capture structural heterogeneity in symptoms beyond summary scores, mindGAP may support personalized mental health assessment and precision psychiatry.

Related Links:

• 📄 Paper: *In preparation *

🪰 DOLO (Drosophila tracking with YOLO)