Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning Approaches for Blood Disease Diagnosis Across Hematopoietic Lineages

Gabriel Bo, Justin Gu, Christopher Sun

TL;DR

This work builds a foundation model for blood disease diagnosis by embedding high-dimensional gene expression from hematopoietic cells into a $256$-dimensional latent space via a fully connected autoencoder trained on multipotent progenitors. The latent representations are then consumed by fully connected networks, Transformer self-attention, and graph convolutional networks to classify diseases, with a zero-shot evaluation on downstream cell types. The approach achieves multi-class accuracy above 95% on progenitors and attains a zero-shot F1-score greater than $0.7$ for lymphocytes, though cross-cell-type transfer to lymphocytes remains challenging. Overall, the study demonstrates cross-lineage utility of a single foundation model for hematopoietic disease diagnosis and highlights directions to improve lymphocyte-specific embeddings and generalization across cell types.

Abstract

We present a foundation modeling framework that leverages deep learning to uncover latent genetic signatures across the hematopoietic hierarchy. Our approach trains a fully connected autoencoder on multipotent progenitor cells, reducing over 20,000 gene features to a 256-dimensional latent space that captures predictive information for both progenitor and downstream differentiated cells such as monocytes and lymphocytes. We validate the quality of these embeddings by training feed-forward, transformer, and graph convolutional architectures for blood disease diagnosis tasks. We also explore zero-shot prediction using a progenitor disease state classification model to classify downstream cell conditions. Our models achieve greater than 95% accuracy for multi-class classification, and in the zero-shot setting, we achieve greater than 0.7 F1-score on the binary classification task. Future work should improve embeddings further to increase robustness on lymphocyte classification specifically.

Deep Learning Approaches for Blood Disease Diagnosis Across Hematopoietic Lineages

TL;DR

This work builds a foundation model for blood disease diagnosis by embedding high-dimensional gene expression from hematopoietic cells into a -dimensional latent space via a fully connected autoencoder trained on multipotent progenitors. The latent representations are then consumed by fully connected networks, Transformer self-attention, and graph convolutional networks to classify diseases, with a zero-shot evaluation on downstream cell types. The approach achieves multi-class accuracy above 95% on progenitors and attains a zero-shot F1-score greater than for lymphocytes, though cross-cell-type transfer to lymphocytes remains challenging. Overall, the study demonstrates cross-lineage utility of a single foundation model for hematopoietic disease diagnosis and highlights directions to improve lymphocyte-specific embeddings and generalization across cell types.

Abstract

We present a foundation modeling framework that leverages deep learning to uncover latent genetic signatures across the hematopoietic hierarchy. Our approach trains a fully connected autoencoder on multipotent progenitor cells, reducing over 20,000 gene features to a 256-dimensional latent space that captures predictive information for both progenitor and downstream differentiated cells such as monocytes and lymphocytes. We validate the quality of these embeddings by training feed-forward, transformer, and graph convolutional architectures for blood disease diagnosis tasks. We also explore zero-shot prediction using a progenitor disease state classification model to classify downstream cell conditions. Our models achieve greater than 95% accuracy for multi-class classification, and in the zero-shot setting, we achieve greater than 0.7 F1-score on the binary classification task. Future work should improve embeddings further to increase robustness on lymphocyte classification specifically.

Paper Structure

This paper contains 15 sections, 5 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction and Related Work
  2. Dataset and Features
  3. Methods
  4. Fully-Connected Autoencoder
  5. Fully-Connected Neural Network
  6. Transformer Model with Multi-head Self-Attention
  7. Graph Convolutional Networks
  8. Experiments and Results
  9. Experiments
  10. Fully-Connected Neural Network
  11. Transformer Model with Multi-head Self-Attention
  12. Graph Convolutional Network
  13. Improving Downstream Classification Diagnosis
  14. Conclusion and Future Work
  15. Contributions

Figures (2)

  • Figure 1: We used single-cell sequencing data from Hematopoietic Multipotent Progenitor Cells, Monocytes, and Lymphocytes in this analysis. The PCA-transformed latent embeddings for these cell types are plotted accordingly.
  • Figure 2: We trained different architectures on hematopoietic multipotent progenitor cells and simultaneously plotted generalization capacity to downstream disease classification in monocytes and lymphocytes.