Integrating Spatial Transcriptomics in single cell resolution with Explainable Machine Learning for Enhanced Insights in Lung Cancer Biology.

Haris Manousaki, GEORGIOS GEORGAKILAS, Theodore Dalamagas

Keywords: Machine Learning, Spatially Resolved Transcriptomics, Explainable AI, AI in Life Sciences, NSCLC

Abstract:

Introduction
In the rapidly advancing field of transcriptomics, Spatial Transcriptomics (SRT) signifies a
transformative advancement, particularly within the scope of Precision Medicine. SRT
enables gene expression profiling at the single-cell level and captures the location of
transcriptional activity within the under-study tissue. Despite SRT’s potential, the field faces
challenges similar to its precursor, single-cell RNA sequencing (scRNA-seq), including
uncertainties related to procedural parameters and biases in cell labeling due to manual
marker-gene annotation [1]. Thus, a central focus of this project is to observe the
uncertainty in cell classification and uncover insights into cell-to-cell communication using
explainable machine learning.
Methods
This study uses non-small cell lung cancer tissue samples obtained from the CosMx platform
repository [2], to train machine learning (ML) models for cell classification by integrating SRT
gene expression data with both the conventional marker gene approach, utilizing the Seurat
package, and an enhanced version provided by Nanostring. By comparing the classification
accuracies of these models, we can quantify the uncertainties associated with
marker-gene-based labeling. Additionally, we explore the impact of cell-to-cell interactions
by initially identifying boundaries of distinct cell populations, using Shannon’s entropy, and
subsequently comparing the misclassification rate of the ML models between cells that are
localized at the boundaries and at the center of populations.
Results
Significant differences in classification performance were observed. The Seurat-mediated
marker gene-based annotation yielded a lower Matthews correlation coefficient (MCC
0.6623) and Precision (0.7327) compared to the Nanostring-based annotation (MCC 0.868,
Precision 0.8642), highlighting the limitations of marker gene approaches. Furthermore,
areas of high entropy were found to align with regional and cell class boundaries, verifying
the impact of cell-to-cell communication in gene expression. This observation unveils the
potential of our method to extract biological insight from SRT data and leverage explainable
ML to explore the tumor microenvironment.

[1] S. Fang et al., “Computational Approaches and Challenges in Spatial
Transcriptomics,” Genomics Proteomics Bioinformatics, vol. 21, no. 1, p. 24, Feb. 2023,
doi: 10.1016/J.GPB.2022.10.001.
[2] “CosMx SMI NSCLC FFPE Dataset | NanoString.” Accessed: Aug. 06, 2024. [Online].
Available:
https://nanostring.com/products/cosmx-spatial-molecular-imager/ffpe-dataset/nsclc-
ffpe-dataset/

Submission Number: 173