Predicting Single Cell Expression from Multiplex-Immunofloresent Imaging and Bulk-Count Data
(2024) BINP52 20232Degree Projects in Bioinformatics
- Abstract
- Spatial omics allows for the investigation of the Tumor Immune Microenvironment (TIME). Stratifying patients by their TIMEs contributes with insights on tumor immune escape, and help in tumor drug targeting. Tools to investigate such spatial omics approaches suffer from high costs and limited direct clinical applicability. We address this issue by proposing a Machine Learning model which learns to deconvolute bulk expressions based on immunofluorescence (IF) imaging. By utilizing selected, specific celltype stainings, we aim to improve predictive power over Hematoxylin and Eosin (H&E) imaging. Explicitly, our model learns visual representations of cells in a contrastive manner, utilizing Graph Neural Networks (GNN) to predict expressions... (More)
- Spatial omics allows for the investigation of the Tumor Immune Microenvironment (TIME). Stratifying patients by their TIMEs contributes with insights on tumor immune escape, and help in tumor drug targeting. Tools to investigate such spatial omics approaches suffer from high costs and limited direct clinical applicability. We address this issue by proposing a Machine Learning model which learns to deconvolute bulk expressions based on immunofluorescence (IF) imaging. By utilizing selected, specific celltype stainings, we aim to improve predictive power over Hematoxylin and Eosin (H&E) imaging. Explicitly, our model learns visual representations of cells in a contrastive manner, utilizing Graph Neural Networks (GNN) to predict expressions from a cell graph of learned cell representations, enabling the trained model to predict high-plex single cell (sc) expression from just four marker images. We validate our model performance in depth using our in-house ovarian cancer (OC) GeoMx digital spatial profiling (DSP) dataset, as well as the CycIF (Cyclic IF) colorectal cancer (CRC) 2022 Atlas up to the single cell level. While the model struggles with accurate sc expressions, accuracy is dramatically improved when considering groupings of cells, enabling future TIME stratification based on just low-plex IF images, while allowing for standard sc analysis. We also highlight potential avenues for future improvements. (Less)
- Popular Abstract
- Decoding Tumor Microenvironments using AI and Imaging
Understanding the complex interactions between immune and tumor cells is crucial for advancing cancer treatments. Our project leverages cutting-edge machine learning to decode these intricate environments using cost-effective imaging techniques.
Cancer’s ability to evade the immune system, known as tumor immune escape, complicates the selection of effective immunotherapy treatments. Understanding the Tumor Immune Microenvironment (TIME) is vital for identifying drug targets and tailoring treatments to individual patients. Traditional spatial omics methods, which involve high-resolution imaging and detailed molecular profiling of tissue samples, offer valuable insights but are... (More) - Decoding Tumor Microenvironments using AI and Imaging
Understanding the complex interactions between immune and tumor cells is crucial for advancing cancer treatments. Our project leverages cutting-edge machine learning to decode these intricate environments using cost-effective imaging techniques.
Cancer’s ability to evade the immune system, known as tumor immune escape, complicates the selection of effective immunotherapy treatments. Understanding the Tumor Immune Microenvironment (TIME) is vital for identifying drug targets and tailoring treatments to individual patients. Traditional spatial omics methods, which involve high-resolution imaging and detailed molecular profiling of tissue samples, offer valuable insights but are often prohibitively expensive and challenging to implement in clinical settings.
To address these limitations, we developed a machine learning model designed to predict detailed molecular profiles from more accessible imaging data. Specifically, our model uses immunofluorescence (IF) imaging to estimate the complex gene and protein expression patterns typically obtained through more costly techniques. By focusing on key visual cell-type markers, our approach aims to maintain accuracy while reducing the need for extensive resources. Our model employs advanced techniques such as contrastive learning and Graph Neural Networks (GNN) to analyze IF images. We train the model to predict molecular expressions of single cells. We validated our approach using a unique dataset from ovarian cancer (OC) and colorectal cancer (CRC), achieving promising results in group-level analysis.
Methods and Findings
We used a combination of machine learning techniques to achieve our objectives. First, we trained a Convolutional Neural Network model in a contrastive manner to learn visual representations of individual cells. These representations served as nodes in a cell graph, where a GNN aggregated information of neighboring cells to predict cell expressions. Our validation process involved extensive testing with our in-house OC dataset and the CRC 2022 Atlas. While the model faced challenges in predicting individual single cell expressions, it performed significantly better when analyzing groups of cells. This group-level accuracy is crucial to analyze or predict patient responses to treatments and could potentially guide immunotherapy decisions based on readily available IF images.
Applications and Future Directions
Our research demonstrates that it is possible to glean detailed cell expression insights from IF imaging techniques, potentially aiding how we study and treat cancer. By enabling patient stratification with just low-plex IF images, our model offers a more practical and cost-effective tool for clinical settings. Future improvements could focus on enhancing single-cell prediction accuracy and expanding the model to other types of cancers and higher plex imaging. Additionally, integrating this approach with other diagnostic tools could further refine patient-specific treatment strategies.
Master’s Degree Project Bioinformatics 60 credits 2024
Department of Biology, Lund University
Advisor: Anna Gerdtsson
Department of Lund Immunotechnology (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9176797
- author
- Heidrich, Markus
- supervisor
- organization
- course
- BINP52 20232
- year
- 2024
- type
- H2 - Master's Degree (Two Years)
- subject
- language
- English
- id
- 9176797
- date added to LUP
- 2024-10-18 14:40:01
- date last changed
- 2024-10-18 14:40:01
@misc{9176797,
abstract = {{Spatial omics allows for the investigation of the Tumor Immune Microenvironment (TIME). Stratifying patients by their TIMEs contributes with insights on tumor immune escape, and help in tumor drug targeting. Tools to investigate such spatial omics approaches suffer from high costs and limited direct clinical applicability. We address this issue by proposing a Machine Learning model which learns to deconvolute bulk expressions based on immunofluorescence (IF) imaging. By utilizing selected, specific celltype stainings, we aim to improve predictive power over Hematoxylin and Eosin (H&E) imaging. Explicitly, our model learns visual representations of cells in a contrastive manner, utilizing Graph Neural Networks (GNN) to predict expressions from a cell graph of learned cell representations, enabling the trained model to predict high-plex single cell (sc) expression from just four marker images. We validate our model performance in depth using our in-house ovarian cancer (OC) GeoMx digital spatial profiling (DSP) dataset, as well as the CycIF (Cyclic IF) colorectal cancer (CRC) 2022 Atlas up to the single cell level. While the model struggles with accurate sc expressions, accuracy is dramatically improved when considering groupings of cells, enabling future TIME stratification based on just low-plex IF images, while allowing for standard sc analysis. We also highlight potential avenues for future improvements.}},
author = {{Heidrich, Markus}},
language = {{eng}},
note = {{Student Paper}},
title = {{Predicting Single Cell Expression from Multiplex-Immunofloresent Imaging and Bulk-Count Data}},
year = {{2024}},
}