Predicting Metabolic Pathway Activity Using Gene Expression Values in Colorectal Cancer Tissues
Kaavya Borra, Anmol Karan, Rohan Matta & Anmol Karan
Lay Summary:
Our project uses neural networks to predict molecular pathway activity in colorectal cancer tissues using tissue images and gene expression counts throughout the tissue. By exploring deep learning methods for multiple pathways, we aim to predict pathway activity without the need for expensive genetic testing.
Abstract:
This study aims to evaluate the ability of neural networks to predict different molecular pathways from H&E imaging using spatial transcriptomics images as the ground truth and to identify the best strategy for improving these neural network predictions. We preprocessed our data with three Python libraries- Decoupler, single-strand Gene Set Enrichment Analysis (ssGSEA), and Scanpy. We used PyTorch to train and test our models using Spatial Transcriptomics data collected by the Dartmouth-Hitchcock Medical Center. Using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database, we identified the Hallmark p53, Hallmark Apoptosis, Hallmark WNT Beta Catenin Signaling, Hallmark MTORC1 Signaling as target gene sets, and the KEGG Vascular Smooth Muscle pathway as a sanity check. Out of all three preprocessing methods, Decoupler performed the best, with an average Mean Absolute Error (MAE) of 0.106, whereas ssGSEA did the worst with an MAE of 0.164. We confirmed that a ResNet18 could predict the module score for all three methods and that it is an effective device for predictions like this. We achieved these results through training on limited data and plan to upscale the training size to improve performance. Because ResNet18 offers promising results, we plan to utilize other model architectures, such as a vision transformer model or graph neural network, and training methods to optimize performance further.
Q&A:
Bios: Kaavya Borra,Anmol Karan,Rohan Matta,Anmol Karan
Program Track: Advanced Research
GitHub Username:
ksborra -Kaavya Borra
Name: Anmol Karan AnmolKaran Anmol Karan AnmolKaran Name: GitHub Username:, dtype: object -Anmol Karan
rohanmatta11 -Rohan Matta
Name: Anmol Karan AnmolKaran Anmol Karan AnmolKaran Name: GitHub Username:, dtype: object -Anmol Karan
What was your favorite seminar? Why?
Zarif Azher -Kaavya Borra
Name: Anmol Karan My favorite seminar was Zarif’s presentation, … Anmol Karan My favorite seminar was Zarif’s presentation, … Name: What was your favorite seminar? Why?, dtype: object -Anmol Karan
My favorite seminar was the presentation about copper interference because it seemed like a somewhat simple idea, where the complexity lied in implementation. It was particularly memorable because nearly everyone talked about their own projects and asked questions about the seminar near the end, so I could see different ways of interpreting what the speaker said. -Rohan Matta
Name: Anmol Karan My favorite seminar was Zarif’s presentation, … Anmol Karan My favorite seminar was Zarif’s presentation, … Name: What was your favorite seminar? Why?, dtype: object -Anmol Karan
If you were to summarize your summer internship experience in one sentence, what would it be?
I loved being able to explore the intersections between python and pathology through bioinformatics -Kaavya Borra
Name: Anmol Karan The internship experience allowed me to make n… Anmol Karan The internship experience allowed me to make n… Name: If you were to summarize your summer internship experience in one sentence, what would it be?, dtype: object -Anmol Karan
It was a fascinating experience that taught me more about specific biological processes investigated in machine learning, and it also helped me better understand data preprocessing. -Rohan Matta
Name: Anmol Karan The internship experience allowed me to make n… Anmol Karan The internship experience allowed me to make n… Name: If you were to summarize your summer internship experience in one sentence, what would it be?, dtype: object -Anmol Karan