Integrating Transformer-Based Pathway Reconstruction with Virtual RNA Interference for Enhanced Spatial Gene Expression Prediction
Ashwin Menachery, Rishi Pathuri, Shreyas Pendem, Faraaz Rattani & Amrit Singh
Lay Summary:
We built an AI system that learns to read cancer tissue slides and predict how genes behave across different regions of the tumor. By teaching the AI to see tissue as a continuous landscape instead of isolated blocks, our approach creates smoother, more realistic maps that could help doctors better understand cancer growth and treatment response.
Abstract:
Spatial transcriptomics enables the mapping of gene expression within intact tissue architecture, but accurate prediction of expression across unmeasured regions remains challenging. We present two transformer-based strategies for enhancing Virtual RNA Interference (VRI) predictions by integrating spatial context at the pathway level. Method 1 pretrains a transformer to reconstruct masked pathway values from neighboring spots and penalizes VRI predictions that deviate from these transformer inferences. Method 2 jointly trains a transformer and VRI model to directly predict pathway expression from tissue image patches. Using colon whole-slide images and pathway-level data, we demonstrate that both methods improve spatial coherence and prediction accuracy compared to baseline VRI. Method 1 yielded a 7.3% reduction in MSE relative to baseline, while Method 2 achieved a 9.1% reduction, particularly benefiting low-expression regions. These results highlight the utility of transformer-based spatial priors for improving spatial transcriptomic inference.
Q&A:
Bios: Ashwin Menachery,Rishi Pathuri,Shreyas Pendem,Faraaz Rattani,Amrit Singh
Program Track: Advanced Research
GitHub Username:
ashwinmenachery -Ashwin Menachery
garrafan -Rishi Pathuri
Shazzey -Shreyas Pendem
faraazrattani -Faraaz Rattani
AmritSingh10 -Amrit Singh
What was your favorite seminar? Why?
Zarif’s seminar was my favorite since he explored machine learning in the context of entrepreneurship, a combination I find especially exciting. I connected with his talk because it showed how computer science can be applied in business settings, which matches my broader interests. -Ashwin Menachery
I really enjoyed this seminar because it connected environmental factors, like trace metals, to the development of ALS, showing how biology and chemistry interact in real-world health problems. It was fascinating to see how some metals, like selenium and zinc, could be protective while others, like copper, increased risk. -Rishi Pathuri
Zarif’s multimodal seminar because it illustrated how combining data of different types can lead to better diagnoses. -Shreyas Pendem
I liked how Hideki Furuya presented the study clearly and visually, making the complex data across three instruments easy to follow, and how he demonstrated the reliability of the Oncuria assay, showing consistent results with minimal variance. I also appreciated how he highlighted the clinical relevance, showing high analyte levels in BC patients while sparing controls and explaining why it matters for patients. Finally, he made the topic interesting and practical by connecting the technical results to real-world applications and showing that existing lab equipment can be used without compromising accuracy. -Faraaz Rattani
My favorite seminar was Zarif’s because he talked about entrepreneurship along with ML, which is what I am interested in broadly. Through business applications and computer science integration, I really resonated with Zarif’s seminar. -Amrit Singh
If you were to summarize your summer internship experience in one sentence, what would it be?
My summer internship was a great learning experience where I gained new skills and explored real-world applications of computer science. -Ashwin Menachery
This internship was a deeply rewarding experience as it challenged me to implement and compare multiple advanced, spatially-aware machine learning architectures, significantly deepening my skills in model development and custom loss function design for bioinformatics. -Rishi Pathuri
The internship provided the opportunity to work with advanced machine learning models on high-performance compute nodes. -Shreyas Pendem
I learned a lot dealing with transformers and how to collaborate with a team in machine learning setting. -Faraaz Rattani
A tremendous opportunity to learn about Computer Science applications in the field of Bioinformatics. -Amrit Singh