Bayesian ST Inference- Quantifying Uncertainty in Spatial Transcriptomics
Nehan Mohamed & Thanosan Prathifkumar
Lay Summary:
We built a framework that is not only able to predict what genes are active within tissues give a tissue slide image, but also quantify how certain we are about the predictions. This allows us to understand what is reliable and needs to be developed further.
Abstract:
Spatial transcriptomics data provides gene expression information within spatial contexts. However, this is expensive and time consuming. Thus, virtual RNA inference (VRI) provides an alternative to this by predicting ST-level information from hematoxylin and eosin stained tissues. One problem with these models is that these models do not quantify uncertainty of their predictions. In this study, we introduce a Bayesian framework for VRI models that can estimate aleatoric and epistemic uncertainty. We implement a Bayesian neural network with Monte Carlo sampling and heteroscedastic loss to decompose gene-level uncertainty. Upon a pathway analysis, our results show that core metabolic and cell-cycle pathways have lower uncertainty, while stress response and regulatory pathways have much higher uncertainty. There is aleatoric uncertainty in metabolic and homeostatic processes, while epistemic uncertainty in immune signaling and RHO GTPase pathways. Additionally, when decomposing uncertainty by architecture, cancerous tissues have higher uncertainty in cell death pathways, reflecting tumor microenvironment heterogeneity. These findings show that uncertainty quantification is a powerful tool for interpreting VRI predictions.
Q&A:
Bios: Nehan Mohamed,Thanosan Prathifkumar
Program Track: Advanced Research
GitHub Username:
NehanMohamed -Nehan Mohamed
thanosan23 -Thanosan Prathifkumar
What was your favorite seminar? Why?
My favorite seminar was by Dr. Hideki Furuya from Cedars-Sinai. I found his discussion on cancer biomarker research, particularly his work on bladder and colorectal cancers very informative and engaging. His clear explanations helped me connect complex molecular science to meaningful clinical applications. -Nehan Mohamed
Zarif’s seminar about multimodal modelling and perspectives about research was my favourite. Multimodal modelling is something that I think is very powerful and it is great to hear that the lab is working on this. I especially enjoyed the part about research and entrepreneurship and it provided valuable advice and takeaways for the rest of the internship. -Thanosan Prathifkumar
If you were to summarize your summer internship experience in one sentence, what would it be?
My summer internship was a transformative journey that pushed me beyond my comfort zone and opened up entirely new opportunities for learning and growth. -Nehan Mohamed
I learned a lot throughout the summer experience. I learned a lot about spatial transcriptomics and machine learning. I also learned a lot about research itself and the journey from start to finish when working on a research project. -Thanosan Prathifkumar