Enhancing Spatial Transcriptomic Inference with LLM Loss Functions
Sidh Jaddu
Lay Summary:
My work aims to make skin cancer detection more effective and efficient using AI.
Abstract:
Skin cancer affects one out of every five Americans during their lifetimes, making it the most common cancer in the United States. Common diagnostic methods, which are largely based on biopsies and imaging, exist, but these methods often miss early-stage tumors or are inaccurate in mapping tumor behavior at the molecular level. One known method for understanding molecular-level tumor behavior is traditional spatial transcriptomic (ST) methods, which map gene activity in tissues; however, these methods are costly, and thus, difficult to upscale. With the increasing popularity and use of deep learning systems, deep learning-based ST methods that use hematoxylin and eosin (H&E) tissue data to infer ST data have been implemented, but these systems often lack a contextual understanding of the data. Hence, LLM loss functions are being employed to help provide a better contextual understanding. This research sought to improve existing deep learning inference models by enhancing their LLM loss functions, effectively bettering the model’s understanding. Existing LLM loss functions are based on GPT2. This study aims to further improve the LLM loss functions by exploring the use of additional data and other LLM models, namely ALBERT and XLNET. The addition of patient data showed improvements in model performances regardless of model type used. In the future, this methodology can be applied to other types of cells beyond skin cells and more advanced models can also be explored.
Q&A:
Bios: Sidh Jaddu
Program Track: Advanced Research
GitHub Username:
amIAIguru -Sidh Jaddu
What was your favorite seminar? Why?
Jason Wei’s seminar was very interesting. Given that he works at OpenAI, his work was very relevant to my LLM project. -Sidh Jaddu
If you were to summarize your summer internship experience in one sentence, what would it be?
It was a highly enjoyable experience and I certainly learned a lot about LLMs.
-Sidh Jaddu