Text-based Supervised Fine-tuning for Virtual RNA Inference Models
Ishan Ramrakhiani & Pranay Kancharla
Lay Summary:
We are trying to enhance a virtual RNA inference model, which is a model that predicts gene expressions by introducing text-based supervision through histological descriptions of the key features in the patch.
Abstract:
We study a minimal yet strong baseline for predicting spatial gene expression from histology patches using a frozen pathology foundation model (UNI2-h) and a light two-layer regression head. In parallel, we incorporate pathology-language supervision via MedVLM-R1-generated captions and a positive-only alignment objective (cosine alignment) to encourage morphological interpretability. Across a systematic ablation of depth, width, explicit activations, dropout, and batch normalization in the regression head, we observed no substantial or seed-consistent gains over the simplest configuration (1536→1024→1000; MSE loss; Adam 10^−4). Given its consistently strong performance across seeds, we select this baseline for subsequent experiments.
Q&A:
Bios: Ishan Ramrakhiani,Pranay Kancharla
Program Track: Advanced Research
GitHub Username:
himanalot -Ishan Ramrakhiani
PRK109 -Pranay Kancharla
What was your favorite seminar? Why?
I loved the seminar on writing by Dennis Hazelett because I was working at Bindwell at the time, helping to improve the PLAPT series of models for protein-ligand affinity prediction, and I created written analysis reports on model runs, recommending architectural/hyperparameter decisions. This experience, along with the seminar, helped me understand that the work I do in life doesn’t have significance unless I can spread it to others and have it live beyond my work. Writing takes findings and magnifies them into impact through future use of the information contained in the piece. -Ishan Ramrakhiani
My favorite seminar was the one where Rishy Deosthali and Vismay Ravikumar presented because there project was interesting and it was also neat to hear from an intern POV. I think after listening to this seminar, I got a better understanding on how to move forward and think about what changes I would’ve made. -Pranay Kancharla
If you were to summarize your summer internship experience in one sentence, what would it be?
The internship was taught me to ask others for help when I need it, and it was thought-provoking in that it taught me how to look for new ways to improve models when data is limited (mostly when we didn’t know where the data was), as well as to question the way we’re approaching our problem (as we have been once we realized our current structure for fine-tuning hasn’t improved the models). -Ishan Ramrakhiani
An informative and fun experience to help work on research to better the world. -Pranay Kancharla