Latent Diffusion-Based Cross-Model Learning for Spatial Transcriptomics in Human Skin
Vedhsai Thiriveedi
Lay Summary:
I developed an AI system that learns to connect what tissue looks like under the microscope with the genes that are active in that same region. This approach could one day help doctors better understand diseases by combining what they see in pathology slides with the molecular signals hidden inside the cells.
Abstract:
Spatial transcriptomics (ST) has transformed molecular biology by enabling genome-wide gene expression measurement while preserving tissue architecture, but integrating molecular signals with histological context remains a central challenge. In this study, we present a generative deep learning framework that combines Contrastive Language–Image Pretraining (CLIP) with a Latent Diffusion Model (LDM) to achieve cross-modal learning in human skin ST. Using the 10x Genomics Visium platform, we tiled histological images into 256 × 256 patches, mapped them to corresponding transcriptomic spots, and extracted the top 50 expressed genes per region. Both image patches and gene expression vectors were embedded into a shared latent space with CLIP, and these embeddings conditioned the LDM to reconstruct histological features directly from transcriptomic input. Our model achieved strong cross-modal alignment, with a mean cosine similarity of 0.933, LPIPS of 1.095, and SSIM of 0.102, confirming that reconstructed patches preserved transcriptomic signal despite limited structural fidelity. Reconstructions visualized as heatmaps revealed molecular–morphological correspondence, providing interpretable outputs even under noisy conditions. This proof-of-concept demonstrates that latent diffusion can unify transcriptomics and histology in skin tissue. Future work will incorporate pathologist-verified labels, connected component analysis for cluster-level modeling, and subtype classification, advancing generative methods toward clinical integration.
Q&A:
Bios: Vedhsai Thiriveedi
Program Track: Advanced Research
GitHub Username:
Name: Vedhsai Thiriveedi Vedhsai-codes Vedhsai Thiriveedi Vedhsai-codes Name: GitHub Username:, dtype: object -Vedhsai Thiriveedi
What was your favorite seminar? Why?
Name: Vedhsai Thiriveedi My favorite seminar was the one where we learn… Vedhsai Thiriveedi My favorite seminar was the one where we learn… Name: What was your favorite seminar? Why?, dtype: object -Vedhsai Thiriveedi
If you were to summarize your summer internship experience in one sentence, what would it be?
Name: Vedhsai Thiriveedi This summer, I developed and tested artificial… Vedhsai Thiriveedi My summer internship focused on developing and… Name: If you were to summarize your summer internship experience in one sentence, what would it be?, dtype: object -Vedhsai Thiriveedi