EPCOT Foundation Model — TSS-to-Expression Data Pipeline
Jan 1, 2026
·
1 min read
Pseudobulk pipeline from raw snMultiome (1 bp sparse ATAC → 600 kb HDF5 windows; CP10K-normalized 19,264-gene RNA vectors across 7 cell types). Five bit-equivalent training optimizations for a 2–5× speedup at identical numerics. Reset-head fine-tuning for small-data generalization. Located and fixed a ~50 kb TSS-bin offset.
Authors
B.S.E. Candidate, Data Science
I study Data Science at the University of Michigan, Ann Arbor. I came here through the UM-SJTU dual-degree program: the first half of my undergraduate degree was Mechanical Engineering at Shanghai Jiao Tong University’s Joint Institute, the second half is Data Science at Michigan. It was less a change of heart than a change of tools.