Computational RNA Biology & Medicine Lab

Overview

icSHAPE gives you a per-nucleotide reactivity profile for an RNA, but the signal is noisy and expensive to collect, and it only exists for the cell types someone has already paid to measure. The question in the lab was whether base-pairing probabilities from thermodynamic folding could stand in for that experimental signal, which would let structure-aware binding models generalise to cell types with no icSHAPE data at all.

The substitute is pSHAPE: fold the region, take the base-pairing probability matrix, and compute 1 minus each column's sum, which estimates how unpaired each nucleotide is and so carries the same meaning as a reactivity value. I built the pipeline that produces it — tracing each 101-nucleotide binding window back to its transcript through the GENCODE v42 annotation and transcriptome, expanding the window on both sides for folding context, truncating at transcript boundaries rather than padding, mapping icSHAPE bigWig signal onto the expanded window, folding, then cropping the result back to the original 101-nt core so it drops straight into CellRBP's training format. Folding runs on RNAstructure rather than ViennaRNA, which I switched to after ViennaRNA's Python API produced results that disagreed with its own web tool.

Separately I measured how much binding is actually conserved between cell types, computing Jaccard overlap of binding sites for 62 proteins between K562 and HepG2. Doing that at a single threshold is misleading, because two peaks a few bases apart are the same site in biology and different sites in a set-intersection, so I ran it at 1, 5, 10, 30 and 100 bp and compared the curves. Those overlap figures then got lined up against cross-cell AUC — train on one cell line, test on the other — to ask whether proteins that bind consistently are the ones whose models transfer.

The supporting work was less glamorous and mattered just as much: stripping pseudoknots out of reference structures, clipping reactivities to a maximum of 3, zeroing negatives and marking missing values, and sweeping the Deigan SHAPE-directed folding parameters around their m = 1.8, b = −0.6 defaults. Reactivity data is full of values that are physically meaningless and will happily train a model on nothing.