PLP1
Mutations in PLP1 cause Pelizaeus-Merzbacher disease, a disorder of myelin. Most work treats them as one problem. I sorted 371 pathogenic variants by mechanism and matched each class to a therapy. Meet the Twin, a way into the digital twin I built to score diseased cells against a healthy baseline.
independent computational projectA protein that builds myelin
Oligodendrocytes wrap nerve fibers in layers of myelin so signals travel fast. PLP1 is a major protein in that wrap. Too little, too much, or a misfolded version, and the wrap fails.
371 variants, three mechanisms
I pulled 701 PLP1 records from ClinVar and kept the 371 that are pathogenic or likely pathogenic. By consequence: 173 loss of function, 63 duplications, 85 misfolding candidates, and 50 I could not classify.
Each class points to a different strategy: replacement for loss of function, suppression for duplications, stabilization for misfolding.
Match the therapy to the mechanism
The digital twin I built is a reusable framework. It scores a disease cell's gene expression against a validated healthy baseline across six interpretable biological measures. On new, independent datasets it picked out a real disease-specific treatment effect, and where no data exists for a condition, it says so instead of guessing. I cross-validated it on human data (GSE118257) and tested it on held-out data (GSE277705).
The brain and panel here illustrate the idea: pick a variant, then a therapy, and watch matched and mismatched pairings play out. The panel numbers are illustrative, not output from the twin.
Which missense variants destabilize it?
I ran FoldX on the AlphaFold model of PLP1 for the 83 missense variants I could score. 27 destabilize it (ΔΔG above 1.5 kcal/mol), 14 mildly, and 42 not at all. Glycine-to-bulky and proline-introducing changes dominate. The worst is Gly246Trp at 48.28. Positions here are schematic.
Do they stress the cell? Can one be rescued?
In jimpy mouse oligodendrocyte progenitors (GSE111605, 3 vs 3), 19 unfolded-protein-response genes shift up as a group (p = 0.0073). No single gene passes correction, so it is a pathway-level signal.
For Gly246Trp I scanned the 14 neighboring residues. Leu31Gly came out on top, predicted to relieve about 24 kcal/mol. It is a computational hypothesis, not tested.
What this does not show
Everything here is computational. FoldX scores are predictions, 56 misfolding variants stay uncertain, the therapy pairings are hypotheses rather than recommendations, and the 3D models and the panel numbers are schematic. The rescue idea is untested.