Library compounds whose LINCS consensus expression profiles most strongly anti-correlate one transcriptional signature associated with Huntington's Disease, after a cytotoxicity-alignment filter and clinical-maturity ranking. Compounds our structural screen flags are labelled below rather than hidden. This is a property of that signature, not of the disease. Computational and hypothesis-generating, pending wet-lab. Method & limits →
Quantified reversal ranking · per-candidate mechanistic rationale · ClinicalTrials.gov whitespace detail · S3 (505(b)(2)+Orphan) regulatory framing · structure-function intelligence.
What this ranking is built on. The expression signature behind this indication is human and shares 940 genes with the 978-gene LINCS landmark set, clearing the 300-gene floor below which our rankings stop being reproducible under resampling. Twenty-one of our other indications fail that floor. Six more are withheld for different reasons: four rest on mouse or rat signatures, and two carry a signature whose values point almost entirely one way, leaving nothing for a reversal score to oppose. The cross-species cases are described in the cross-species note. None of the twenty-seven is for sale at any price.
The one control we have passed. We permute a signature's values among its own genes — same genes, same depth, same distribution, disease content destroyed — and re-rank. Across six signatures the candidate list does not survive it: the median overlap with the real top twelve falls to zero of twelve. Of 60 individual draws, 9 recovered one compound and one recovered two — we report that rather than rounding it into “none”. The ranking tracks the disease measurement rather than the identity of the assayed genes. That is a specificity result, and it is the only control this method has passed.
What you are not buying. This is a computational hypothesis, pending wet-lab. We have not shown these candidates work, and we cannot yet show the list is reproducible across independent signatures of the same indication: 23 indications in our corpus hold two signatures each, and their top-twelve lists agree no more often than lists drawn from unrelated diseases (14 of 23 pairs share nothing; p = 0.92 against 3,173 different-indication pairs). One side of every one of those comparisons was too shallow to measure anything, so that result does not condemn the method — but it does mean the replication question is open, and you should price it accordingly.