A detective arrives at a crime scene and finds a shoe print. He doesn't ask, "What would happen if this boot walked through mud?" He asks, "Which of the known owners could have left exactly this print?" That's how RevPert works too, except instead of prints it has activity profiles of thousands of genes, and instead of a criminal database it has a catalog of genes that scientists already know how to switch off in cells.
When a cell changes state, for example becoming resistant to a drug, its transcriptome, the complete list of active genes, shifts. Previously, researchers looked at which genes changed the most and suspected them. But a strong change doesn't mean a causal one: a gene can be noisy in the background while the real culprit sits quietly in the middle of the list. RevPert acts differently: it compares the entire shift with the footprints of known gene knockouts and ranks them by similarity.
In practice, RevPert first computes a simple similarity measure that takes the sign of the changes into account. Then a neural network adds a "residual" to it, a fine correction it has learned from thousands of examples. The final score is the sum of these two parts. This hybrid turned out to be stronger than either component alone: a pure neural network is good at finding genes in laboratory data but loses biological meaning on real-world tasks, while a simple correlation is reliable but crude.
The method was tested on identity recovery: a cell was taken, a known gene was switched off, and then researchers looked at where RevPert placed that gene in its list of suspects. In four cell lines it proved more accurate than previous approaches: the median rank improved severalfold, and the share of cases where the true gene landed in the top ten rose from 3.7% to 12.5% in one of the lines.
Then RevPert was applied to real medical problems. In liver cancer cells resistant to sorafenib, it placed the AURKA gene 29th among possible culprits, whereas by the strength of expression change it was only 137th. In leukemia cells resistant to imatinib, the MYB gene came out on top. This is no longer an academic test but a short list for biologists who can check these genes in an experiment.
The detective doesn't claim to have found the culprit for sure. He brings the investigator a list of ten names and says, "Check these first." RevPert does the same for a cell, and that's its value: instead of thousands of genes that changed during disease, the researcher gets a handful of candidates to work with further.
Think of how a detective tracks down a culprit from fingerprints: RevPert does the same, but instead of fingers it uses the expression profiles of thousands of genes. And it doesn't just look for a match; it accounts for the direction of the shift, as if the detective could tell who left a print and who wiped it away.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.