This work sits somewhat off the beaten path: its closest neighbors focus on predicting the response to a perturbation rather than on inverse ranking. Around it there are already maps from perturbations to expression (Perturb-seq, LINCS), graph-based and foundation models, and cross-modal bridges such as UniPert–G2CP. But none of them answers the intervention question: which of the cataloged knockouts is most likely to cause or reverse an observed transition. For chemical perturbations, inverse search has long been established through Connectivity Map and L1000, whereas for genetic ones it is only beginning to take shape. RevPert fills this niche, offering a unified ranking interface for any transcriptomic contrast.
RevPert builds a gallery of knockout profiles (predicted by a linear model for Essential or averaged across folds for LINCS-KO) and compares them with the query ΔY = YB − YA. First, the signed Pearson correlation is computed between the query and each profile in the gallery. Then a shared encoder (a multilayer perceptron) compresses the profiles via PCA to 256 dimensions and produces a learned similarity in a 128-dimensional normalized space. The final score is a weighted sum of the z-scores of correlation and learned similarity, where the weight α is tuned during training. Training maximizes an InfoNCE loss over the entire catalog with soft regularization from the Pearson teacher. For external resistance contrasts, the same two-arm interface is used: arm A ranks by phenocopy (sign match), and arm B ranks by opposite sign (an activation heuristic).
On Essential Perturb-seq, RevPert achieved the best median rank in all four cell lines (for example, HepG2: 266 versus 572 for Pearson), improving Recall@10 by a factor of 1.5–3.4 (HepG2: 12.5% versus 3.7%). Wilcoxon paired tests rejected equality in favor of RevPert in every cell line (P < 10⁻¹⁸), and bootstrap intervals did not include zero. On the ten LINCS-KO cell lines, RevPert showed average median ranks of 345–560, which is 7–15 times lower than the official PDGrapher (thousands). Recall@10 was 4.14% versus 2.18% for Pearson and 1.64% for PDGrapher. On external resistance contrasts: for HCC/sorafenib, AURKA landed at rank 29 on arm B (versus 137 by DEG magnitude), and METTL3 at rank 65 on arm A (versus 4040 by DEG). For CML/imatinib on the GWPS catalog, MYB took 1st place on arm A, STAT5A 16th on arm B, STAT5B 27th on arm A, RUNX1 18th on arm B, and BCR 34th on arm A, whereas DEG ranks were considerably worse. Ablation showed that purely learned similarity improves identity but collapses resistance anchors, while residual fusion preserves both.
RevPert differs from its neighbors in that it solves the inverse problem of ranking a catalog for an observed transition rather than predicting the response to a given perturbation. It is the first to systematically test external use, namely the nomination of drug-resistance drivers, as a primary result rather than a side test. The method shows that Pearson's signed geometry is essential for preserving biologically meaningful anchors, and that pure retrieval without it loses directionality. This sets a standard for future inverse scorers: don't discard correlation, learn a residual on top of it.
As Perturb-seq atlases expand, the same interface can be retrained on broader catalogs, including tissue-specific cell lines to improve transfer. It may be possible to incorporate graph-based or foundation encoders into the residual component while keeping the linear gallery. External applications will require stricter validation protocols with independent experimental checks.
The method could speed up the search for candidate driver genes in studies of drug resistance, aging, and cancer progression. It also provides shortlists for downstream drug-target models and focused wet-lab assays.
The next step is to validate the nominated drivers in independent experiments and extend to other state transitions, such as colorectal progression or senescence. It is also worth exploring encoder transfer between contexts with linearly matched catalogs.
In genomics, a question remains open: how to move from an observed transcriptomic shift to causal genes rather than merely differentially expressed ones. RevPert attacks this problem directly by ranking a catalog of interventions, but it does not solve it completely: the ranks remain hypotheses that require experimental verification. The link to the unsolved problem of identifying causal regulators in complex transcriptomic programs is clear.
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.