RIDDEN

Receptors are the most frequently drugged class of proteins, and most computational methods for studying cell-cell communication ask whether a ligand and its receptor are co-expressed in a pair of cells. Co-expression is a weak proxy: transcript abundance does not track protein abundance, and an abundant receptor sitting unliganded is inactive. What a transcriptomics experiment can see directly is the consequence of receptor activation — the shift in the genes downstream of that receptor.

RIDDEN (Receptor actIvity Data Driven inferENce) applies that footprint logic to receptors. It was built by Szilvia Barsi, Bence Szalai and colleagues at the HUN-REN Research Centre for Natural Sciences and Semmelweis University, with Daniel Dimitrov and Julio Saez-Rodriguez at Heidelberg University, and was published in PLOS Computational Biology after a December 2024 preprint. The approach sits in the same tradition as PROGENy and DoRothEA for pathways and transcription factors, and its nearest neighbour is CytoSig, which covers 43 cytokines; RIDDEN extends the idea to 229 receptors by mining a perturbation compendium rather than curated stimulation studies.

The model is a fixed artifact. Fitting happened once, over the compendium; the released tool multiplies a new expression matrix by that matrix and reports activities, with no training step and no user-supplied checkpoint.

#Key Features

• Downstream footprint, not receptor mRNA: Activity is inferred from the expression of receptor-regulated genes, which distinguishes an activated receptor from a merely present one.
• Breadth over depth of assay: Five perturbation modalities — ligand stimulation, drug treatment, shRNA, CRISPR knockout and overexpression — are pooled into a single model, so a receptor supported by any of them enters the signature set.
• Ligands mapped onto receptors: Curated ligand-receptor interactions translate ligand perturbations into receptor perturbation labels, multiplying the usable training signal.
• Confidence grading: Cross-validation assigns each receptor a level from A to E, letting users weight or filter activities by how well that receptor's signature reproduces.
• Bulk and single-cell in one interface: The command-line tool takes any normalized expression table with genes as columns and samples or cells as rows.

#Technical Details

Curated ligand-receptor interactions came from OmniPath; perturbation profiles came from the LINCS L1000 collection, filtered to receptor and ligand perturbations and aggregated into 38,989 consensus signatures over 599 receptors and ligands. For every receptor-gene pair, an ordinary least squares regression relates gene expression to a perturbation indicator coded +1 for stimulation, −1 for inhibition and 0 for no perturbation; the fitted coefficients form the model matrix. Receptors were then filtered by cross-modality validation, keeping only those whose activity can be predicted from a modality other than the one that trained them. The surviving model summarizes 14,463 perturbation profiles — 229 receptors, 228 cell lines, 747 distinct perturbations — into a 229 × 978 matrix over the LINCS landmark gene set, which is what ships in the repository. Inference is a dot product of the expression profile with that matrix, standardized against a background from 1,000 permutations of gene labels.

Cross-validated across the 229 receptors, median ROC AUC is 0.71. Benchmarked on CytoSig's own cytokine perturbation data, RIDDEN reaches a median ROC AUC of 0.61 against CytoSig's 0.59, and 0.68 for its highest-confidence receptors. On the single-cell Immune Dictionary data, the overlapping-receptor comparison gives 0.64 against CytoSig's 0.67, while A-confidence receptors reach 0.84. Receptor weight vectors cluster by family, and transcription factors sharing a KEGG pathway with a receptor show higher inferred activity in that receptor's weights than those that do not. In a renal cell carcinoma cohort, RIDDEN's pretreatment PD-1 activity separated overall survival under nivolumab (log-rank p = 0.0004; Cox β = −0.36) where PD-1 and PD-L1 transcript levels did not, and the association was absent under everolimus.

#Applications

The tool suits anyone reading transcriptomics for signaling state: annotating which receptors are active in which cell populations of a single-cell atlas, ranking receptor-level mechanisms behind a drug response, or testing receptor activity as a patient stratification biomarker where expression alone is uninformative. It installs as a Python package and runs from the command line under an MIT license.

#Impact

RIDDEN's contribution is showing that a large perturbation compendium can substitute for the curated stimulation experiments that footprint methods normally depend on, pushing receptor coverage from tens to hundreds. It is a linear model, deliberately so, and its interpretability is part of the argument — weights can be read as receptor-to-gene regulatory statements and checked against pathway databases. The limits follow from the training data: coverage stops at the 229 receptors the compendium supports, the LINCS landmark set is 978 measured genes rather than a full transcriptome, so receptors with similar downstream programs are hard to separate, and confidence varies enough across receptors that the A-to-E grades are not decoration.