Motif priors carry single-cell post-transcriptional regulon inference; footprints calibrate it, co-expression cannot replace it

Question: How do recent methods infer post-transcriptional regulons and RNA-binding protein targets from single-cell transcriptomes, and how do motif-based, footprint-based, and co-expression-based approaches compare?

Answer

Taken together, the retrieved papers support a layered rather than a competitive answer. Motif-based inference is the only paradigm in the set that has been implemented at cell-type, gene and isoform resolution directly from single-cell transcriptomes, and it earns its credibility by being scored against the other two paradigms: eCLIP-derived footprints supply the target definitions it recovers, and RBP-knockdown Perturb-seq supplies the functional arbiter it passes [2]. Footprint-based evidence is, in this literature, less a stand-alone inference route from single cells than a calibration layer — CLIP inputs are being pushed down in material requirements but remain bulk and context-limited [21], while footprint-trained sequence models transfer across cell types and outperform single-task tools [19]. Co-expression-based inference is the weakest link: RBP transcript abundance is an unreliable proxy for activity, correlation-based network recovery is fragile and often fails causal tests in the adjacent GRN literature [12], and normalization choices change what "coregulation" is even detected [4]. Across these papers, the position of this survey is that motif logic is necessary but insufficient, footprint data are necessary but unavailable per cell, and co-expression is neither necessary nor sufficient — the productive architecture is motif-defined, region- and isoform-resolved regulons validated by footprint and perturbation evidence.

Motif-based inference is currently the only paradigm that scales to cell-resolved post-transcriptional regulons

scRBP infers motif-supported RBP regulons from short- and long-read single-cell transcriptomes at gene and isoform resolution, drawing on 20,746 motifs for 616 RBPs, scoring regulon activity per cell, and resolving targets by transcript region (5'UTR, CDS, intron, 3'UTR) [2]. Region resolution is not cosmetic: across 30 tissues, transcript-region preferences separated 3'UTR-associated stability factors from intron-dominant splicing regulators [2]. The same regional logic is independently supported by single-cell CRISPR work in which knockdown of 15 RBPs produced thousands of perturbation-driven alternative splicing events and skipped exons were enriched for binding sites of the perturbed RBPs [20]. Two further strands extend the motif-bearing sequence space that single-cell methods must address: cell-state-specific polyadenylation site usage, where APA-regulated genes shared conserved 3'UTR motifs often containing Alu elements [27], and de novo polyadenylation site mapping from normally discarded Read1, which yielded stronger AAUAAA signal and tissue-specific polyadenylation preference alongside RBPs [5].

Motif models are being replaced by context-aware sequence models, which changes what "motif-based" means

Replacing independence assumptions with flexible, context-aware motif models improves RBP binding prediction: a TFFM-based PTBP1 model capturing positional dependencies correlated with experimental binding affinities at r = -0.95 versus r = -0.91 for the earlier HMM [1]. At the far end of this trajectory, a foundation model trained exclusively on 223 eCLIP-seq experiments spanning 150 RBPs predicts base-resolution binding profiles from sequence, substantially outperforms its single-task predecessor in binding profile and motif recovery, generalizes to unseen cell types and iCLIP data, and remains traceable to the specific RBPs and motifs driving predictions [19]. Denoising CLIP signal into a generalizable predictive model is likewise the explicit goal of kernelized regression approaches that outperform prior methods transcriptome-wide [23].

Footprint data are the reference standard and the transferable prior, but not yet a single-cell assay

The most direct footprint-based inference route in routine use is gene-set enrichment: RBP target genes extracted from eCLIP-seq for 168 RBPs assembled into a collection for over-representation, GSEA and GSVA on any gene list [6]. Where measured and predicted binding were compared head-to-head inside a functional model, in vivo binding sites had higher importance for predicting RNA degradation rate and translation level than computationally predicted sites, likely due to confounding effects [17]. Footprint acquisition is moving toward lower input — soniCLIP recovers targets comparably to ENCODE eCLIP while requiring 10% of the protein input [21] — and toward isoform and compartment resolution, with direct nanopore CLIP revealing isoform-selective binding and variant-specific interactions [24] and proximity-CLIP showing that YBX1 and ELAVL1 target subsets differ by compartment yet share identical binding motifs [25]. None of these retrieved methods is a per-cell assay, which is why footprints function as priors and benchmarks rather than as the single-cell inference engine.

Co-expression inference is undercut by the absence of an activity readout and by fragile correlation structure

RBP mRNA level is a poor stand-in for RBP activity, and this is the same objection that motivated activity-aware TF modelling: TF mRNA levels may not represent protein activity, and regularized latent activity estimation is introduced to overcome it [13], while deconvolution approaches argue that activity estimates should replace arbitrary enrichment scores [14]. Correlation-based network recovery is unstable in ways that matter for regulon claims: GRNs inferred from multimodal single-cell data were highly sensitive to method choices including random seeds and pipeline step substitution, and although they performed well predictively and partially recovered known interactions, they struggled to capture causal relationships from perturbation assays [12]. For co-expression specifically, selecting gene pairs by p-value rather than correlation strength is more robust and false-positive control is often neglected in benchmarking [15], normalization method choice differentially affects detection of positive versus negative coregulation [4], and simulation work shows intracellular regulatory signals rarely propagate to detectable co-expression even before aggregation [16]. Direct measurement confirms the underlying gap: protein-RNA correlations peak dynamically in transitioning progenitors and transcriptomics alone misses fundamental regulatory programs [3]. An adjacent kinetic route — inferring transcription, export and degradation rates from single-cell multiome data — quantifies post-transcriptional regulation directly and finds degradation and export prominent in shaping mRNA accumulation, but it is not RBP-specific [18].

Motif hits alone are a weak target definition, and the retrieved papers quantify exactly how weak

Canonical seed pairing is not required for repression: top 3'-only sites bind as well as top canonical sites and impart similar repression, constituting an estimated 0.5–1% of the endogenous targetome for miRNAs that recognize them [9]. Seedless sites can lead to substantial downregulation in a high-throughput survey of 4,994 miRNA:3'UTR regulatory outputs, motivating a combined seed-plus-seedless score [11], and 20.0% of miRNAs and 8.2% of isomiRs appear to bind mRNAs independently of their seed regions in chimeric-read data [10]. Specificity is itself quantitative and non-additive: low-specificity RBPs can paradoxically sharpen the target discrimination of high-specificity partners by occupying non-preferred sites, an emergent property not predictable from affinity alone [7]. Structure adds a further layer, as TDP-43 binding and pUG fold formation are mutually exclusive [22] and SERBP1 recognition is driven by G-quadruplex topology rather than nucleotide sequence [26]. Motif-based single-cell inference therefore should be read as probabilistic, competition- and structure-conditioned target nomination.

Conflicts and gaps

No retrieved paper benchmarks motif-based, footprint-based and co-expression-based RBP regulon inference against one another on the same single-cell data with a common ground truth; scRBP is the only single-cell RBP regulon framework in the set, and its own validation is enrichment-based rather than a head-to-head comparison [2]. The strength of its functional validation is modest and should be stated as such: inferred targets were over-represented among genes responding to RBP knockdown at 1.4–2.5-fold across 161 RBPs [2]. There is a live disagreement about whether measured binding or sequence models should carry the prior: measured in vivo sites outperform computationally predicted sites in a stability model, attributed to confounding [17], whereas a CLIP-trained model generalizes to unseen cell types and outperforms motif recovery, implying footprint information can be abstracted away from its original context [19]. Isoform-resolved footprinting exists but is bulk [24], and single-cell regulon inference at isoform level rests on motif assignment to isoforms rather than on measured isoform-specific occupancy [2]. Finally, whether combinatorial RBP competition can be inferred at all from a single transcriptome is unresolved, given that competition effects are emergent and not predictable from affinity [7].

What would change the answer

A single benchmark matrix — hundreds of RBPs, one perturbed single-cell dataset, three prediction sets — would settle the paradigm comparison. Genome-scale Perturb-seq or Perturb-multiome RBP knockdowns with matched motif-based predictions, eCLIP footprint gene sets and co-expression-derived regulons, scored on the same cells, is the decisive design, and the pieces exist: 161 RBPs already interrogated by knockdown with scRBP predictions available [2], and long-read single-cell CRISPR perturbation with isoform-level readout demonstrated for 15 RBPs [20]. Lower-input footprint chemistry would test whether per-cell occupancy can replace motif priors rather than calibrate them [21]. Prospective validation through APA or 3'UTR-isoform readouts would discriminate region-specific regulon claims, since cell-state APA programs with shared 3'UTR motifs are already measurable from scRNA-seq [27], and quantitative in vivo binding-to-stability coupling assays could supply the affinity-weighted link that motif scores lack [8]. Absent such a benchmark, the defensible reading of the current literature is that motif-based, region-resolved inference is the operating standard, footprint data are its ground truth and its most transferable prior, and co-expression should be used to score regulon activity only after an external target definition has been fixed.

References

[1] Han, A., Lee, J.. "Updated binding model for a splicing factor, PTBP1 (Polypyrimidine Tract Binding Protein1)". bioRxiv. 2025-08-12T00:00:00.000Z. URL: https://doi.org/10.1101/2025.08.08.669377

[2] Ma, Y., Tang, M., Sriram, V., Li, J., Jops, C., England, J., Vo, D., Liao, Y., Smith, R. L., Margolis, M., Shalem, O., Ziller, M., Vuong, C. K., de la Torre-Ubieta, L., Gandal, M.. "scRBP infers cell-specific RNA-binding protein regulons from single-cell transcriptomes at gene and isoform resolution". bioRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.13.758934

[3] Gollin, J., Garcia-Gisbert, N., Isobe, T., Laurenti, E., Gottgens, B., Yang, A., Wilson, N. K.. "Single-Cell Proteo-Transcriptomic Profiling Uncovers Post-Translational Regulatory Layers in Human Hematopoiesis". bioRxiv. 2026-10-12T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.09.757909

[4] Svensson, V.. "A single-cell RNA-seq catalog of ground truth gene coregulation". bioRxiv. 2026-09-21T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.18.752692

[5] Liu, Y., Xiong, P., Li, S., Liu, X., Liu, T., Yang, Q., Wu, S., Peng, H., Li, Y., Zhang, L., Deng, Y., Zhu, Y., Wang, J., Deng, Y.. "Identification and quantification of alternative polyadenylation sites in single cell RNA-seq data using scPAISO". bioRxiv. 2025-08-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.08.20.669565

[6] Glass, M., Hüttelmaier, S.. "A gene set collection of RBP target genes". bioRxiv. 2025-10-15T00:00:00.000Z. URL: https://doi.org/10.1101/2025.10.14.682362

[7] Yi, S., Singh, S. S., Ye, X., Krishna, R., Kothwela, V., Jankowsky, E., Luna, J. M.. "Inherent Specificity and Variation Sensitivity as Quantitative Metrics for RBP Binding". bioRxiv. 2026-04-29T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.28.646018

[8] Kretov, D. A., Sanborn, O., McIsaac, T., Park, E., Imrat, I., Wu, S., Cifuentes, D.. "RBPscan: A Quantitative, In Vivo Tool for Profiling RNA-Binding Protein Interactions". bioRxiv. 2025-01-04T00:00:00.000Z. URL: https://doi.org/10.1101/2025.01.03.631239

[9] Hall, M. H., Wang, P. Y., Pham, T. M., Bartel, D. P.. "Functional microRNA targeting without seed pairing". bioRxiv. 2025-03-11T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.03.641270

[10] Weston, M., Ripan, R. C., Li, X., Hu, H.. "Novel features of miRNA and isomiR-mRNA interactions". bioRxiv. 2025-05-14T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.12.653582

[11] Mastriano, S., kanoria, S., Rennie, W., Liu, C., Li, D., Cheng, J., Ding, Y., Lu, J.. "High-Throughput Quantification of miRNA-3'-Untranslated-Region Regulatory Effects". bioRxiv. 2024-12-06T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.05.626985

[12] Badia-i-Mompel, P., Casals-Franch, R., Wessels, L., Müller-Dott, S., Trimbour, R., Yang, Y., Ramirez Flores, R. O., Saez-Rodriguez, J.. "Comparison and evaluation of methods to infer gene regulatory networks from multimodal single-cell data". bioRxiv. 2025-02-13T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.20.629764

[13] Suvojit Hazra, Marina Kotvanova, Kirstan Gimse, Sushmita Roy. "MERLIN-SUITE: Probabilistic modular GRN inference from multi-omics data integrating regulatory priors and transcription factor activity". arxiv. 2026-07-03T00:00:00.000Z. URL: https://arxiv.org/abs/2607.01791

[14] Van den Berge, K., Bickel, P. J., Dudoit, S.. "transfactor: Transcription factor activity estimation via probabilistic gene expression deconvolution". bioRxiv. 2025-03-19T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.19.644088

[15] Shan, X., Lin, Y., Zhao, H.. "A Unified Framework for Selecting and Evaluating Cell-Type-Specific Gene Co-expressions in Single-Cell Data". bioRxiv. 2025-12-08T00:00:00.000Z. URL: https://doi.org/10.1101/2024.11.08.622700

[16] Chu, C. P., Morin, A. A., Pavlidis, P.. "Limits to the inference of gene regulation from bulk tissue expression data". bioRxiv. 2025-09-14T00:00:00.000Z. URL: https://doi.org/10.1101/2024.10.24.619521

[17] Saran, S., Lebedeva, S., Hirsekorn, A., Ohler, U.. "Cell-type specific prediction of RNA stability from RNA-protein interactions". bioRxiv. 2024-11-19T00:00:00.000Z. URL: https://doi.org/10.1101/2024.11.19.624283

[18] Nariya, M. K., Santiago-Algarra, D., Tassy, O., Cerciat, M., Ye, T., Riba, A., Molina, N.. "Single-cell multiomics reveals the oscillatory dynamics of mRNA metabolism and chromatin accessibility during the cell cycle". bioRxiv. 2024-11-19T00:00:00.000Z. URL: https://doi.org/10.1101/2024.01.11.575159

[19] Moyon, L., Tirabassi, A., Baranowskii, A., Capitanchik, C., Kuret Hodnik, K., Wilkinson, L., Londhe, S., Dumbovic, G., Gagneur, J., Ule, J., Horlacher, M., Marsico, A.. "PARNET: A CLIP-SEQ-BASED FOUNDATION MODEL FOR RNA SEQUENCE REPRESENTATION LEARNING". bioRxiv. 2026-08-17T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.08.743506

[20] Müller, S., Andrews, N., Yan, R. E., Sookdeo, A., Burrell, W. H., Dai, X., Rughani, P., Daniloski, Z., Juul, S., Sanjana, N. E.. "Transcriptome-wide profiling of alternative splicing regulators with CRISPore-seq". bioRxiv. 2025-11-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.25.690515

[21] Sommerkamp, P., Sahadevan, S., Sekaran, T., Colucci, S., Ferring-Appel, D., Hentze, M. W.. "Identification of RNA Targets of Classical and Non-Canonical RNA-binding Proteins by soniCLIP". bioRxiv. 2026-08-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.17.745202

[22] Vivek, R., Kume, T., Roschdi, S., Record, T., Hoskins, A., BUTCHER, S.. "TDP-43 controls RNA structure through high affinity lattice interactions". bioRxiv. 2025-11-26T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.26.690829

[23] Gianluca Corrado, Michael Uhl, Rolf Backofen, Andrea Passerini, Fabrizio Costa. "ProtScan: Modeling and Prediction of RNA-Protein Interactions". arxiv. 2024-12-31T00:00:00.000Z. URL: https://arxiv.org/abs/2412.20933

[24] Bahrudeen, M., Scaravilli, M., Zahir, Z., Woodward, K., Ahola, T., Ratnadiwakara, M., Kaslin, J., Loughlin, F., Shirokikh, N. E., Anko, M.-L.. "dirCLIP profiles variant-specific RNA-protein interactions via nanopore long-read sequencing". bioRxiv. 2026-01-22T00:00:00.000Z. URL: https://doi.org/10.64898/2026.01.22.701028

[25] Nowak, I., Polash, A. H., Huynh, H. T., Kaur, M., Lobo, V., Scutenaire, J., Fong, M., Alluhaibi, G., Anastasakis, D. G., Hafner, M., Benhalevy, D., Sarshad, A.. "RBProximity-CLIP Enables Subcellular Mapping of RNA-Binding Protein Interactions at Nucleotide Resolution". bioRxiv. 2025-12-15T00:00:00.000Z. URL: https://doi.org/10.64898/2025.12.12.693770

[26] Baudin, A., Dinh, H. H., Breunig, K., Lei, X., Xu, X., Penalva, L. O., Libich, D. S.. "Intrinsically disordered SERBP1 regulates translation through topology-driven G-quadruplex recognition.". bioRxiv. 2026-03-05T00:00:00.000Z. URL: https://doi.org/10.64898/2026.03.03.709370

[27] Le, N. B., Sona, S., Santo, B., Schweickart, R. A., Kochat, V., Padron, W., Rai, K., Lee, S.-H., Yi, J. M., Wessely, O., Lee, B. H., Ting, A. H.. "Alternative polyadenylation regulates human urothelial differentiation". bioRxiv. 2025-03-11T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.03.641311

How this survey was made

This survey answers one question: “How do recent methods infer post-transcriptional regulons and RNA-binding protein targets from single-cell transcriptomes, and how do motif-based, footprint-based, and co-expression-based approaches compare?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 27 papers whose work this text draws on, chosen from the 165 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.