CASREL: de novo RNA-binding protein circuitry from single-cell splicing without binding priors

Why binding priors and bulk assays left a gap in splicing circuitry

Earlier work established that sequence-specific RBPs bind pre-mRNA and control alternative splicing in a highly context-dependent, combinatorial manner, with each cell type expressing a distinct RBP repertoire that reinterprets the same regulatory elements differently [8]. Reviews of the field catalogued how SRSF and hnRNP families exert opposing effects on exon inclusion, and how RBP–ncRNA crosstalk feeds into tumor progression, drug resistance, and metastasis [2][6]. Computational prediction of RBP binding sites has depended on experimentally defined motifs, as in RBPmap, which maps binding sites only when a motif is already available [3]. CLIP-seq and RBP perturbation experiments remain the reference standard for physical and functional evidence, but they are low-throughput, often performed in a few model cell lines, and bulk-resolved, so they miss cellular heterogeneity [1]. Single-cell studies of alternative splicing heterogeneity during early mouse development confirmed that AS events vary substantially from cell to cell, reinforcing the case for single-cell resolution [4].

The practical consequence is that most existing computational approaches are restricted to a few preselected RBPs in limited cell lines, and bulk RNA-seq dilutes signals from both RBP expression and AS events, masking cell-type-dependent regulatory machinery [1]. CASREL is positioned against this baseline: it does not require CLIP annotations, predefined motifs, or prior knowledge of RBP function, and it operates on naturally occurring splicing variation across individual cells in complex tissues [1].

What CASREL actually does: ensemble learning plus SHAP attribution

CASREL takes single-cell RBP expression profiles and AS profiles as input and runs a three-step pipeline: preprocessing, gradient-boosted decision tree (GBDT) modeling of AS events from RBP expression, and SHAP-based model interpretation to infer RBP–AS regulation [1]. The authors benchmarked logistic regression, a fully connected neural network, and GBDT on two TNBC datasets and one HCC tumor immune cell dataset; GBDT achieved the highest weighted-F1 across all three (Wilcoxon rank-sum, all P < 0.05), with median weighted-F1 consistently above 0.8, and also led on balanced accuracy and macro-F1 [1]. Because single-cell splicing data are sparse, dropout-prone, and imbalanced, balanced accuracy and macro-F1 are expected to be substantially lower than overall accuracy, and the authors note that recent benchmarking studies of sparse scRNA-seq classification report similarly moderate values [1].

The interpretability comparison is the more consequential result. Logistic regression coefficients and FCN SHAP values assigned the same RBP both promoting and inhibiting effects on the same AS event—contradictory, non-monotonic attributions—whereas GBDT produced long-tailed SHAP distributions and mutually exclusive RBPs with monotonic effects [1]. For CD44s (exon 3–14 skipping in CD44), GBDT recovered ESRP1 as a repressor and HNRNPL as an activator, both previously established regulators of CD44s in breast cancer and EMT; for FAS exon 6 skipping, GBDT recovered RBM5, TRA2B, and HNRNPC, known FAS regulators, while LR and FCN did not [1]. The authors explicitly caution that SHAP values are model-attribution scores reflecting each feature's contribution to the predictive output, not direct measurements of biochemical regulatory strength [1].

How the predictions hold up against ENCODE knockdowns and CLIP-seq

The strongest external test uses ENCODE RNA-seq from 262 RBP knockdown experiments in K562 and HepG2, each with two independent shRNAs and two biological replicates, yielding four knockdown samples per RBP per cell line [1]. Differential splicing was called with Wilcoxon rank-sum tests and Benjamini–Hochberg correction (FDR < 0.05, |ΔPSI| > 0.1), and an RBP–AS pair was counted as validated if significant differential splicing appeared in at least one of the two cell lines, a deliberately permissive criterion because splicing magnitude is often cell-type-dependent while the underlying regulatory relationship may be conserved [1]. Permutation-based null distributions (1000 iterations per RBP) were used to confirm that validation rates exceeded chance, with enrichment fold change computed against the null mean and one-sided empirical P-values [1]. Orthogonal CLIP-seq support came from POSTAR3, covering binding site annotations for 267 RBPs (221 human, 46 mouse), with the supporting rate defined as the fraction of CASREL-predicted target genes containing at least one validated binding peak for that RBP [1].

A specificity analysis selected PCBP1, GRSF1, and HNRNPK and compared splicing changes upon knockdown of the target RBP versus 100 randomly selected RBPs as negative controls [1]. In the FUS-knockout neuron experiment, models trained on unperturbed control neurons predicted differential splicing in independent FUS-knockout neurons with a validation rate of ~0.8, versus 0.5–0.6 when FUS targets were inferred from other cellular contexts, suggesting CASREL captures cell-type-specific regulatory information beyond the conserved component shared across tissues [1]. The authors also benchmarked against SFpointer (CLIP-seq-based), SAM-AM (sequence conservation-based), and DeepRBP (population-scale machine learning) using standardized datasets and pipelines from the respective publications [1].

Cell type-specific circuits, conserved regulators, and batch behavior

In tumor-infiltrating T cells from HCC and CRC, CASREL identified PCBP1 as a strong promoter of STAT3β formation in both CD4+ and CD8+ T cells, consistent with prior evidence that PCBP1 binds STAT3 exon 23 to enhance STAT3β generation in multiple cell lines; GRSF1 and TRA2A acted as inhibitors, consistent with RBP-knockdown data from HepG2/K562 [1]. Cell type-specific patterns also emerged: PCBP2, which shares RNA-binding preferences with PCBP1, inhibited STAT3β formation selectively in CD8+ T cells [1]. Regulators of STAT3β in CD4+ and CD8+ T cells were highly consistent between HCC and CRC, which the authors interpret as supporting cell type dependency of RBP–AS regulation with minimal batch bias [1]. For CLK1 exon 4 skipping—which generates the inactive CLK1T1 isoform—CASREL identified conserved positive regulators including POLR2G, HNRNPA1, and SNRPC across CD4+ and CD8+ T cells, plus distinct cell type-specific RBPs [1].

These examples illustrate the intended use case: recovering known regulators as a sanity check while surfacing candidate cell type-specific circuitry that would be difficult to reach with bulk assays. The cross-tissue consistency between HCC and CRC is a computational robustness observation, not independent experimental replication, and the authors frame the output as putative cell-specific regulatory circuitry [1].

Where the inference stops: candidates, not causal validation

The central boundary is explicit in the paper: SHAP values are model-attribution scores, not biochemical regulatory strength measurements, and the top-ranked associations constitute candidate cell type-specific splicing regulatory circuitry [1]. Validation against ENCODE knockdowns and POSTAR3 CLIP-seq is orthogonal and quantitative, but it is still computational comparison against pre-existing datasets, not prospective experimental testing of CASREL's novel predictions [1]. The permissive one-of-two-cell-lines validation criterion and the coverage restriction to AS events quantifiable in ENCODE both widen the effective validation window [1]. The broader literature on RBP function reinforces why this matters: RBP–RNA interactions are combinatorial and context-dependent, with RBPs controlling each other's binding activities, so a ranked association list cannot by itself establish directionality or mechanism [8]. Reviews of RBP–ncRNA crosstalk in tumors document extensive regulatory layers—m6A modification, alternative polyadenylation, phase separation—that operate alongside splicing and are not modeled by CASREL [2].

The comparison class also has inherent limits. Motif-based tools such as RBPmap require an experimentally defined binding motif and therefore cannot operate de novo [3]. Experimental approaches that do establish causality, such as CRISPR perturbation screens, carry their own unknown false-negative rates and off-target concerns, as documented in a CAD variant screen where six of eight prioritized sgRNAs validated but expected loci were not recovered [7]. Small-molecule efforts against RBPs such as Lin28 illustrate the difficulty of therapeutically targeting these interactions and the value of structural priors that CASREL does not use [5]. The natural next step for CASREL predictions is prospective perturbation of the top-ranked cell type-specific RBP–AS pairs in the cell types where they were inferred, which the paper does not report [1].

About These Sources

This research page is built on 8 peer-reviewed studies — published from 2014 to 2026, 5 from 2024 or later, collectively cited 1,335 times — selected as the most relevant from 13 studies that passed quality screening, drawn from 88 papers retrieved from a database of over 500 million.

Sources used in this answer