---
title: 'SiCmiR Atlas: Single-Cell miRNA Inference'
url: https://www.emergentmind.com/topics/sicmir-atlas
type: topic
---

# SiCmiR Atlas: Single-Cell miRNA Inference

Searching arXiv for the specified papers and closely related work on single-cell miRNA inference and atlas resources.
SiCmiR Atlas is a human single-cell resource for mature microRNA abundance inferred directly from single-cell RNA sequencing data by SiCmiR, a lightweight two-layer neural network that maps the expression of 977 LINCS L1000 landmark genes to a full mature-miRNA profile. Its stated purpose is to overcome sparsity and dropout in single-cell transcriptomes and the technical barriers of single-cell small-RNA profiling while supporting interactive visualization, differential biomarker discovery, and cell-type-resolved miRNA-target interaction networks in cancer and related disease settings [2508.05692].

## 1. Definition, scope, and resource identity

SiCmiR Atlas is presented as a dedicated database of single-cell mature miRNA expression derived from scRNA-seq rather than from direct small-RNA sequencing. The resource is human-only and integrates scRNA-seq counts or CPMs with model-predicted mature miRNA abundance per cell. Its coverage is described in detail as 362 publicly available datasets spanning 189 tissues from 26 major organs, 84 physiological or disease conditions grouped into 12 categories, and 726 annotated cell types, totaling 9,358,863 cells. The Atlas is freely accessible at `https://awi.cuhk.edu.cn/~SiCmiR/` [2508.05692].

| Component | Specification |
|---|---|
| Species | Human |
| Modalities | scRNA-seq counts/CPMs and predicted mature miRNA abundance per cell |
| Coverage | 362 datasets, 9,358,863 cells, 726 cell types, 189 tissues, 84 conditions |
| Core functions | Interactive visualization, biomarker discovery, cell-type-resolved MTI networks |

The paper characterizes the resource as the first dedicated database of single-cell mature miRNA expression. A notable textual discrepancy exists within the paper itself: the abstract states that the Atlas contains 632 public datasets, whereas the detailed resource description specifies 362 datasets. Both figures are reported in the same source, and the detailed breakdown is the one attached to the fuller description of tissues, conditions, and cell types [2508.05692].

The underlying rationale is methodological as well as biological. Mature miRNAs are post-transcriptional regulators central to cancer biology, but direct single-cell miRNA profiling remains technically challenging because of polyadenylation dependence, adaptor dimers, data sparsity, difficulty distinguishing small RNA species, and protocol variability. SiCmiR Atlas addresses this by inferring miRNA abundance from mRNA-based single-cell data, thereby transferring bulk-derived statistical power to single-cell resolution [2508.05692].

## 2. Predictive foundation: the SiCmiR model

The Atlas is built on SiCmiR, a multi-output regression deep neural network trained on paired bulk miRNA-mRNA profiles from TCGA pan-cancer data. The model uses expression from the 977 LINCS L1000 landmark genes as input, with XBP1 excluded because of zero counts in TCGA, and predicts continuous abundance for 1,298 mature miRNAs in RPM-like units. The network architecture is explicitly described as two fully connected layers with dimensions \(977 \rightarrow 1024 \rightarrow 1298\), with batch normalization, ReLU activation, and dropout at rate \(0.3\) in the hidden layer. Training uses stochastic gradient descent with learning rate \(0.4\), early stopping after 20 epochs without validation-loss improvement, and grid-searched hyperparameters [2508.05692].

The training corpus comprises 6,462 paired miRNA-mRNA bulk profiles from 33 TCGA cancer types, with 2,768 independent samples reserved for testing. mRNA inputs were obtained from UCSC Xena and mapped to the 977 landmark genes; miRNA-seq reads were normalized to RPM. For single-cell inference, counts are library-size normalized as CPM \(\times 10^4\), log-transformed, and in some case studies z-scored across cells to stabilize scale. No explicit batch correction is applied, although stratified 3-fold cross-validation is reported as a robustness check [2508.05692].

The optimization target is mean squared error without auxiliary losses or explicit \(L_2\) regularization:

$$
\min_{\Theta} \sum_i \lVert f(x_i; \Theta) - y_i \rVert_2^2.
$$

The choice of the 977 landmark genes is central to the design. The paper argues that these genes are reproducible, broadly responsive to perturbation, and sufficient to infer the wider transcriptome while reducing sensitivity to zero inflation at single-cell resolution. A plausible implication is that the model treats feature compactness not merely as computational compression, but as a robustness mechanism for sparse scRNA-seq inputs [2508.05692].

The framework also exposes two inference regimes. Direct single-cell prediction is supported, but the method explicitly offers pooled “pseudo-bulk” modes, including cell-type averages and bootstrap pooling, to increase signal and mitigate dropout. When some landmarks are missing in a scRNA-seq dataset, inference still proceeds, although the paper states that performance benefits from careful gene mapping, standard normalization, and pooling [2508.05692].

## 3. Atlas organization, interface, and analytical workflows

The Atlas couples predicted miRNA abundance with harmonized metadata. Cell-type labels are retained from original studies, tissue and organ assignments are mapped using Cell Ontology, and condition or disease annotations are grouped into consistent categories. The web application is implemented with Apache Wicket on a CentOS HPC environment and supports browsing by dataset source, tissue, condition, and cell type [2508.05692].

The interface exposes several linked analytical layers. Interactive UMAP or t-SNE views display cluster distributions within a dataset. An expression explorer allows browsing of mRNA and miRNA expression across clusters and provides quick visualization of cell-type enrichment. Differential analysis modules use Wilcoxon rank-sum tests for single-cell contrasts and DESeq2-like pseudo-bulk contrasts for bulk-style comparisons, with Benjamini-Hochberg adjustment available for FDR control. The network builder constructs cell-type-specific miRNA-target interaction networks by integrating TargetScan, miRWalk, miRDB, and miRTarBase with SiCmiR-predicted miRNA abundance and observed mRNA expression or correlations [2508.05692].

The paper gives a concrete usage pattern. A researcher may query a cancer such as glioblastoma, inspect the UMAP to isolate malignant and stromal clusters, visualize a candidate miRNA such as hsa-miR-21-5p, run differential analysis to obtain \(\log_2\) fold-change, \(p\)-values, and \(q\)-values, then build a malignant-cell MTI network and prioritize hubs by degree, betweenness, or PageRank centrality. Outputs can be exported for downstream validation or integration with GEPIA2 or OncoMiR [2508.05692].

This workflow positions the Atlas as more than a visualization layer. It functions as a structured inference-and-prioritization system in which predicted mature miRNA abundance, differential enrichment, and network context are analyzed together. The design is particularly oriented toward cell-type-resolved biomarker discovery in heterogeneous tumor microenvironments [2508.05692].

## 4. Performance, benchmarking, and generalization

SiCmiR is evaluated against alternative neural architectures and against miRSCAPE. In the architecture comparison, the two-layer neural network outperformed ResNet and Transformer baselines on the test set, with Pearson correlation coefficient \(0.673\) versus \(0.578\) and \(0.359\), and mean squared error \(0.522\) versus \(0.648\) and \(0.863\). Feature-selection experiments further showed that large feature sets produced training PCC values around \(0.79\) to \(0.82\), but test PCC values clustered around \(0.66\) to \(0.68\); the L1000 landmark set achieved the stated best trade-off at roughly 1,000 features, with training PCC \(0.75\) and test PCC \(0.67\) while reducing computational cost [2508.05692].

Cross-validation results are reported both miRNA-wise and sample-wise. Three-fold cross-validation yielded miRNA-average PCCs of \(0.75 \pm 0.00067\) on training and \(0.67 \pm 0.00073\) on test folds, and sample-average PCCs of \(0.72 \pm 0.09583\) on training and \(0.63 \pm 0.13707\) on test folds. Against miRSCAPE, SiCmiR achieved higher independent-test PCC, \(0.67\) versus \(0.61\), and much faster inference, with 2.23 seconds for the pre-trained pan-cancer model versus more than 2 hours for miRSCAPE because of per-dataset training. The pan-cancer model also generalized better than cancer-specific models across the 33 TCGA cancer types and was selected as the default model [2508.05692].

The case studies are used to substantiate cross-domain generalization. In pancreatic ductal adenocarcinoma scRNA-seq comprising 57,530 cells, SiCmiR predicted 66 of 90 dysregulated miRNAs from pooled DC2 versus DC1/acinar contrasts, corresponding to sensitivity \(0.73\), and 28 by direct single-cell prediction, corresponding to \(0.31\). For the 39 miRNAs covered by miRSCAPE, SiCmiR correctly predicted 29 of 37, or \(0.78\), versus 29 of 39, or \(0.74\), for miRSCAPE. Reported examples include higher hsa-miR-30b-3p in DC1 than DC2 and elevated hsa-miR-21-5p in DC2 and MUC5\(^+\) DC1 versus MUC5\(^-\) DC1, consistent with PDAC progression [2508.05692].

In ACTH-secreting pituitary adenoma, a cancer type absent from TCGA training, the model recovered 55 of 75 reported dysregulated miRNAs at \(|\log_2 FC| \ge 0.25\) in pooled contrasts, giving sensitivity \(0.73\); at \(|\log_2 FC| \ge 1\), sensitivity was \(0.53\). Direct single-cell predictions identified 46 miRNAs, of which 34 matched the expected direction, corresponding to \(0.69\). In hepatocellular carcinoma bulk data from the AA cohort, 19 of 24 DEmiRs were recovered with sensitivity \(0.79\), and among 13 DEmiRs with test PCC at least \(0.8\), 12 were correctly predicted, or \(0.92\). In an A549 drug-perturbation setting, seven DEmiRs were inferred, five with test PCC at least \(0.8\), and the decreases of hsa-miR-27a-3p and hsa-miR-24-3p with increasing *Cinnamomi Ramulus* were stated to be validated by qPCR [2508.05692].

The glioblastoma extracellular-vesicle case study extends the model from abundance prediction to network inference. Incorporating SiCmiR mature-miRNA predictions via miRTalk increased high-confidence MTIs from 5 to 114,501 edges and increased the fraction showing expected negative miRNA-mRNA correlation from 15.6% to 36.9%, with Fisher’s odds ratio \(3.17\), 95% confidence interval \(2.94\) to \(3.42\), and \(P < 2 \times 10^{-16}\). Average interaction score rose from \(0.00095\) to \(0.04494\), and the repression effect size modestly strengthened, with Cliff’s \(\delta = -0.21\) and one-sided Wilcoxon \(P \approx 0.011\) [2508.05692].

## 5. Hub-miRNAs, SHAP modules, and cell-type-resolved network signatures

A distinctive feature of SiCmiR Atlas is the use of predictive reproducibility as a criterion for identifying candidate hub-miRNAs. The paper defines hub miRNAs as those achieving PCC at least \(0.80\) on the independent test set, which yields 414 hub miRNAs, including 105 mature pairs from common precursors. These hubs are reported to form denser cancer-associated networks than non-hubs, with mean degree \(11.12\) versus \(4.63\), and their targets are enriched in hallmark processes including angiogenesis, extracellular matrix remodeling, epithelial-mesenchymal transition, proliferation, and immune modulation [2508.05692].

Model interpretation is performed with SHAP. The reported SHAP analysis identifies 12 functional modules linking landmark-gene contributions to inferred miRNA behavior. One prominent module is driven by COL1A1, CDC25A, and GLI2 and contains 32 miRNAs clustered on chromosome 14, including hsa-miR-127-3p/5p, hsa-miR-134-5p, and hsa-miR-136-3p/5p. The targets of this module are enriched for metastasis-related processes, and high expression of the associated miRNA set predicts poorer survival in renal cancers, specifically KIRC and KIRP, while COL1A1, TGFB3, CDC25A, TNIP1, STAT5B, and TRAK2 are listed among the top prognostic contributors [2508.05692].

The network layer in the Atlas integrates target predictions from TargetScan, miRWalk, and miRDB with experimentally validated interactions from miRTarBase. These are overlaid with predicted miRNA abundance and observed mRNA expression or correlations, and edges are defined by combined target-site confidence, expression support, and directionality consistency. Standard centrality measures, including degree, betweenness, and PageRank, are then used to prioritize hub-miRNAs or hub genes within cell-type-specific MTI networks [2508.05692].

The paper presents these network signatures as sets of cell-type-enriched miRNAs and targets supported by recovered literature associations, direction-consistent negative correlations, and survival links. This suggests that the Atlas is intended not only to recover isolated biomarkers but also to map higher-order regulatory programs that differ across tumor, stromal, and immune compartments [2508.05692].

## 6. Distinction from earlier resources, interpretation, and limitations

SiCmiR Atlas should be distinguished from earlier miRNA-related exploratory tools in cancer genomics. In particular, the 2018 paper “Single nucleotide polymorphisms that modulate microRNA regulation of gene expression in tumors” introduced an open-source interactive R Shiny application named mirApp, together with complete analysis code, for querying significant SNP-miRNA-gene interaction results, termed regQTLs, across four TCGA tumor types: breast, liver, lung squamous cell carcinoma, and prostate [1803.03189]. That resource is pathway-guided, genotype-centric, and based on matched SNP genotype, miRNA-seq, and mRNA RNA-seq in bulk tumors; SiCmiR Atlas is single-cell-oriented, expression-predictive, and organized around inferred mature miRNA abundance and MTI networks rather than genotype-dependent interaction terms. Conflating the two obscures a substantial change in modality, scale, and inferential goal.

Several limitations of SiCmiR Atlas are explicit. The model is bulk-trained on TCGA pan-cancer profiles, so domain shifts to non-cancer tissues or rare niches may require cautious interpretation. No explicit batch correction is used in training, and heterogeneous cross-dataset integration may therefore remain sensitive to preprocessing differences. Predictions depend on the measurement quality of the 977 L1000 landmarks, and single-cell direct inference is acknowledged to have lower sensitivity than pooled or pseudo-bulk prediction under substantial dropout. MTI networks depend on incomplete target databases and on correlational support, so they are not definitive mechanistic maps [2508.05692].

The biomarker-discovery framework also has interpretive boundaries. Single-cell differential testing uses Wilcoxon rank-sum tests and pseudo-bulk contrasts use DESeq2 with Wald tests; Benjamini-Hochberg adjustment is available, and AUC may be used to assess classification performance. However, the paper’s primary validation emphasis is not prospective clinical testing but recovery of literature-supported dysregulation patterns, direction-consistent correlations, and survival links. A plausible implication is that the Atlas is best viewed as a hypothesis-generation and prioritization platform rather than as a substitute for direct small-RNA measurement or perturbational validation [2508.05692].

Within those limits, SiCmiR Atlas defines a distinct computational paradigm for miRNA analysis at single-cell scale: mature-miRNA abundance is inferred from a compact landmark-gene representation, organized into a browsable human atlas, and connected to biomarker, hub, and network analyses. Its significance lies less in replacing direct profiling than in furnishing a scalable intermediate layer between conventional scRNA-seq data and miRNA-centered hypotheses in heterogeneous human cancers [2508.05692].

Source: https://www.emergentmind.com/topics/sicmir-atlas