This document serves as an erratum for the research paper "MicroRNA target prediction using thermodynamic and sequence curves," originally published in BMC Genomics. The primary purpose of this correction is to rectify errors in Figures 6 and 7 and their respective captions. These figures are central to understanding the performance and biological insights provided by Avishkar, a computational tool designed for high-accuracy microRNA (miRNA) target prediction.
Abstract
The original publication contained errors in the presentation of key data visualizations. This erratum provides the corrected versions of Figure 6 (seed alignment pattern frequencies) and Figure 7 (ROC performance curves). These corrections underscore the prevalence of non-canonical miRNA-mRNA interactions and validate the predictive robustness of the Avishkar framework across different genomic regions.
The Challenge of Non-Canonical Binding
MicroRNAs regulate gene expression by binding to messenger RNA (mRNA). While many predictive models rely on "canonical seed matches"—perfectly complementary sequences—biological reality is more complex. Functional interactions frequently involve mismatches, gaps, or GU wobble base pairs. Avishkar addresses this by moving beyond simple sequence matching, utilizing thermodynamic stability and sequence-specific structural curves to capture these non-canonical interactions.
Evolutionary Conservation of Binding Patterns
One of the most significant findings clarified by the corrected data is the prevalence of non-canonical binding. By categorizing seed alignment patterns into matches, mismatches, gaps, and GU wobbles, the authors compared frequencies across species.
As shown in Figure 6, there is a remarkably high correlation ($\rho = 0.923$) between the binding patterns found in humans and mice. Notably, among the top 10 most frequent patterns, only two were canonical 6-mer or 7-mer matches. The remaining eight involved various forms of imperfect matching, suggesting that these non-canonical patterns are evolutionarily conserved and biologically functional rather than stochastic noise.
Predictive Performance Across Gene Regions
The Avishkar tool was evaluated using 10-fold cross-validation on large-scale human and mouse datasets. While traditional research often focuses exclusively on the 3’ Untranslated Region (3'UTR), this study demonstrates that miRNA regulation is active across the entire gene body.
The corrected ROC curves in Figure 7 illustrate consistent predictive accuracy across different regions:
- 3'UTR: AUC $\approx$ 0.75
- Coding Sequence (CDS): AUC $\approx$ 0.73
- 5'UTR: AUC $\approx$ 0.71
These results confirm that Avishkar maintains significant predictive capability (compared to a random baseline of 0.5) regardless of the target region, supporting the hypothesis that miRNA-mediated regulation is more widespread than previously assumed.
Conclusion
The corrections provided in this erratum reinforce the core thesis of the original work: thermodynamic modeling and the inclusion of non-canonical binding patterns are essential for accurate miRNA target prediction. The high cross-species correlation and regional consistency of the Avishkar tool highlight its utility as a robust framework for genomic research.
Summary updated