Rajkumar, N. | bioRxiv Science

bioRxiv2025

SiaRNA: A Siamese Neural Network with Bidirectional Cross-Attention for Pairwise siRNA-mRNA Efficacy Prediction

Small interfering RNA (siRNA) therapeutics have extraordinary potential for targeted gene silencing. They mediate post-transcriptional gene regulation by binding to complementary messenger RNA (mRNA) sequences and degrading them, thereby preventing the production of unwanted proteins. Recent machine learning and deep learning frameworks for predicting siRNA efficacy have only achieved moderate success as these models solely rely either on handcrafted features or on sequential relations and therefore cannot capture the full complexity of siRNA-mRNA interactions. In this context, we propose SiaRNA, which uses a Siamese Neural Network for feature-derived representations and a bidirectional cross-attention mechanism for sequence-level relationships. It uniquely identifies mRNAs and their corresponding siRNAs as paired entities, allowing unified and context-aware modeling. Unlike previous models, which discard 2-nucleotide (2-nt) overhangs at the 3 end while using 21-nt efficacy labels, SiaRNA both trains and tests on 21-nt sequences to ensure biologically consistent predictions. Our model sets a new performance benchmark, outperforming previous state-of-the-art models. SiaRNA is trained on the HUVK dataset achieving an accuracy of 0.881, while its generalization has been confirmed by testing on the independent Simone dataset. These results prove SiaRNAs potential as a reliable and biologically accurate framework to guide siRNA design and improve therapeutic outcomes.

bioinformatics↗

bioRxiv2025

Assessing genomic reproducibility of read alignment tools

Genomic research relies on accurate and reproducible computational analyses of DNA sequencing data to draw reliable biological conclusions. Read mapping, the process of aligning reads to a reference genome, is central to many applications, including variant detection and comparative genomics. While several tools have been developed for this task, genomic reproducibility1, defined as the consistency of results across replicates, remains underexplored. Here, we address this question by introducing a methodology based on synthetic replicates of sequencing data, generated by perturbing the original reads through shuffling, reverse complementing, or combined shuffling and reverse complementing. Our approach is able to simulate variability observed across sequencing runs due to differences in library preparation techniques. We evaluated the reproducibility of eight alignment tools (BWA-MEM22, Bowtie23, HISAT24, minimap25, NextGenMap6, SNAP7, SMALT7,8 and Subread9) under these perturbations using whole-genome sequencing (WGS) data. Synthetic replicates were aligned and compared to the original sample to quantify discrepancies. Mapping accuracy changes ranged from 0.0001% to 4.4% for primary reads, which are alignments not marked as secondary, supplementary, or duplicates, and up to 12.2% for high-quality primary reads. For primary reads, the percentage of reads commonly mapped in both the original and synthetic replicate ranged from 91.66% to 100% relative to the total number of mapped reads in the original dataset. High-quality filtering improved consistency, though some tools still failed to recover more than 70% of the original alignments. Within the set of common reads, the incidence of inconsistent mappings was as high as 13.53% for primary reads and 6.73% for high-quality primary reads. Bowtie 2 was fully reproducible under the shuffling replicate, Subread was fully reproducible under the reverse-complement replicate, whereas NextGenMap exhibited only minor inconsistencies. By contrast, SNAP and minimap2 displayed the most significant variability under reverse complementing. We further demonstrate that alignment inconsistencies propagate to the downstream task of calling structural genomic variants. Using Manta for structural-variant (SV) calling, we observed that Bowtie 2, HISAT2, and minimap2 maintained perfect SV concordance between original and replicate alignments, whereas other tools exhibited lower concordance, with Subread scoring only 87% concordance. In conclusion, our comprehensive evaluation demonstrates that synthetic perturbations reveal critical differences in how alignment tools handle technical variability and how these differences propagate to downstream variant analyses, underscoring the necessity of incorporating reproducibility benchmarks into the selection and validation of read mappers to ensure robust and reliable genomic interpretations.

bioinformatics↗