BeetleFormer: A Sequence Transformer for Identifying Effective Double-Stranded RNAs in the Red Flour Beetle, Tribolium castaneum

Selecting double-stranded ribonucleic acid (dsRNA) reagents remains a bottleneck in insect gene-function research and RNA interference (RNAi)-based pest management. We developed BeetleFormer to learn sequence patterns associated with RNAi screening outcomes in Tribolium castaneum. The model integrates multiscale convolutions, long-range attention, complementary-strand symmetry and sequence descriptors. Following previous dsRNA prescreening research, 14,821 iBeetle reagents meeting mortality or phenotype criteria were used as positives and 14,825 messenger RNA (mRNA) fragments as negative samples. Source genes and overlapping sequences were separated across training, validation and testing. On 6083 test sequences, BeetleFormer achieved an area under the receiver operating characteristic curve (AUC) of 0.978, accuracy 0.913 and Matthews correlation coefficient 0.831, compared with 0.888, 0.817 and 0.636 for an adapted convolutional–recurrent comparator. Ablations supported local processing and strand symmetry; AUC remained 0.976 after exact-length matching. External testing on 20 published dsRNAs and 20 mRNA negatives yielded AUC 0.928. Model interpretation revealed terminally concentrated sequence sensitivity and organized learned representations. These findings establish an improved sequence-based prescreen that turns genome-scale RNAi observations into candidate-selection evidence, supporting functional-genomics experiments and the development of RNAi-based pest control.