Residual based anomaly detection framework for variant caller dispatch in DNA sequencing data

Reliable and scalable variant analysis is an enabling component of genomic studies involving microbial communities, host-associated microorganisms, and their hosts, and may support future investigations of genetic heterogeneity within symbiotic systems. Widely used workflows incorporating BWA and GATK provide standardized default processing routes, but their performance may vary across genomic regions containing repetitive sequences, complex structures, or atypical local sequence characteristics. Applying a context-aware software-recommendation procedure to every genomic region, however, can substantially increase computational demand. Here, we present LSTM-EWMA, a screening-and-dispatch framework designed to identify genomic regions that should be considered for specialized downstream evaluation. The framework represents ordered genomic regions as a sequence of feature vectors, uses a Long Short-Term Memory (LSTM) network trained exclusively on predefined in-control (IC) regions to model baseline patterns, and applies an Exponentially Weighted Moving Average (EWMA) control chart to standardized prediction residuals. Regions exceeding prespecified control limits are operationally labeled as out-of-control (OC) and designated as candidates for downstream software recommendation, whereas unflagged regions remain on the default processing path. These labels describe computational workflow states and do not independently confirm genomic variants or biological abnormalities. Using simulated sequencing data derived from the human reference genome as an initial methodological benchmark, LSTM-EWMA distinguished predefined OC regions from IC regions while maintaining a low observed false-alarm rate under the evaluated settings. These findings support the feasibility of the dispatch strategy within the current simulation design and provide a defined basis for subsequent evaluation in microbial, metagenomic, and host-associated sequencing contexts. With further validation across taxonomically diverse and biologically characterized datasets, LSTM-EWMA could support scalable variant-analysis workflows for microbial community and symbiosis research. The source code is publicly available at https://github.com/Icarus200110/Lstm-EWMA.