Haplomatic: A Deep-Learning Tool for Adaptively Scaling Resolution in Genetic Mapping Studies

Genomic mapping studies face a fundamental trade-off between accuracy and resolution: increasing resolution improves localization of genetic signals but typically reduces the accuracy of frequency estimates due to increased statistical noise. In pooled sequencing studies, this trade-off impacts the accuracy of haplotype frequency estimates, a primary statistic used to identify genetic associations.To mitigate this trade-off we introduce Haplomatic, a novel deep-learning-based tool that adaptively adjusts genomic resolution by predicting haplotype frequency estimation error. Haplomatic generates simulated population data from known recombinant inbred line populations, predicts error through a transformer-based neural network, and adjusts resolution until a target error is achieved. Haplomatic achieves significant resolution gains (15% on average) over previous methods without sacrificing accuracy across multiple evaluated sequencing depths (10x, 50x, and 100x). To our knowledge, this is the first instance of applying deep learning to directly predict estimation error and dynamically scale resolution in trait mapping studies.