Lessons Learned from Predictive Modeling of Biogeographical Ancestry: Why Initial Data Analysis Matters?

In high-dimensional Single Nucleotide Polymorphism (SNP)-based prediction of BioGeographical Ancestry (BGA), performance variability is often attributed to algorithmic choices. However, evidence from recent supervised learning applications suggests that key limitations may originate from data properties. In particular, the integration of heterogeneous projects to overcome the unavailability of extensive data can induce differential missingness patterns and project effects that may influence downstream predictive performance even when extensive bioinformatics quality-control procedures are applied. This short contribution revisits these issues within the framework of Initial Data Analysis (IDA), defined as the set of inspection, screening, and documentation steps performed before the start of statistical analyses. Using a combined SNP dataset from multiple genome sequencing projects, we focus on two domains that proved particularly influential in predictive modeling: (i) missing data patterns; and (ii) project structure. Our results show project-specific missingness and between-project heterogeneity after bioinformatic quality controls, with potential implications for model training and generalization. We argue that systematic IDA provides a principled way to identify and document these characteristics, enhancing robustness, interpretability, and reproducibility in SNP-based BGA prediction.