Binary logistic regression support early-stage selection in tropical potato breeding for industrial processing (P-02)

The identification of superior potato genotypes relies heavily on accurate empirical selection, especially during early stages of breeding when thousands of progenies are evaluated. At this stage, subjective, ideotype-oriented breeder preferences play a key role. Statistical modeling and machine learning (ML) techniques such as binary logistic regression (BLR) are powerful tools to support biological decision-making and reduce human bias. To address early-selection subjectivity we employed BLR on data of two field years (FY1 and FY2), pinpointing the most impactful features for classification. A total of 179 breeding lines were evaluated across FY1 and FY2 for 17 traits. Separate models were fitted for each year using the statsmodels library. Sensitivity, specificity, AUC-ROC, Global Model Efficiency (GME), and Youden’s Index (J) were extracted. Five-fold cross-validation was conducted using the scikit-learn library. Four significant predictors for FY1 (total yield, marketable tuber percentage, tuber eye depth, and skin color) and two for FY2 (marketable tuber percentage and shape) were retained. Overall, the FY1 model exhibited superior predictive power. High specificity (0.97) was achieved for FY1 at a cutoff of 0.5, while FY2 reached its highest specificity (0.75) at a cutoff of 0.2. The AUC-ROC was 0.87 for FY1 and 0.75 for FY2. The highest Youden’s Index (J) for FY1 was 0.59 at a cutoff of 0.1, while FY2 peaked at 0.35 with a cutoff of 0.2. Confusion matrices indicated better model performance at lower cutoffs (0.1–0.2), with fewer misclassifications. Cross-validation showed higher precision (0.6) for FY1 at a cutoff of 0.8, whereas FY2 showed null precision at this same cutoff. For breeding purposes, precision should be prioritized, lowering advancement of poor-performing clones. In our dataset, cutoff definition was a critical factor and must be addressed carefully. These findings highlight the potential of BLR to enhance decision-making in breeding programs.