Partially connected neural networks for complex trait prediction: application to human height.

Polygenic risk scores (PRS) are fundamental tools for complex trait prediction, yet conventional methods often struggle to capture non-linear genetic interactions and ancestry-specific genetic architectures. Here, we propose a chromosome-aware, partially connected neural network (PCNN) that models localized genetic contributions and non-linear interactions across the genome. Through simulations with varying heritability levels and causal variant proportions, we demonstrate PCNN’s robust performance in capturing polygenic architectures. Applied to height prediction in 51,164 Han Chinese individuals using multi-ancestry genome-wide association studies summary statistics, PCNN achieves an R² = 0.2718 for males and 0.2603 for females, showing comparable or improved predictive accuracy relative to existing PRS methods such as Lassosum and PRScs. By structuring PRS inputs at the chromosome level, PCNN reduces input dimensionality and improves computational efficiency without compromising predictive performance. Our sensitivity analyses reveal the model’s ability to optimize non-linear relationships through LeakyReLU activation functions. This work establishes PCNN as an effective framework for polygenic trait prediction, particularly valuable for modeling population-specific genetic architectures and precision medicine applications.