Transfer learning with false negative control improves polygenic risk prediction

Polygenic risk score (PRS) is a quantity that aggregates the effects of variants across the genome and estimates an individual's genetic predisposition for a given trait. PRS analysis typically contains two input data sets: base data for effect size estimation and target data for individual-level prediction. Given the availability of large-scale base data, it becomes more common that the ancestral background of base and target data do not perfectly match. In this paper, we treat the GWAS summary information obtained in the base data as knowledge learned from a pre-trained model, and adopt a transfer learning framework to effectively leverage the knowledge learned from the base data that may or may not have similar ancestral background as the target samples to build prediction models for target individuals. Our proposed transfer learning framework consists of two main steps: (1) conducting false negative control (FNC) marginal screening to extract useful knowledge from the base data; and (2) performing joint model training to integrate the knowledge extracted from base data with the target training data for accurate trans-data prediction. This new approach can significantly enhance the computational and statistical efficiency of joint-model training, alleviate over-fitting, and facilitate more accurate trans-data prediction when heterogeneity level between target and base data sets is small or high. Polygenic risk score (PRS) can quantify the genetic predisposition for a trait. PRS construction typically contains two input datasets: base data for variant-effect estimation and target data for individual-level prediction. Given the availability of large-scale base data, it becomes common that the ancestral background of base and target data do not perfectly match. In this paper, we introduce a PRS method under a transfer learning framework to effectively leverage the knowledge learned from the base data that may or may not have similar background as the target samples to build prediction models for target individuals. Our method first utilizes a unique false-negative control strategy to extract useful information from base data while ensuring to retain a high proportion of true signals; it then applies the extracted information to re-train PRS models in a statistically and computationally efficient fashion. We use numerical studies based on simulated and real data to show that the proposed method can increase the accuracy and robustness of polygenic prediction across different ranges of heterogeneities between base and target data and sample sizes, reduce computational cost in model re-training, and result in more parsimonious models that can facilitate PRS interpretation and/or exploration of complex, non-additive PRS models.