As a researcher I have a few questions and concern about this study:
1. While the spectral similarity and statistical congruence (e.g., mean, IQR) between real and simulated data are convincingly shown, no evidence is provided to demonstrate that the simulated GPC values maintain the multigenic and environmentally influenced architecture of the original trait. GPC is well known to be a complex, quantitative trait with significant G×E interaction. Given that the DCGAN was trained on a one-year dataset, and phenotypic variance components were not decomposed (e.g., through REML or mixed models), how can we be certain that the generated values are not biased toward simplified or over-smoothed representations of the trait space?
2. The key claim that simulated data improve GWAS power is potentially confounded by the fact that the DCGAN learns from and amplifies existing correlations in a dataset with limited sample diversity (n = 276). Notably, the increase in detected SNPs in the augmented GWAS (estimated value 2) could reflect learned spectral–genomic artifacts rather than biologically meaningful associations. The detection of OsmtSSB1L as a candidate gene is intriguing, but the absence of its signal in the measured-GPC GWAS suggests either insufficient replication or potential overfitting by the model. The authors should be encouraged to perform permutation testing or false discovery rate correction across traits to distinguish real signals from DCGAN-induced noise.
3. Despite conducting field trials in two consecutive years, the study ultimately relies on GPC data from single environments for GWAS, ignoring substantial environmental variance and genotype-by-environment interaction—both of which are well documented in rice GPC studies. The absence of BLUE or BLUP estimations substantially limits the ability to derive genetically stable phenotypes, which directly impacts the accuracy and repeatability of detected QTLs. This omission is particularly critical given the stated goal of improving breeding selection.
4. The wavelet features (e.g., WF1743,2) identified as highly predictive of GPC are localized around the protein-sensitive NIR/SWIR bands (~1510 and 1690 nm), yet no biochemical or spectral deconvolution validation is provided to verify the specificity of these signals to protein absorption rather than correlated biochemical constituents (e.g., starch, moisture). Given the known overlapping absorptance in these regions, particularly with amide and O–H bands, could the model be inadvertently capturing proxy traits, thereby compromising its specificity?
5. The finding that estimation accuracy peaks at 200 simulated samples before deteriorating raises the question of how to determine optimal augmentation volume a priori. Was any regularization mechanism applied to constrain the variance of simulated samples, or were all epochs treated equally?
6. The study focuses exclusively on japonica varieties from the Taihu Basin and Nanjing breeding programs, which raises concerns about the transferability of the model to broader rice germplasm pools. How would this model perform in indica backgrounds or tropical japonicas?