BGPT: Paper Review: Deep learning empowers genomic selection of pest-resistant grapevine

Explore by Goal

Quick Explanation Copied

What the paper claims (and data it reports): A DL phenotyping pipeline (VGG16 + a continuous-score model “DCNN-PDS”) converts grape-leaf images of Spodoptera litura damage into (i) binary “mild vs severe” labels and (ii) continuous 0–5 scores, then feeds both into GWAS and finally into ML-based genomic selection (GS). Reported performance includes 95.3% binary accuracy (VGG16 test set) and r=0.94 (R²=0.88) for continuous regression, while GS reports 95.7% binary accuracy and corr=0.90 for continuous traits.

Long Explanation

Paper Review (critical & evidence-based)

Deep learning empowers genomic selection of pest-resistant grapevine — Horticulture Research — DOI: 10.1093/hr/uhaf128

Key inputs

231 grape accessions; leaf-image phenotyping under greenhouse infestation; genome resequencing used for GWAS and GS; transcriptomics used for DEG intersection

Core outputs

DL phenotypes → GWAS QTLs (binary + continuous) → candidate genes; DL phenotypes + SNP sets → ML-based GS predictions

Figure 1 — Reported DL performance anchors (binary + continuous)

These are the headline metrics explicitly stated in the provided text.

Interpretation caution: mixed metric types (accuracy vs correlation) are visualized together by scaling r/corr/R² to percent for readability; the underlying reported values remain as stated above.

Figure 2 — Reported dataset sizes for DL steps

The paper states 1448 images used for CV (binary) and 362 for test; and 2181 for training/CV (regression) and 439 for test.

Figure 3 — GWAS summary counts (binary vs continuous)

The paper reports 33 significant QTLs for binary and 36 for continuous traits, with 67 and 85 candidate genes respectively, totaling 139 nonredundant candidate genes.

Figure 4 — Claimed signaling contexts for candidate genes

This figure is a concept map of which pathway families the paper says its candidate genes are involved in.

The paper states its candidate genes participate in pest resistance pathways including jasmonic acid (JA), salicylic acid (SA), ethylene, and other defense signaling, and it highlights CRK3 and ACA12 in relation to herbivore responses.

1) What the paper did (pipeline check)

Phenotyping with DL: It trains multiple classic CNNs (AlexNet, VGG16, ResNet50/101, InceptionV3, DenseNet121) and selects VGG16 as best for binary mild/severe classification, reporting 90.1% average accuracy during CV and 95.3% on an independent test set.
Regression for continuous severity: It uses a VGG16-based DCNN-PDS architecture with residual blocks and reports MSE/MAE plus Pearson r=0.94 and R²=0.88 on the test set.
Genome-wide association: It performs GWAS on DL-derived binary and continuous traits and reports significant QTL counts and candidate genes.
Integration with transcriptomics: It uses RNA-seq under herbivory to identify DEGs and intersects these with GWAS candidate genes to prioritize pest-resistant candidates (e.g., ACA12, CRK3).
Genomic selection (GS): It tests ML GS models for binary and continuous traits using LD-pruned SNP sets selected by stepwise GWAS P-value ranking; it reports 95.7% accuracy (binary) and 0.90 correlation (continuous) as final test performance.

2) Scientific quality appraisal (skeptical review)

2.1 Strengths

Two-level phenotyping: The paper argues that binary traits can miss minor-effect genetic architecture, so it also uses continuous regression labels derived directly from images (0–5 scale).
Independent test sets for DL metrics: It reports separate held-out test sets for both classification and regression.
Pipeline plausibility: The reported workflow (DL phenotyping → GWAS → RNA-seq overlap → GS) is coherent and, if correctly implemented, matches common integrative plant genetics logic.

2.2 Red flags / potential failure modes (what could break the conclusions)

Phenotype-to-genotype circularity risk: GWAS and GS depend on DL-derived labels. Even with a held-out DL test set, any systematic bias in labeling (e.g., leaf imaging conditions, background cues, subtle confounds) can propagate into downstream association and prediction. The paper states multi-rater labeling to reduce subjective bias, but it is still a human labeling process used to supervise the DL model.
Generalization across grape panels/environments not shown: The results report test-set performance within the paper’s experimental panel and imaging regime. The core claim of breeding utility would be strengthened by external validation on independent grape populations/environments, but this is not explicitly provided in the supplied text.
Overfitting via SNP subset search: GS SNP counts are chosen using stepwise selection based on GWAS P-value ranking, evaluated over iterations. Such workflows can inflate apparent predictive performance if the same data influence both marker selection and evaluation, unless strict nested validation is used. The paper describes cross-validation and repeated iterations, but the supplied text does not fully specify the nesting rigor.
Population structure & LD: GWAS uses a mixed model framework (GEMMA) and includes SNP filtering, but with 231 accessions, resolution and false positives can remain concerns, especially with complex structure. The paper does report phylogenetic/regression nonrandomness tests and population structure analysis summaries, but independent replication is not evidenced in the supplied text.
Mechanistic claims remain hypothesis-level: The paper proposes candidates like CRK3 and ACA12 and links them to JA/calcium-related defense logic and prior plant studies. However, the grape-specific functional validation of these candidates against SLF is not part of the supplied text.

3) Mechanistic synthesis (careful: what is claimed vs supported)

What is supported in the paper’s own data:

Association signals: It reports QTLs and candidate genes from GWAS for both binary and continuous phenotypes.
Expression under herbivory: It reports DEGs under herbivory (with specific directionality counts) and overlaps these with GWAS candidates.
Model-to-biology plausibility: It interprets candidates in known plant defense pathway logic (JA/SA/ethylene and phosphorylation/calcium).

What would most strongly change the story (disproving points):

If DL phenotypes are biased: re-labeled phenotypes (independent imaging/labeling pipelines) would need to preserve the GWAS loci and GS predictive power. The paper’s reliance on image labeling and DL supervision makes this a key hypothesis.
If gene expression overlap is not SLF-specific: transcriptomics in the paper uses herbivory contexts; if those DEGs don’t reflect the same causal mechanisms for SLF leaf damage in the GWAS panel, prioritization could be misled.

4) Reproducibility & data availability (based on provided text)

Resequencing data: deposited in NCBI BioProject PRJNA994294.
Code: scripts available on GitHub (linked in text as https://github.com/zhouyflab/Pest-Resistance).

5) Author-review jump links (for deeper critique)

Open targeted author reviews in BGPT.

Feedback:

Updated: March 22, 2026