Local SNP encoding
Shared 1D-CNN layers encode ordered SNP regions using genotype dosage, heterozygosity, call masks and positional information.
MAIZE · DROUGHT RESPONSE · GENOMIC PREDICTION
From genomic variation
to drought-related traits.
BioGNet: a deep learning framework for maize phenotype prediction
Explore a model that brings local SNP patterns, biological annotation and genome-wide relationships together to support research on maize breeding.
1,319 materials456 inbred lines · 863 F1 hybrids
Four historical environments2022–2025 drought phenotyping
Seven prediction tasksASI · PH · SEW · composite D
Manual research reviewCompatibility checked before analysis
THE MODEL
Local signals. Biological context. Genetic background.
BioGNet predicts drought-related phenotypes using a shared SNP encoder, biological attention and a single genomic relationship branch. The outputs are predicted trait values.
Biological attention
Two attention blocks connect regions with LD, distance and annotation priors. Six functional slots and two background slots summarize regional information.
GRM & material fusion
A single 64-dimensional GRM input and material-type embedding join the regional representations before phenotype regression.
ASIAnthesis–silking interval · days
PHPlant height · cm
SEWSingle-ear weight · g/ear
DComposite drought coefficient · index
WW and WS denote well-watered and drought-stress tasks. D is one composite endpoint built using a frozen historical transformation. The eight attention slots are distinct from the single GRM branch.
HELD-OUT BENCHMARK
Compare models within the target materials.
Explore 13 models across seven tasks. Choose a material cohort and metric to inspect the archived test results.
These archived comparisons use the same frozen predictions as the manuscript. BioGNet uses a single seed selected post hoc by test Pearson r; three-seed mean results are a separate evaluation. R² is 1 − SSE/SST and can be negative. Pooled results include differences between material types.
PREPARE YOUR DATA
A clear starting point for review.
Submit one ZIP using the documented CSV/TSV structure. Basic file checks run at submission; full genotype alignment and scientific compatibility are reviewed by the team.
- Download the template.
Include
genotypes.csv,markers.tsv,samples.tsvandREADME.md. The example uses synthetic materials and a few example markers. - Describe the materials.
Identify F1 hybrids or inbred lines. Preserve mother and father direction for hybrids. Use AGPv4 positions and explicit REF/ALT alleles.
- Choose the analysis.
Prediction only requires compatible genotypes and metadata. Evaluation also needs observed phenotypes. Evaluating observed D requires the three traits paired under WW and WS within the same environment.
Use ALT hard-call dosages 0, 1 and 2; missing calls are NA, blank or dot. Markers and alleles must be aligned to the relevant frozen task panel and training-reference GRM before inference. The template demonstrates format and is not an inference-ready dataset.
Track your request
Use your reference ID and submission email to check the review status. You can also record a deletion request for the team.
About the project
BioGNet studies genomic prediction of maize drought-related traits using four years of historical phenotyping. Model attribution and candidate-gene annotations provide research leads rather than experimental proof of variant function.
Predictions, rankings and any validation metrics depend on input compatibility and the target genetic background. Metrics require observed phenotypes and are not generated for genotype-only submissions.