Hyperparameter Optimization for Deep Learning in Soybean Genomic Prediction

A dataset containing results from 58,320 hyperparameter combinations for deep learning models applied to genomic prediction of soybean agronomic traits. The study by Diego Jarquin, last updated in April 2026, compares deep learning architectures against parametric models using two fivefold cross-validation schemes. Results indicate performance improvements of up to ~20% for certain models.

Use Cases

• Benchmarking deep learning hyperparameter optimization strategies based on the 58,320 tested combinations.
• Comparing parametric and deep learning models for genomic prediction based on the described M1, M2, and M3 model architectures.
• Analyzing the impact of activation functions and learning rates on model performance for agronomic traits.
• Designing cross-validation schemes for predicting new genotypes or incomplete field trials based on the CV1 and CV2 methods described.

Strengths

• Systematically tested 58,320 hyperparameter combinations, providing a broad search space.
• Includes comparisons to three established parametric models (M1, M2, M3) for benchmarking.
• Results quantify performance improvements, such as a ~20% gain for yield prediction over a baseline model.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the specific soybean study and experimental conditions.

Provenance

Source

Diego Jarquin via figshare

Collection Method

Results from a computational study optimizing deep learning architectures for genomic prediction.

Freshness

Last updated 2026-04-15 20:01:59; freshness should be verified.