Eye-color and Type-2 diabetes phenotype prediction from genotype data using deep learning methods.

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Machine Learning in Healthcare

Background:

• Genotype-phenotype predictions are crucial for identifying genetic mutations linked to human variations.
• Existing methods include statistical techniques for pinpointing causal SNPs and machine learning for classification.
• A hybrid approach integrating both statistical and machine learning components is proposed.

Purpose of the Study:

• To develop and evaluate a hybrid genotype-phenotype prediction model.
• To assess the model's performance on eye-color and Type-2 diabetes datasets.
• To explore the utility of machine learning classifiers in genotype-phenotype association studies.

Main Methods:

• Preprocessed genotype data (SNPs) into multiple datasets based on mutation differences.
• Calculated mutation types (none, partial, full) for each Single Nucleotide Polymorphism (SNP).
• Applied various machine learning classifiers, including Random Forest, Gradient Boosting, ANN, LSTM, GRU, BILSTM, and 1DCNN, alongside ensemble methods.

Main Results:

• For eye color, stacked ensembles of LSTM achieved the highest accuracy (0.96) with AUCs of 0.98 (brown) and 0.97 (blue-green) for 1560 SNPs.
• Different classifiers yielded accuracies ranging from 0.91 to 0.96 for eye-color prediction.
• The model achieved 0.97% accuracy for Type-2 diabetes prediction using optimized SNPs.

Conclusions:

• Genotype-phenotype predictions, particularly using machine learning, are valuable for identifying SNP-trait associations, with forensic applications noted.
• The hybrid model's non-linearity and combined SNP mutation analysis enhance prediction accuracy.
• The approach is effective for binary classification and can be extended to multi-class problems.