AAIC Conference (2025)
Doyeong Hwang, Kiyoung Kim, Soorin Yim, Kyungwook Lee, Sungjoon Park, Daniel Gatti(Jackson Laboratory), Elissa Chesler(Jackson Laboratory), Amy Dunn(Jackson Laboratory), Kristen O'Conell(Jackson Laboratory)
Abstract
Background
The AD-BXD mouse dataset comprises of Single-nucleotide polymorphism(SNP), cognitive phenotype data, and non-cognitive phenotype data such as indirect calorimetry(IDC), glucose tolerance testing(GTT), and Amyloid-beta. While these data are promising for predicting AD progression and facilitating discovery of AD targets, challenges persist. Non-cognitive features, though critical, are sparse and pose difficulties in handling missing data points for prediction tasks. Additionally, the high dimensionality of SNP features relative to the limited number of available instances presents a significant computational challenge, requiring robust dimensionality reduction and feature selection methods to maximize predictive power.
Method
We developed AD prediction model for AD-BXD mouse. The model uses SNP and non-cognitive phenotype features as input, and targets cognitive phenotype. First, we performed bootstrapped k-NN imputation on non-cognitive phenotype features. Second, we used autoencoder for SNP dimension reduction, training encoder to extract necessary information needed to reconstruct the original SNP information. Third, latent vectors from autoencoder is concatenated with imputed non-cognitive phenotype features. MLP predictor predicts cognitive label using the concatenated input, training encoder and predictor to reduce the mean squared error between the predicted and groundtruth cognitive label. Second and third steps are iterated.
Result
Bootstrapped k-NN imputation with non-cognitive phenotypes produced 300 instances for old-age and 1018 instances for age-wise settings. MLP predictor using all features concatenated produced PCC(Pearson Correlation Coefficient) of 0.587 for CFM and 0.504 for CFA on old-age setting, as predictive performance. SNP dimension produced PCC 0.69 for CFM and 0.606 for CFA on old-age setting. (The feature importance analysis revealed biomarkers that are consistent with previously reported findings in the literature.)
Conclusion
Imputation with bootstrapped k-NN enabled full leverage of AD-BXD data for AD prediction. SNP dimension reduction boosted predictive power of the model. We assume that the dimensionality reduction enabled model to learn meaningful pattern from SNP feature. (Results from feature importance analysis shows that the prediction model can also be used for further target identification researches.)