Abstract
Identifying complex phenotypes from high-dimensional biological data is challenging due to the intricate interdependencies among different physiological indicators. Traditional approaches often focus on detecting outliers in single variables, overlooking the broader network of interactions that contribute to phenotype emergence. Here, we introduce ODBAE (Outlier Detection using Balanced Autoencoders), a machine learning method designed to uncover both subtle and extreme outliers by capturing latent relationships among multiple physiological parameters. ODBAE's revised loss function enhances its ability to detect two key types of outliers: influential points (IP), which disrupt latent correlations between dimensions, and high leverage points (HLP), which deviate from the norm but go undetected by traditional autoencoder-based methods. Using data from the International Mouse Phenotyping Consortium (IMPC), we show that ODBAE can identify knockout mice with complex, multi-indicator phenotypes - normal in individual traits, but abnormal when considered together. In addition, this method reveals novel metabolism-related genes and uncovers coordinated abnormalities across metabolic indicators. Our results highlight the utility of ODBAE in detecting joint abnormalities and advancing our understanding of homeostatic perturbations in biological systems.
Introduction
In biological systems, "normal" is typically defined by specific ranges for individual variables, such as blood glucose levels or heart rate. However, these systems are governed by homeostasis—a dynamic state of balance where physiological variables are deeply interdependent. A disease state may not manifest as an extreme value in a single indicator but rather as a breakdown in the relationship between multiple indicators.
Traditional outlier detection methods, such as z-scores, evaluate variables in isolation. Even standard deep learning approaches like Autoencoders (AEs) face a significant limitation: they are often too effective at data reconstruction. By minimizing Mean Squared Error (MSE) to achieve near-perfect reconstruction, standard AEs can inadvertently learn to reconstruct outliers, rendering them invisible. This is particularly true for High Leverage Points (HLP)—data points that deviate from the norm but follow the general trend of the data.
Methods: Outlier Detection using Balanced Autoencoders
To address these limitations, the authors developed ODBAE (Outlier Detection using Balanced Autoencoders). The framework introduces several key innovations to move beyond simple reconstruction error:
- Revised Loss Function: ODBAE incorporates a penalty term ($L_{EIG}$) based on the eigenvalues of the data's covariance matrix. This term intentionally prevents "complete reconstruction," ensuring that outliers—both Influential Points (IP) and High Leverage Points (HLP)—produce significantly higher reconstruction errors.
- Principal Component Balancing: The model balances reconstruction error across all principal component directions, ensuring that anomalies in "quieter" dimensions are not overlooked.
- Explainability: Using Kernel-SHAP, ODBAE identifies the specific parameters contributing to an abnormal score, facilitating biological interpretation.
Results and Insights
ODBAE was validated using synthetic datasets, benchmark machine learning sets (Dry Bean, Breast Cancer), and the International Mouse Phenotyping Consortium (IMPC) dataset.
Identifying Hidden Phenotypes
A key finding is that "better" reconstruction accuracy often leads to poorer outlier detection. By forcing the model to be slightly less accurate via the $L_{EIG}$ penalty, it becomes more sensitive to manifold deviations. For instance, in Ckb gene-knockout mice, individual traits like body weight and length appeared normal, but ODBAE flagged them due to an abnormal BMI—a multi-indicator phenotype missed by univariate tests.
Novel Gene Discovery and Validation
Applying ODBAE to IMPC metabolic data identified 203 important metabolic genes, 91 of which were undetectable by traditional z-score methods. The study identified 43 genes with no prior metabolic links in the Mouse Genome Informatics (MGI) database. These were further validated by showing that their human orthologues are associated with Type 2 Diabetes through SNP (Single-Nucleotide Polymorphism) analysis.
Discussion and Future Work
While ODBAE demonstrates superior performance over standard AEs and models like DAGMM, it currently assumes a Gaussian distribution for "normal" data. Future work will focus on optimizing the model for non-Gaussian, multimodal, or skewed distributions, as well as extending the framework to time-series and imaging data. ODBAE represents a significant step forward in detecting joint abnormalities and understanding homeostatic perturbations in complex biological systems.
Summary updated