Interpreting biologically informed neural networks for enhanced biomarker discovery and pathway analysis

Abstract

The advent of novel methods in mass spectrometry-based proteomics allows for the identification of biomarkers and biological pathways which are crucial for the understanding of complex diseases. However, contemporary analytical methods often omit essential information, such as protein abundance and protein co-regulation, and therefore miss crucial relationships in the data. Here, we introduce a generalized workflow that incorporates proteins, their abundances, and associated pathways into a deep learning-based methodology to improve biomarker identification and pathway analysis through the creation and interpretation of biologically informed neural networks (BINNs). We successfully employ BINNs to differentiate between two subphenotypes of septic acute kidney injury (AKI) and COVID-19 from the plasma proteome and utilize feature attribution-methods to introspect the networks to identify which proteins and pathways are important for distinguishing between subphenotypes. Compared to existing methods, BINNs achieved the highest predictive accuracy and revealed that metabolic processes were key to differentiating between septic AKI subphenotypes, while the immune system was more important to the classification of COVID-19 subphenotypes. The methodology behind creating, interpreting, and visualizing BINNs were implemented in a free and open source Python-package: https://github.com/InfectionMedicineProteomics/BINN.

Explainable machine learning for the identification of proteome states via the data processing kitchen sink

“…Here, we refine and present our previous methods in a coherent package to reanalyze a dataset consisting of patients with COVID-19 at varying levels of severity (1-7 WHO-grade) [15] using the full suite of DPKS methods. As we have previously reanalyzed this data to demonstrate the application of BINNs [13], here, we additionally show that explainable machine learning with recursive feature elimination can be used to identify a panel of 10 highly accurate predictive protein biomarkers in minutes that may be overlooked using classic statistical analysis.…”

“…Since we know the expected ratios of yeast and mouse proteins between sample groups, we can use this data to verify that the algorithms implemented in DPKS are working correctly to quantify proteins. Previously, we have used DPKS to select a panel of protein biomarkers for septic AKI [12] and applied similar feature attribution methods to provide high accuracy classifiers and intelligent pathway analysis using biologically informed neural networks (BINN) [13] to show how protein biomarker panels can be identified using explainable machine learning that outperform proteins selected via classic statistical analysis. Here, we refine and present our previous methods in a coherent package to reanalyze a dataset consisting of patients with COVID-19 at varying levels of severity (1-7 WHO-grade) [15] using the full suite of DPKS methods.…”

“…The calculated importance of each protein towards a models prediction allows for the data driven selection of the most important features for a particular proteome state instead of using p-value thresholds. As we have previously described [12,13], proteins found as most important for the classification of a proteome state may not be the most statistically significant, and those with low fold-change may also be important predictive drivers (Supp. Figure7a).…”

“…As we have previously described [12,13], this panel of proteins would have been difficult to identify using classic differential abundance analysis, and emphasizes the importance of integrating machine learning methods into existing analysis (Figure 4b). Interestingly, some of the important features do not have significant p-values, and some of the more significant (low p-value) proteins are not considered most important in differentiating between severity subphenotypes of COVID-19.…”

“…Recently, the application of machine learning to proteomics experiments has been shown to improve the identification and selection of prognostic protein biomarkers in disease [12,13]. However, the implementation of machine learning algorithms in clinical protein biomarker detection experiments remains non-trivial, and expert knowledge must be used to select the correct hyperparameters, deal with class imbalances, and correctly deal with low sample numbers to ensure the models are trained properly and provide results that can generalize to new data.…”