BATL: Bayesian annotations for targeted lipidomics

Abstract

Motivation: Bioinformatic tools capable of annotating, rapidly and reproducibly, large, targeted lipidomic datasets are limited. Specifically, few programs enable high-throughput peak assessment of liquid chromatography-electrospray ionization tandem mass spectrometry data acquired in either selected or multiple reaction monitoring modes.

Results: We present here Bayesian Annotations for Targeted Lipidomics, a Gaussian naïve Bayes classifier for targeted lipidomics that annotates peak identities according to eight features related to retention time, intensity, and peak shape. Lipid identification is achieved by modeling distributions of these eight input features across biological conditions and maximizing the joint posterior probabilities of all peak identities at a given transition. When applied to sphingolipid and glycerophosphocholine selected reaction monitoring datasets, we demonstrate over 95% of all peaks are rapidly and correctly identified.

Availability and implementation: BATL software is freely accessible online at https://complimet.ca/batl/ and is compatible with Safari, Firefox, Chrome and Edge.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Common challenges associated with SRM, MRM and PRM peak identification. (a) Ambiguity occurs when multiple lipid isomers, isobars, and isotopes are detected within the same matrix at a given transition, yet technical variations in flow rate, composition of the mobile phase, temperature, pH, etc., cause their retention times to vary across samples. Data represent XICs of the same matrix (murine plasma) in animals fed different diets. Note six peaks are observed in one sample at a given transition. Seven peaks are observed in a different sample shifted by 1 min. Matching retention time would not align these shifted species. (b) Assigning lipid identities based on peak elution order (picking the nth eluting peak) will also lead to misidentifications when comparing lipid species across matrices. Data represent XICs of plasma and brain (temporal cortex) lipidomes from the same animal. Note both the retention time shift and the fundamentally different number of species within each matrix. Matching by either retention time or peak elution order would confound identification. (c) Matching lipids based on peak intensity features is complicated by pathological changes detected in lipid metabolism. Data represent XICs of the human plasma lipidome of patients with different neurodegenerative diseases. Note the marked change in abundances between conditions that impacts on lipid identification. While algorithms exist to address each of these challenges, few are applicable to datasets wherein all differences manifest simultaneously. BATL addresses these challenges

Fig. 2.

Schematic of the BATL lipid identification workflow. BATL follows three steps: (i) users are asked to identify training datasets for which they have unambiguous knowledge of peak identities. (ii) These datasets are used to train BATL, constructing a naïve Bayes statistical model based on the peak features users select. (iii) The model and associated metadata are used by the BATL algorithm to annotate peaks in subsequent query SRM, MRM or PRM datasets

Fig. 3.

Classifier performance on 10-fold cross validation sphingolipid and glycerophosphocholine datasets. The 95% confidence intervals are shown in panels (a, b, d and e). In a and b),data represent mean accuracies of BATL models trained on retention time with each decision rule and retention time mean/window matching algorithms for (a) sphingolipids or (b) glycerophosphocholines (***Q < 0.001, t-test adjusted with the Benjamini–Hochberg method of all models against the MWBM decision rule). (c) Lipid assignment differences between MAP, constrained MAP, and MWBM decision rules during cross validation and trained using retention time. In the top panel, data represent the Gaussian likelihoods of five glycerophosphocholine isomers based on the retention time feature. The rows of gray dots indicate the retention times of four peaks from the same sample in the validation set. Each row indicates the outcome of the three decision rules. Arrows indicate the lipid assignments; checkmarks indicate correct assignments; and Xs indicate incorrect assignments. The numbers for constrained MAP indicate the order of peak assignments. In d and e,data represent mean accuracies of the BATL models using MWBM decision rule trained on several features and feature combinations for (d) sphingolipids or (e) glycerophosphocholines. The feature name codes are described in Table  1