[Paper Review] Learning Concave Conditional Likelihood Models for Improved Analysis of Tandem Mass Spectra

John T. Halloran, David M. Rocke|PubMed|Sep 4, 2019

Advanced Proteomics Techniques and Applications24 references3 citations

TL;DR

This paper introduces Convex Virtual Emissions (CVEs), a novel class of emission distributions that ensure concave conditional log-likelihood scoring in dynamic Bayesian networks, enabling global convergence during parameter learning. By integrating CVEs into the Didea peptide-spectrum matching algorithm, the authors achieve state-of-the-art scoring accuracy and a 64.2% speedup in inference, outperforming DRIP and MS-GF+ by identifying 16% more spectra at 1% FDR.

ABSTRACT

The most widely used technology to identify the proteins present in a complex biological sample is tandem mass spectrometry, which quickly produces a large collection of spectra representative of the <i>peptides</i> (i.e., protein subsequences) present in the original sample. In this work, we greatly expand the parameter learning capabilities of a dynamic Bayesian network (DBN) peptide-scoring algorithm, Didea [25], by deriving emission distributions for which its conditional log-likelihood scoring function remains concave. We show that this class of emission distributions, called <i>Convex Virtual Emissions</i> (CVEs), naturally generalizes the log-sum-exp function while rendering both maximum likelihood estimation and conditional maximum likelihood estimation concave for a wide range of Bayesian networks. Utilizing CVEs in Didea allows efficient learning of a large number of parameters while ensuring global convergence, in stark contrast to Didea's previous parameter learning framework (which could only learn a single parameter using a costly grid search) and other trainable models [12, 13, 14] (which only ensure convergence to local optima). The newly trained scoring function substantially outperforms the state-of-the-art in both scoring function accuracy and downstream Fisher kernel analysis. Furthermore, we significantly improve Didea's runtime performance through successive optimizations to its message passing schedule and derive explicit connections between Didea's new concave score and related MS/MS scoring functions.

Motivation & Objective

• To overcome the limitations of previous Didea parameter learning, which relied on costly grid searches and could only optimize a single parameter.
• To develop a general class of emission distributions that ensure concave conditional log-likelihood for efficient, globally convergent parameter learning in Bayesian networks.
• To significantly improve the runtime efficiency of Didea’s sum-product inference without sacrificing accuracy.
• To enhance downstream discriminative analysis by deriving conditional log-likelihood gradients for use in kernel-based postprocessors.
• To establish theoretical connections between Didea’s new scoring function and widely used methods like XCorr.

Proposed method

• Derive a class of emission distributions called Convex Virtual Emissions (CVEs) by solving a nonlinear differential equation that enforces convexity conditions on general Bayesian network emissions.
• Show that CVEs generalize the log-sum-exp function and preserve concavity in both maximum likelihood and conditional maximum likelihood estimation.
• Integrate CVEs into the Didea dynamic Bayesian network model to enable efficient, global optimization of a large number of parameters.
• Optimize Didea’s message passing schedule through algorithmic refinements, reducing runtime by 64.2%.
• Derive conditional log-likelihood gradients of Didea’s score to enrich the feature space for kernel-based discriminative postprocessing.
• Establish a theoretical lower bound of Didea’s score on the widely used XCorr scoring function, enabling future parameter learning for XCorr.

Experimental results

Research questions

• RQ1 Can a general class of emission distributions be derived that ensures concave conditional log-likelihood for scalable parameter learning in dynamic Bayesian networks?
• RQ2 How can Didea’s parameter learning be extended beyond single-parameter grid search to enable efficient, global optimization of multiple parameters?
• RQ3 To what extent can Didea’s inference speed be improved without compromising scoring accuracy?
• RQ4 Can the gradients of Didea’s conditional log-likelihood provide richer discriminative features than existing methods for post-processing?
• RQ5 Is there a theoretical relationship between Didea’s new scoring function and established MS/MS scoring functions like XCorr?

Key findings

• The proposed Convex Virtual Emissions (CVEs) class generalizes the log-sum-exp function and ensures concave conditional log-likelihood, enabling global convergence during parameter learning.
• The new Didea model with CVEs outperforms DRIP and MS-GF+ by identifying 16% more spectra at a strict 1% FDR across benchmark datasets.
• The optimized Didea implementation achieves a 64.2% reduction in inference time, reducing search time to under 7 seconds per spectrum—two orders of magnitude faster than DRIP.
• The conditional log-likelihood gradients of Didea provide significantly more informative features than DRIP’s log-likelihood gradients, enabling Percolator to achieve superior recalibration performance.
• The trained Didea scoring function identifies 12.3% more spectra than DRIP and 13.4% more than MS-GF+ at 1% FDR, outperforming all state-of-the-art methods in downstream analysis.
• A theoretical lower bound is established between Didea’s score and the XCorr scoring function, enabling future parameter learning for XCorr using the concave Didea framework.