Protein abundance inference via expectation-maximization in fluorosequencing

TL;DR

This paper introduces a new method using expectation-maximization to estimate protein abundances from fluorosequencing data, improving accuracy and scalability.

Contribution

A novel probabilistic framework using EM for fluorosequencing data to infer protein abundances is introduced.

Findings

The algorithm processes one million reads in under ten seconds and reduces error significantly.

Ten million reads are processed in under four hours on a GPU, showing scalability.

Improved fluorosequencing chemistry could lead to more accurate protein abundance estimates.

Abstract

Fluorosequencing generates millions of single peptide reads, yet a principled route to quantitative protein abundances has been lacking. We present a probabilistic framework that adapts expectation–maximization (EM) to the fluorosequencing measurement process, using posterior peptide probabilities from existing classifiers to estimate relative protein abundances. The algorithm iteratively updates abundances to maximize the likelihood of observed reads. We first evaluate five-protein simulations with realistic labeling and system errors. A simple Python implementation processes one million reads in under ten seconds on a standard workstation and reduces the mean absolute error by over an order of magnitude relative to a uniform-abundance guess, indicating robust performance in small-scale settings. We also assess scalability with full human-proteome simulations (20 642 proteins). Ten million reads are processed in under four hours on an NVIDIA DGX with a single Tesla V100 GPU, confirming tractability at proteome scale. Under current fluorosequencing error rates, the method yields modest accuracy gains, but when error rates are reduced, estimation error drops markedly, indicating that chemistry improvements would translate directly into more accurate quantitative proteomics. Overall, EM-based inference provides a scalable, model-driven bridge from peptide-level classification to protein-level quantification in fluorosequencing. Furthermore, the framework can also serve as a refinement step within other inference methods.

The code and data utilized to produce all the results of this paper is at https://github.com/JavierKipen/ProtInfGPU.

Figures