Sparse Proteomics Analysis - a compressed sensing-based approach for feature selection and classification of high-dimensional proteomics mass spectrometry data

Tim O. F. Conrad, Martin Genzel, Nada Cvetkovic, Niklas Wulkow, Alexander B. Leichtle, Jan Vybíral, Gitta Kutyniok, Christof SchütteBMC Bioinformatics 2017 · OpenReview

TL;DR

This paper introduces Sparse Proteomics Analysis (SPA), a novel method that leverages compressed sensing principles to identify a minimal set of disease-specific features from high-dimensional mass spectrometry data. SPA offers a robust and interpretable approach to disease fingerprinting and classification, outperforming existing methods in "very-sparse" scenarios.

What Problem Does This Paper Address?

High-throughput proteomics techniques, like mass spectrometry (MS), generate massive datasets with millions of data points (features) per sample. In clinical settings, the goal is to identify subtle differences in these spectra between healthy individuals and those with a particular disease. This requires machine learning algorithms to pinpoint "discriminating features" that form a "disease fingerprint" and then use this fingerprint to classify new samples. However, MS data is notoriously noisy, and the sheer dimensionality makes traditional feature selection and classification challenging, often leading to overfitting and difficulty in interpreting the results. The core problem is to find a small, robust set of features that accurately distinguishes between disease states, enabling easier interpretation and potentially lower-cost diagnostics.

Methods

The authors propose Sparse Proteomics Analysis (SPA), a framework built upon the theory of 1-bit compressed sensing. The core idea is to treat the classification problem as a sparse recovery problem. SPA involves several key steps:

  1. Preprocessing:
    • Normalization: Spectra are normalized (e.g., by total ion count) to ensure comparability.
  2. Smoothing: A Gaussian kernel convolution is applied to reduce noise while preserving spectral structure.
  3. Standardization: Data is centered and scaled to have unit standard deviation, ensuring features are equally weighted.
  4. Sparse Feature Selection: This is the heart of SPA. It uses a 1-bit compressed sensing optimization (specifically, a convex relaxation of the 1-bit recovery problem) to find a sparse feature vector. This optimization aims to maximize the agreement between the predicted and true class labels while enforcing sparsity constraints on the feature vector.
  5. Postprocessing:
    • Connected Components Detection: Since MS peaks are broad, significant features often appear as intervals rather than isolated points. This step identifies these "connected components" of non-zero features and reduces each interval to its most significant entry, further enhancing sparsity and interpretability.
  6. Dimension Reduction (Optional): The identified sparse feature vector can be used to project the original data into a lower-dimensional space, reducing the risk of overfitting and enabling the use of standard ML tools.

The key novelty lies in adapting 1-bit compressed sensing, which typically deals with sign-only measurements, to the classification problem and incorporating specific preprocessing and postprocessing steps tailored for MS data.

Datasets

The paper evaluates SPA on:

  • Simulated Data-Sets: Two types of simulated datasets (DS1 and DS2) were generated. These datasets mimic MS data with a known ground-truth sparse feature vector (five "condition positive" peaks) and varying levels of noise (SNR of 10 and 3.33). DS2 introduces a non-trivial correlation structure among features to test robustness.
  • Real-World Data-Sets:
    • Spiked Data: A labeled dataset created by spiking human blood samples with varying concentrations of specific proteins. This dataset offers a semi-ground truth, as the spiked peaks are known, but their presence in all spectra is not guaranteed.
  • Pancreas Cancer Data (P. CA): Two real-world datasets from patients with pancreatic cancer and controls, acquired from different clinical centers (Leipzig and Heidelberg). For this dataset, the true-positive feature positions are unknown, and evaluation relies solely on classification performance.

Key Insights and Ideas

  • 1-Bit Compressed Sensing for Classification: The paper demonstrates that the principles of 1-bit compressed sensing, originally designed for signal recovery from sign-only measurements, can be effectively adapted for robust feature selection and classification in high-dimensional biological data.
  • Sparsity as a Virtue for Interpretability: The emphasis on identifying a minimal set of features is crucial for clinical applications. Sparse features are easier for medical experts to interpret, validate, and translate into diagnostic assays, reducing the cost and complexity of follow-up experiments.
  • Handling MS Data Characteristics: SPA's preprocessing (smoothing, standardization) and postprocessing (connected components) are specifically designed to address the inherent noise and peak structure of MS data, leading to more reliable feature identification.
  • Robustness to Noise: The compressed sensing framework inherently provides robustness to noise, a critical factor in real-world MS data.

Key Results

  • Competitive Performance: SPA demonstrates competitive performance against standard algorithms like Lasso and l1-SVM on both simulated and real-world datasets.
  • Superiority in Sparse Scenarios: In situations where a very small number of features are truly discriminative (the "very-sparse" regime), SPA often outperforms other methods, achieving high accuracy with fewer selected features.
  • High Accuracy on Real Data: On the pancreas cancer dataset, SPA achieved high balanced accuracy (e.g., ~0.98 for UHL, ~0.97 for UHH) with a relatively small number of features (around 20-30), suggesting its effectiveness in identifying clinically relevant biomarkers.
  • Robustness to Noise: The simulated data experiments showed that SPA maintains good performance across different noise levels (SNR 10 and 3.33).
  • Spiked Data Challenges: Performance on the spiked data was relatively lower across all methods, attributed to the inherent variability in the presence of spiked peaks in real samples.

Limitations and Future Work

  • Assumption of Sparse Ground-Truth: A core assumption of SPA is that a sparse ground-truth feature vector exists. If the underlying biological mechanisms are too complex to be captured by a sparse model, SPA (like other sparsity-promoting methods) will likely fail.
  • Missing Data Handling: While the paper acknowledges the "missing-data problem" in MS, it primarily relies on preprocessing to mitigate its impact on feature identification. For classification, it suggests stopping the process if data is missing, rather than inferring it. This is noted as a crucial point for future pipeline development.
  • Generalization of Feature Space: The paper mentions that the analysis coefficients from the Gaussian dictionary offer opportunities for generalization, suggesting future work could explore redundant dictionaries or more sophisticated functions.
  • Computational Scalability: While not the primary focus, the paper notes that standard convex optimization methods can scale poorly with very large datasets, hinting at potential areas for algorithmic improvement.

Summary updated

Abstract

Background High-throughput proteomics techniques, such as mass spectrometry (MS)-based approaches, produce very high-dimensional data-sets. In a clinical setting one is often interested in how mass spectra differ between patients of different classes, for example spectra from healthy patients vs. spectra from patients having a particular disease. Machine learning algorithms are needed to (a) identify these discriminating features and (b) classify unknown spectra based on this feature set. Since the acquired data is usually noisy, the algorithms should be robust against noise and outliers, while the identified feature set should be as small as possible. Results We present a new algorithm, Sparse Proteomics Analysis (SPA), based on the theory of compressed sensing that allows us to identify a minimal discriminating set of features from mass spectrometry data-sets. We show (1) how our method performs on artificial and real-world data-sets, (2) that its performance is competitive with standard (and widely used) algorithms for analyzing proteomics data, and (3) that it is robust against random and systematic noise. We further demonstrate the applicability of our algorithm to two previously published clinical data-sets.
BibTeX
@article{conrad2017sparse,
  author = {Tim O. F. Conrad and Martin Genzel and Nada Cvetkovic and Niklas Wulkow and Alexander B. Leichtle and Jan Vybíral and Gitta Kutyniok and Christof Schütte},
  title = {{Sparse Proteomics Analysis - a compressed sensing-based approach for feature selection and classification of high-dimensional proteomics mass spectrometry data}},
  journal = {BMC Bioinformatics},
  year = {2017},
  url = {https://openreview.net/forum?id=XVsk3B0Dkf},
}
Figure