SparseCodePicking: feature extraction in mass spectrometry using sparse coding algorithms

Mass spectrometry (MS) is an important technique for chemical profiling which calculates for a sample a high dimensional histogram-like spectrum. A crucial step of MS data processing is the peak picking which selects peaks containing information about molecules with high concentrations which are of interest in an MS investigation. We present a new procedure of the peak picking based on a sparse coding algorithm. Given a set of spectra of different classes, i.e. with different positions and heights of the peaks, this procedure can extract peaks by means of unsupervised learning. Instead of an $l_1$-regularization penalty term used in the original sparse coding algorithm we propose using an elastic-net penalty term for better regularization. The evaluation is done by means of simulation. We show that for a large region of parameters the proposed peak picking method based on the sparse coding features outperforms a mean spectrum-based method. Moreover, we demonstrate the procedure applying it to two real-life datasets.

💡 Research Summary

Mass spectrometry (MS) generates high‑dimensional, histogram‑like spectra in which biologically or chemically relevant information is concentrated in a relatively small number of sharp peaks. Accurate identification of these peaks—known as peak picking—is a prerequisite for downstream tasks such as quantification, classification, and biomarker discovery. Traditional peak‑picking pipelines either apply simple thresholding to each individual spectrum or compute a mean spectrum across all samples and detect peaks on this aggregate representation. Both approaches suffer when the data contain multiple classes with distinct peak positions and intensities, when low‑abundance compounds generate weak peaks, or when noise levels are high.

In this work the authors propose a fundamentally different strategy: they treat peak picking as an unsupervised feature‑learning problem and employ sparse coding to discover a compact set of basis spectra (the “dictionary”) that can linearly reconstruct the entire dataset. Each observed spectrum X_i (a column of the data matrix X) is expressed as X_i ≈ D·α_i, where D ∈ ℝ^{m×K} is a dictionary of K basis vectors (each basis vector can be interpreted as a prototypical peak pattern) and α_i ∈ ℝ^{K} is a sparse coefficient vector indicating which bases are active for that sample.

The classic sparse‑coding formulation uses an ℓ₁ penalty on the coefficients to enforce sparsity. However, pure ℓ₁ regularization can be overly aggressive, discarding correlated peaks that should be jointly selected, and it can be unstable when the number of bases is comparable to the number of samples. To address these shortcomings the authors replace the ℓ₁ term with an Elastic‑Net penalty, i.e., a weighted sum of ℓ₁ and ℓ₂ norms (λ₁‖α_i‖₁ + λ₂‖α_i‖₂²). The ℓ₁ component still drives most coefficients to zero, preserving interpretability, while the ℓ₂ component encourages groups of correlated peaks to be selected together, improving robustness against noise and peak shift.

The learning algorithm proceeds by alternating minimization: (1) with D fixed, each α_i is obtained by solving a convex Elastic‑Net regression problem; (2) with all α_i fixed, D is updated by solving a least‑squares problem under unit‑norm constraints on its columns. This loop repeats until convergence or a preset iteration limit. After training, the non‑zero entries of the coefficient matrix A =