Lasso-MCR–ALS: Sparse MCR-ALS for LC/GC-MS

• Lasso-MCR–ALS is a sparse variant of MCR-ALS that uses an L1 penalty to resolve LC/GC-MS data into chemically interpretable concentration and spectral profiles.
• The method systematically enforces physicochemical constraints such as nonnegativity, closure, and shape restrictions, ensuring realistic decompositions of overlapping mass channels.
• Empirical comparisons demonstrate that the L1 penalty yields a sharply peaked optimization landscape, guiding the algorithm to a unique solution unlike L0 or L2 alternatives.

Lasso-MCR–ALS is a penalized regression variant of multivariate curve resolution–alternating least squares (MCR-ALS) designed to resolve complex chromatography-mass spectrometry (LC/GC-MS) data into chemically meaningful components under sparsity constraints. This methodology replaces the classical least-squares update for spectral profiles with a Lasso ($\ell_1$) penalty, promoting sparse solutions in the mass spectra matrix, thereby addressing rotational ambiguity inherent to MCR solutions, especially in the context of overlapping or embedded mass channels and background continuums (Mani-varnosfaderani et al., 2019).

1. Penalized Optimization Formulation

Let $D \in \mathbb{R}^{n \times m}$ denote the measured data matrix, with $n$ chromatographic retention points and $m$ mass channels. The model assumes a bilinear decomposition,

$D \approx C S^\top$

where $C \in \mathbb{R}^{n \times p}$ encodes $p$ concentration/time profiles and $S \in \mathbb{R}^{m \times p}$ contains the $p$ underlying mass spectra.

Classical MCR-ALS solves

$\min_{C,S} \|D - C S^\top\|_F^2.$

Lasso-MCR–ALS augments this with a sparsity-enforcing Lasso penalty on $S$:

$$\min_{C,S}\;\; \|D - C S^\top\|_F^2 + \lambda \|S\|_1,$$

where $\|S\|_1 = \sum_{i=1}^m \sum_{j=1}^p |S_{ij}|$ and $\lambda \ge 0$ modulates sparsity. The penalty may, in principle, be imposed on $C$, but typical usage targets the spectra $S$ (Mani-varnosfaderani et al., 2019).

2. Enforcement of Physicochemical Constraints

In addition to sparsity, several physicochemical constraints are systematically integrated into the ALS framework:

• Nonnegativity: $C_{ik} \geq 0$, $S_{jk} \geq 0$ for all $i, j, k$.
• Closure: $\sum_{k=1}^p C_{ik} = 1$ for every sample $i$ (when applicable).
• Shape constraints: e.g., unimodality or monotonicity on profiles $C$. After each step in the iterative procedure, these constraints are enforced to maintain physical interpretability and chemical plausibility (Mani-varnosfaderani et al., 2019).

3. Alternating Least Squares Updates with $\ell_1$ Penalty

Each Lasso-MCR–ALS iteration consists of two alternating updates:

• Update for $C$:

$\min_{C \geq 0} \|D - C \hat{S}^\top\|_F^2$

with explicit solution $$C \leftarrow \max (D \hat{S} (\hat{S}^\top \hat{S})^{-1}, 0)$$, followed by normalization and closure as needed.

• Update for $S$: For each spectral component $k = 1, \ldots, p$,

$$\min_{s^{(k)} \geq 0} \|d^{(k)} - C s^{(k)}\|_2^2 + \lambda \|s^{(k)}\|_1,$$

where $d^{(k)}$ is the $k$th column of $D$. This is a nonnegative Lasso regression. Solution strategies include coordinate descent with soft thresholding and proximal methods such as FISTA, always enforcing nonnegativity post-update.

This alternating update scheme enforces sparsity explicitly on $S$ while maintaining the overall bilinear structure and required physicochemical constraints (Mani-varnosfaderani et al., 2019).

4. Area of Feasible Solutions and $L_x$-Norm Grid Search

MCR suffers from rotational ambiguity: for any invertible $p \times p$ matrix $T$, the pair $(C T, S T^{-1})$ yields the same $D$. Imposing nonnegativity restricts $T$ to a subset of parameter space termed the area of feasible solutions (AFS).

Lasso-MCR–ALS assesses the effect of various $L_x$-norm penalties ($x \in [0,2]$) through a grid search over rotation parameters. In two-component systems, $T$ is parameterized as

$$T = \begin{pmatrix} 1 & t_{12} \ t_{21} & 1 \end{pmatrix},$$

and normalized for identifiability. For each parameter grid point, the norm

$E_x(t) = \sum_{k=1}^p \|s^{(k)}(t)\|_x$

is computed and analyzed. It is observed that the optimization surface's steepness grows as $x$ approaches zero, but for $x=0$ (the $L_0$-norm), a flat plateau arises, making minimization ill-posed. For $0$1 < x \leq 2$ they become convex (Mani-varnosfaderani et al., 2019).

5. Empirical Comparison: $L_1$ versus $L_0$ and $L_2$

Case studies on simulated LC/GC–MS datasets, including two- and three-component systems (with overlap and background), reveal:

• $L_2$-norm: Yields shallow gradients of $E_2$ in the AFS, leading to ambiguous minima and ineffective convergence toward the true solution.
• $L_0$-norm: Produces a constant value (flat plateau) within the nonnegativity band, preventing descent and yielding optimization indeterminacy.
• $L_1$-norm (Lasso): Delivers sharply peaked objective surfaces (diamond-shaped level sets in 2D), offering directional gradients that consistently guide the algorithm toward the true sparse, unique solution. For overlapping spectra, the minimum of $E_1$ coincides with the true decomposition, unlike $E_0$ or $E_2$. In three-component data, $E_1$ and $E_2$ may have minima at true rotations, but the $E_2$ surface remains too flat for robust identification, and $E_0$ can be constant within AFS regions (Mani-varnosfaderani et al., 2019).

6. Rationale for Lasso Selection and Theoretical Considerations

The $\ell_2$-norm penalty produces circular or elliptical contours with weak gradients, resulting in poor discrimination of solutions within the AFS. The $L_0$ penalty, counting nonzero entries, creates contiguous regions in the AFS with indistinguishable penalties, lacking any descent path and rendering minimization intractable.

The $L_1$ penalty is the sharpest convex function among the three, with diamond-shaped level sets that guarantee a single global minimum and sufficiently strong gradients to direct the optimization toward the true sparse solution, but without the combinatorial instabilities or plateaus of $L_0$. As $x$ scans from $2 \rightarrow 1^+ \rightarrow 0^+$, gradient sharpness increases, but collapses at $x=0$. $x=1$ thus represents a unique point balancing convexity and optimization tractability (Mani-varnosfaderani et al., 2019).

7. Practical Implementation and Parameter Selection

The regularization parameter $\lambda$ modulates the induced sparsity. Practical strategies for $\lambda$ selection include:

• Cross-validation using held-out chromatographic slices,
• "L-curve" monitoring of the tradeoff between fit ($\|D-C S^\top\|_F^2$) and sparsity ($\|S\|_1$),
• Calibration against reference spectra or internal standards to match expected numbers of nonzero peaks.

For real LC/GC-MS data:

• Preprocess $D$ using baseline correction and normalization, ensuring $S$ entries scale between $[0,1]$.
• Initialization may use Evolving Factor Analysis or PCA.
• Iteratively update $C$ via nonnegative least squares with closure constraint, then $S$ via nonnegative Lasso solvers (coordinate-descent, FISTA).
• Stop iterations when the residual norm change falls below a threshold (e.g., $\epsilon \sim 10^{-6}$), or after a fixed iteration count.
• Typical $\lambda$ values range from $10^{-3}$ to $10^{-1}$ in normalized regimes, with dataset-specific tuning required.

In summary, Lasso-MCR–ALS replaces standard ALS spectral updates with Lasso-penalized, nonnegative regression steps:

$$S \leftarrow \underset{S \geq 0}{\arg\min}\; \|D - C S^\top\|_F^2 + \lambda \|S\|_1,$$

substantially improving sparsity, solution uniqueness, and minimizing rotational ambiguity relative to $L_2$ or $L_0$ alternatives in LC/GC–MS multivariate curve resolution (Mani-varnosfaderani et al., 2019).