ARCHEX, short for ARCHetype-based EXplainer, is built on archetypal analysis, a decomposition technique introduced by Adele Cutler and Leo Breiman in 1994. The method represents every data point as a convex mixture of a small number of extreme profiles, or archetypes, that sit on the boundary of the data cloud. Instead of asking which cluster a point belongs to, archetypal analysis asks which archetypes it is a blend of. A patient record, for example, might be expressed as sixty percent of one extreme profile and forty percent of another. Weinberg’s insight is that these interpretable extremes can serve as a compressed coordinate system for the entire dataset, reducing thousands of raw features to a handful of meaningful dimensions.

The technical pipeline is deliberately simple. ARCHEX first identifies k archetypes from the training data, where k is chosen adaptively and only on the training partition to avoid information leaking into evaluation. Every observation is then projected onto the probability simplex, meaning it receives a set of non-negative membership weights across the archetypes that sum to one. These k-dimensional representations become the inputs to a linear surrogate model trained to reproduce the original black-box model’s prediction target. Because the surrogate is linear and non-black-box, its coefficients can be read directly as a global feature ranking, giving analysts a transparent approximation of how the underlying model behaves across the whole data distribution rather than at a single point.

One of the paper’s most technically interesting contributions is a precise characterization of where ARCHEX’s sparsity comes from. Sparse explanations—ones that highlight only a few features—are prized in interpretability research, and many methods engineer them through L1 regularization, which penalizes the sum of absolute coefficient values. Weinberg shows that ARCHEX needs no such penalty. Because the archetype membership weights lie on the probability simplex, their L1 norm is constant by construction: the weights always sum to one. Sparsity therefore emerges from the simplex projection itself, a geometric constraint rather than a tuning knob. This observation connects the framework to efficient projection algorithms onto the L1 ball and gives the method a form of built-in parsimony that does not have to be traded off against accuracy.

The evaluation is notable for its statistical caution, a quality often missing from explainability research. ARCHEX was tested on five public benchmark datasets spanning four domains, drawn from standard repositories such as UCI and scikit-learn. Rather than reporting single point estimates, the study uses bootstrap confidence intervals on both predictive performance and on the rank correlation between ARCHEX’s archetype-derived feature ranking and SHAP’s local attribution ranking. The results are striking: in four of the five datasets, that rank correlation is small in magnitude and, once sampling uncertainty is quantified, statistically indistinguishable from zero. In other words, there is no reliable evidence that ARCHEX is simply rediscovering what SHAP already tells you.

The fifth dataset complicates the story in an instructive way. On the Digits dataset, the confidence interval excludes zero, and the correlation between the two rankings is moderate and positive rather than low. Weinberg argues that this cuts against a common practice in the field: treating a low correlation point estimate as established proof that a new method offers complementary explanatory content. Without confidence intervals, a researcher might see a weak correlation and claim novelty; with proper uncertainty quantification, the claim may dissolve. The paper thus doubles as a methodological warning about how interpretability methods are compared, echoing earlier sanity-check studies that questioned whether saliency methods measure what they claim.

Beyond the ranking analysis, the paper includes an ablation that tests the value of soft membership. ARCHEX assigns each point a convex blend of archetype weights, while a hard variant based on k-means cluster membership forces each point into a single cluster. Across all five datasets, the soft convex membership outperformed the hard ablation on held-out predictive accuracy. The result makes intuitive sense: real data rarely falls neatly into discrete buckets, and allowing partial membership preserves geometric information that hard assignments discard. For practitioners, it suggests that the smoothness of the archetype representation, not merely the choice of extreme profiles, is doing real work in approximating the decision boundary.

Among the concrete findings, the Breast Cancer Wisconsin dataset offers the most vivid illustration. ARCHEX’s top-ranked features there are measurements related to concavity and concave points—shape characteristics of cell nuclei that describe how irregular a cell’s outline is. Several of these features are ranked far lower by SHAP. Weinberg is careful to present this as a preliminary, descriptive observation rather than a validated clinical finding, and that restraint matters: no prospective clinical study supports a diagnostic claim, and the author explicitly frames the result as a hypothesis-generating signal. Still, it shows how a global, archetype-based lens can surface feature relationships that local attribution methods, focused on individual predictions, may systematically underweight.

Where does ARCHEX fit in the crowded landscape of explainable AI? Weinberg positions it as a global data-structure and subgroup-discovery method that complements, rather than replaces, local attribution tools. SHAP and LIME answer the question of why this particular prediction was made; ARCHEX answers which extreme profiles define the data and how the model’s boundary behaves across them. This distinction echoes a broader debate in the field, from Cynthia Rudin’s argument that high-stakes decisions should rely on inherently interpretable models to concept-based approaches like TCAV that look beyond per-feature attributions. ARCHEX adds a geometric, archetype-centered voice to that conversation, grounded in a decomposition technique with a three-decade pedigree.

The framework does come with caveats. The core optimization implementation is proprietary and under development for commercial use, so the source code is not publicly available—a limitation for a paper whose central claim is about statistical rigor and reproducibility. To mitigate this, the manuscript provides complete algorithmic pseudocode, optimization hyperparameters, preprocessing procedures, and a machine-readable archive of the numerical results behind the tables and figures, allowing independent verification of the reported outcomes even without the exact code. Whether ARCHEX’s archetype lens becomes a standard complement to SHAP will depend on replication by other groups, but the paper’s insistence on confidence intervals before claiming complementarity is a standard the rest of explainable AI would do well to adopt.

Subject of Research: Archetype-based global interpretability framework for explaining machine learning models

Article Title: ARCHEX: explaining models through archetypal decomposition for discovering complementary patterns in model interpretability

Article References: Weinberg, A. I. (2026). ARCHEX: explaining models through archetypal decomposition for discovering complementary patterns in model interpretability. Applied Intelligence, 56(15), Article 475. https://doi.org/10.1007/s10489-026-07506-5

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07506-5

Keywords: explainable AI, archetypal analysis, model interpretability, SHAP, LIME, machine learning, statistical rigor, feature ranking, surrogate models, Applied Intelligence, data science, black-box models

The intellectual foundation of ParTIpy is Pareto Task Inference, or ParTI, a framework grounded in the theory of multi-objective optimality. The core idea is elegant: when evolution drives phenotypes toward Pareto optimality, any improvement in performance on one task necessarily degrades performance on another. Under this assumption, phenotypes such as organisms, proteins, or cells, described by traits that embody trade-offs, are predicted to occupy a simple geometric structure called a polytope, a generalized shape with corners, straight edges, and flat faces, such as a triangle or tetrahedron extended to any number of dimensions. Crucially, the number of vertices of this polytope equals the number of tasks the system is optimizing. Each vertex, known as an archetype, represents a phenotype perfectly specialized for one particular task.

To recover these archetypes from data, the framework relies on archetypal analysis, a dimensionality reduction technique first introduced by Cutler and Breiman in 1994. The method identifies extreme points within a dataset, the archetypes, and then models every observation as a nonnegative linear combination of them. Formally, the archetypes are defined as convex combinations of the data points, ensuring they lie within the convex hull of the observed distribution, and each data point is reconstructed as a convex combination of the archetypes. When applied to single-cell transcriptomics, this means fitting a polytope around cells in gene expression space, where each vertex corresponds to a transcriptional program optimized for a distinct biological task. Positioning a cell relative to the vertices then reveals its bias toward one or several functions.

Previous studies have demonstrated the value of this perspective, revealing trade-offs in hepatocytes, fibroblasts, and cancer cells, and even extending the approach to spatial transcriptomics and proteomics, where groups of co-localized cells cooperate on shared tasks. However, the original ParTI software, written in MATLAB, could not scale to the massive datasets generated by modern high-throughput single-cell technologies, and its dependence on a commercial license hindered broader adoption. ParTIpy addresses both limitations directly, implementing state-of-the-art initialization and optimization strategies in Python and integrating seamlessly with the scverse ecosystem of single-cell analysis tools, including scanpy and AnnData objects.

A central technical challenge in archetypal analysis is that the underlying optimization problem is non-convex, meaning solution quality is highly sensitive to how the algorithm is initialized. The team benchmarked three initialization algorithms and two optimization procedures across twenty-one single-cell transcriptome datasets, each representing a single cell type and drawn from three independent studies spanning multiple sclerosis lesions, spatial Xenium data, and lupus patient blood. They found that combining principal convex hull analysis, or PCHA, for optimization with the Archetypal++, or AA++, initialization scheme delivered the best trade-off between reconstruction accuracy and runtime. When speed is the priority, a more efficient Frank-Wolfe variant provides a practical alternative. The package also supports a convexity relaxation parameter that allows archetypes to extend beyond the observed data cloud, a crucial option when sampling near the boundaries of the true phenotypic space is sparse.

The most striking computational innovation, however, is ParTIpy’s use of coresets. A coreset is a small, cleverly weighted subset of the data such that optimizing on the subset yields results close to those obtained from the full dataset. The researchers incorporated coreset weights directly into the objective function and modified their optimization algorithms accordingly. Their benchmarks showed that the required coreset fraction shrinks as datasets grow: for datasets exceeding one hundred thousand cells, analyzing just one to ten percent of the data was sufficient on average, cutting runtime roughly fourfold. When fitting six archetypes on one million cells using a ten percent coreset, the optimization completed in forty-five to one hundred five seconds. Critically, archetype-specific gene expression profiles derived from coresets remained highly concordant with full-data results, with a median Pearson correlation above 0.9 across coreset fractions greater than one percent.

Because the correct number of archetypes is never known in advance, ParTIpy implements three complementary selection criteria. The first examines the fraction of variance explained as archetypes are added, looking for an elbow where returns diminish. The second employs an information-theoretic criterion that penalizes model complexity, favoring solutions that balance goodness of fit against the number of free parameters. The third assesses robustness through bootstrapping: archetypal analysis is repeated on multiple resampled datasets, archetype positions are aligned across replicates using Hungarian matching, and the positional variance of each archetype is computed. Low variance signals stable, well-defined archetypes, whereas high variance warns of over-parameterization. The team validated that ParTIpy’s results are highly concordant with the original MATLAB implementation across three distinct single-cell datasets.

To demonstrate the framework in action, the researchers applied ParTIpy to single-cell RNA sequencing data of hepatocytes sampled along the liver lobule’s porto-central axis. Three selection criteria converged on a four-archetype model, and pathway enrichment analysis using Reactome gene sets via the decoupler-py univariate linear model revealed distinct specializations: one archetype was associated with drug and xenobiotic metabolism, another with lipid metabolism. When the archetypal programs were projected onto spatial transcriptomics data, the metabolism archetype localized near the central vein while the lipid archetype enriched near the portal node, matching established liver zonation patterns and suggesting that spatial gradients of oxygen and metabolites drive this functional division of labor. The package further supports inference of archetype crosstalk networks by integrating archetype-specific expression profiles with curated ligand-receptor databases.

The second demonstration tackled a longstanding weakness of conventional cross-condition single-cell analysis. Standard approaches discretize cells into clusters and test for changes in cluster abundance, imposing sharp boundaries on what are often intrinsically continuous cell states. ParTIpy instead models cells as continuous mixtures of archetypal programs. Analyzing roughly 147,000 fibroblasts from sixteen non-failing and twenty-six cardiomyopathy donor hearts, the team fit a three-archetype model chosen for superior bootstrap stability. Donor-level pseudobulk profiles revealed significant shifts toward an activated fibrotic archetype in disease, marked by genes such as POSTN, FN1, and COL1A1 and by transcription factors including TWIST1 and SCX, while a quiescent, stress-responsive archetype was closer to healthy donors. Notably, even every non-failing heart contained a small fibroblast population near the fibrotic archetype, indicating that cardiomyopathy redistributes cells among existing programs rather than creating disease-specific states.

The researchers emphasize that ParTIpy’s scalability, downstream utilities, and interoperability make Pareto task inference accessible to the broader single-cell community and beyond, with tutorials covering spatial transcriptomics applications. Unlike deep archetypal analysis approaches that rely on encoder-decoder networks and amortized variational inference, which can introduce an amortization gap, ParTIpy achieves scale through coreset-based direct optimization while preserving archetypes as extremal points in the original trait space. Future integration with RNA velocity, pseudotime, and optimal transport methods could distinguish stable task specializations from transient states. The package, version 0.2.0, is freely available on PyPI and GitHub, with comprehensive documentation at partipy.readthedocs.io, positioning ParTIpy as a versatile new lens for exploring functional trade-offs in high-dimensional biological data.

Subject of Research: Scalable archetypal analysis and Pareto task inference for single-cell transcriptomic data

Article Title: ParTIpy: a scalable framework for archetypal analysis and Pareto task inference

Article References: ParTIpy: a scalable framework for archetypal analysis and Pareto task inference. (n.d.). https://doi.org/10.1038/s44320-026-00209-6

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00209-6

Keywords: ParTIpy, Pareto task inference, archetypal analysis, single-cell transcriptomics, coresets, trade-offs, cell state variability, hepatocyte zonation, cardiac fibroblasts, scverse, polytope geometry, open-source software