The core challenge any model of genetic variation must confront is linkage disequilibrium, the nonrandom association of variants across the genome. Neighboring SNPs, or single nucleotide polymorphisms, tend to be inherited together, but so do many distant ones, and these long-range correlations carry much of the structure that makes a population recognizable in the data. Classical simulators built on the coalescent model capture these patterns by tracing observed variation back to latent genealogies shaped by demographic history, mutation, and recombination. That approach is expressive but computationally punishing, which is why practical tools such as msprime rely on Markovian approximations. A second tradition, descending from the product-of-approximate-conditionals framework of Li and Stephens, dispenses with explicit genealogies and fits the data distribution directly, naturally yielding hidden Markov models. HMMs have been enormously successful in haplotype phasing, genotype imputation, and ancestry inference, but their chain structure forces information between distant SNPs to pass through every intermediate hidden variable, weakening long-range dependencies at each step.

Deep learning promised a way out, and a wave of generative adversarial networks, variational autoencoders, restricted Boltzmann machines, and most recently diffusion models has been applied to genetic data. These models can produce artificial genomes that look plausible in principal component analyses and allele frequency summaries. Yet the UCLA-led team argues they carry structural liabilities for genomics. Generative adversarial networks do not define a probability distribution over the data at all, making likelihood-based inference impossible. Restricted Boltzmann machines define one, but evaluating it requires computing a partition function over an exponentially large space of configurations. Variational autoencoders expose only a lower bound on the marginal likelihood. Diffusion models can in principle compute exact likelihoods but scale poorly to dense SNP data and have not been evaluated on imputation. Crucially, none of these approaches supports efficient estimation of conditional probabilities, the mathematical operation at the heart of genotype imputation, and none offers an objective way to judge whether training has converged.

GPC’s answer is architectural. The model builds on hidden Chow-Liu trees, a class of latent variable models in which every observed SNP is paired with a hidden variable, and the hidden variables are wired together not in a chain but in a tree learned from the data itself. The Chow-Liu algorithm, dating to 1968, computes pairwise mutual information between all variables and connects them in the maximum-weight spanning tree, placing strongly correlated SNPs adjacent to one another regardless of their positions along the chromosome. Where a hidden Markov model would route the dependence between two distant but correlated SNPs through every latent variable in between, losing correlation at each hop, the learned tree carries it over a single edge. In the tree learned from a 10,000-SNP region of chromosome 15 in the 1000 Genomes dataset, the researchers found 4,599 leaf nodes, 397 nodes with degree greater than two, one hub connected to 161 other SNPs, and 18.3 percent of edges spanning more than 1,000 positions apart, a topology fundamentally unlike the unbranched chain of an HMM.

Expressiveness alone would be useless if the model were computationally intractable, and this is where probabilistic circuits enter. A probabilistic circuit represents a joint distribution as a directed acyclic graph of input, sum, and product nodes, with product nodes encoding factorizations and sum nodes encoding weighted mixtures. When the graph satisfies two structural properties, smoothness and decomposability, any marginal probability can be computed exactly in time linear in the circuit’s size. Conditional probabilities, the quantities needed for imputation, follow as simple ratios of two marginal queries, each evaluated in a single feedforward pass. Sampling artificial genomes proceeds by ancestral traversal of the circuit, also linear in size. By compiling hidden Chow-Liu trees into this circuit representation and training them with expectation-maximization on GPUs using the PyJuice package, the team trained models with more than 88 million parameters, with each EM epoch taking under two seconds on the 1000 Genomes data and full training completing in roughly two to six hours on a single NVIDIA RTX A5000 graphics card.

The practical payoff shows up most clearly in genotype imputation, the task of inferring variants a person was not directly genotyped, which underlies much of genome-wide association science. The researchers evaluated GPC across three settings using the 1000 Genomes Project, the UK Biobank, and a high-coverage 1000 Genomes release. In the general setting, where models trained on diverse ancestries impute into similarly diverse test genomes, GPC used for direct conditional imputation achieved a 27.5 percent average improvement in the squared correlation between imputed and true genotypes over the next best method, a restricted Boltzmann machine, and a striking 174 percent improvement for low-frequency variants with minor allele frequency below one percent. Direct imputation through the model itself also beat GPC’s own artificial genomes used as reference panels for the standard tool Impute5, by about 15.5 percent overall, apparently because it targets the imputation objective directly rather than injecting the noise of an intermediate simulation step.

The gains were largest precisely where existing infrastructure is weakest: populations underrepresented in public reference panels. Because large public datasets are overwhelmingly of European ancestry, imputation into African, admixed, or other non-European populations suffers from distributional mismatch. In population-specific experiments, GPC’s direct imputation improved on the next best deep generative method by 33 percent on average, and by 279 percent for low-frequency variants. Remarkably, in the 1000 Genomes experiments GPC trained on ancestry-matched private data outperformed Impute5 running on real European reference genomes, achieving a 12.3 percent average improvement in squared correlation and a 42.1 percent improvement for rare variants. In a realistic array-based scenario, imputing 12,551 SNPs missing from the HumanOmni5Exome genotyping array, GPC outperformed the European-panel Impute5 benchmark by 96.5 percent on average and by more than twelvefold for low-frequency variants. The entire held-out array imputation took about one second in a single conditional query.

Privacy, the original motivation for artificial genomes, also fared better under GPC. Using the nearest-neighbor adversarial accuracy metric, which scores values near 0.5 as the ideal balance between utility and privacy, GPC came closest to the ideal across nearly all splits of both datasets. The failure modes of the competitors were instructive. Restricted Boltzmann machines produced synthetic haplotypes that each sat closer to some individual real genome than to any other synthetic one, meaning samples clustered around training individuals, a memorization pattern that could allow synthetic genomes to be traced back to specific people. The Wasserstein generative adversarial network showed the opposite pathology, with artificial genomes occupying a region largely disjoint from the real data, sacrificing utility. The hidden Markov model failed in the most extreme form, producing genomes perfectly separable from real ones in both directions, achieving maximal privacy by failing to model the data at all. GPC’s artificial genomes were not perfectly indistinguishable, and the authors caution that improved performance on this metric does not guarantee protection against membership inference or attribute inference attacks.

Limitations remain, and the authors are candid about them. GPC currently operates on single genomic regions of roughly 10,000 to 15,000 SNPs, bounded by the memory cost of constructing the Chow-Liu tree, which grows with the square of the SNP count; reaching genome-wide scale will require hierarchical or distributed approaches. The present implementation handles haploid rather than diploid genomes, and like all generative models GPC inherits biases from its training data, leaving fairness across diverse cohorts an open problem. Still, the study marks a notable convergence of two research traditions that have mostly run in parallel: the tractable probabilistic models beloved of population geneticists and the expressive deep architectures driving modern machine learning. By learning the dependence structure of the genome directly from mutual information and then compiling it into a form where exact inference is cheap, GPC suggests that the trade-off between realism, computational tractability, and privacy in synthetic genomic data may be less inevitable than it once appeared. Code and experiments are publicly available as the field moves toward equitable genomic tools for populations long left out of reference panels.

Subject of Research: A tractable deep generative model, Genetic Probabilistic Circuits, for modeling human genetic variation data

Article Title: GPC: An expressive and tractable deep generative model for genetic variation data

Article References: Anand, P., Liu, A., Dang, M., Fu, B., Wei, X., Van den Broeck, G., & Sankararaman, S. (2026). GPC: An expressive and tractable deep generative model for genetic variation data. PLOS Genetics, 22(10), e1012321. https://doi.org/10.1371/journal.pgen.1012321

Image Credits: AI Generated

DOI: 10.1371/journal.pgen.1012321

Keywords: genetic probabilistic circuits, artificial genomes, genotype imputation, linkage disequilibrium, probabilistic circuits, hidden Chow-Liu trees, population genetics, deep generative models, genomic privacy, 1000 Genomes Project, UK Biobank, underrepresented populations

The problem with post hoc explanations is well documented in the machine learning literature. Methods such as LIME, SHAP, and Grad-CAM build simplified surrogate models that approximate a trained network’s behavior, but studies by Alvarez-Melis and Jaakkola, Laugel and colleagues, Slack and colleagues, and Cynthia Rudin have shown that these surrogates can diverge sharply from the underlying model. An explanation model may match the black box’s accuracy while relying on entirely different features, producing explanations that look plausible and are simply wrong. In high-stakes domains such as medical diagnosis and legal decision-making, the researchers argue, such misleading accounts are not merely inconvenient but dangerous. Rudin’s influential 2019 position paper called for abandoning the explain-the-black-box paradigm altogether in favor of models that are interpretable from the ground up, with every component designed for a distinct, human-understandable purpose.

NPCs answer that call by building on concept bottleneck models, an architecture introduced by Pang Wei Koh and colleagues in 2020. Concept bottleneck models split prediction into two stages: a neural network first recognizes human-understandable concepts such as red color or round shape, and a simple predictor then maps those concepts to a final class. The difficulty is that performance and interpretability pull in opposite directions. Variants that replace concept probabilities with high-dimensional concept embeddings, such as the Concept Embedding Model of Zarlenga and colleagues, boost accuracy but destroy interpretability, because the individual dimensions of an embedding carry no semantic meaning. Conversely, architectures that embed logical rules into the predictor improve transparency but typically learn those rules from data alone, leaving no room for human domain knowledge, and they come with no theoretical guarantee linking overall performance to the quality of individual modules.

The new architecture addresses both gaps with two modules working in tandem. The first is an attribute recognition model, a neural network that inspects an input image and outputs probability vectors for a set of high-level categorical attributes. Instead of treating every binary concept independently, NPCs group concepts into attributes: color, shape, symbol, and text, for example, in the case of traffic signs. Each probability vector expresses the likelihood of the various values that attribute can take. The second module is a task predictor built on a probabilistic circuit, a type of tractable graphical model represented as a rooted directed acyclic graph whose internal nodes are weighted sums and products. Because such circuits satisfy structural properties called smoothness and decomposability, any joint, marginal, or conditional probability over attributes and classes can be computed in at most two passes through the graph, in time linear in its size. The circuit embeds logical rules, either learned implicitly from data or supplied explicitly by humans, within its very structure and parameters.

Formally, the model rests on two mild assumptions. The first, sufficient attributes, states that the class label is conditionally independent of the raw input once the attributes are known; for a traffic sign, knowing the shape, color, symbol, and text is enough to identify the class without consulting the pixels again. The second, complete information, states that the attributes are mutually independent given the input, since the image fully encodes them. Under these assumptions, the probability that an image belongs to a class is computed by summing, over every combination of attribute values, the product of the likelihood of that combination and the circuit’s conditional probability of the class given those attributes. The predicted class is simply the one with the highest score. This compositional formula is what makes the system interpretable: every term in the sum corresponds to a human-readable attribute assignment and an explicit rule linking attributes to classes.

Training proceeds in three stages. First, the attribute recognition model is trained in a multi-task learning framework, with one task per attribute and cross-entropy losses weighted by the size of each attribute’s value space so that no single attribute dominates. Second, the probabilistic circuit is constructed in one of two ways. In the data-driven approach, the LearnSPN algorithm recursively clusters variables and data to learn a circuit structure, and maximum likelihood estimation via the CCCP algorithm, which converges monotonically, optimizes its weights. In the knowledge-injected approach, humans supply weighted logical rules, such as the conjunction of circle, red, and slash indicating a no-entry sign, and the circuit is built as a two-layer sum-of-products formula whose edge weights equal the frequency of each rule. The authors prove that this manually constructed circuit exactly models the empirical joint distribution over attributes and classes. In the third stage, the two modules are fine-tuned jointly end to end, using stochastic gradient descent for the neural network and projected gradient descent to keep the circuit weights positive.

A central theoretical contribution is a compositional error bound, which the authors believe is the first performance guarantee of its kind for compositional bottleneck models. Under the two assumptions, the total error of an NPC, measured as the expected total variation distance between the learned and true class distributions, is upper-bounded by the sum of the errors of the individual attribute recognizers plus twice the error of the probabilistic circuit. In plain terms, improving the recognition of any single attribute reliably improves the whole system, and the overall error can be traced to specific, measurable components. This stands in sharp contrast to end-to-end networks, where a change in one layer can have unpredictable effects elsewhere, and it gives practitioners a principled way to decide where to invest in better modules.

Interpretability is delivered through two complementary explanation types. The Most Probable Explanation identifies the attribute assignment that contributes most to a predicted class, answering the question of why the model chose what it did. The authors introduce an alignment rate, the proportion of correct predictions whose most probable explanation matches the ground-truth attributes, as a reliability metric; across most of their experiments this rate approaches one hundred percent, indicating the model reasons much as a human would. Counterfactual Explanations work in the opposite direction: when the model errs, a projected gradient ascent procedure searches for minimal adjustments to the attribute probabilities that would flip the prediction to the correct class. In one striking example, the model misread the text on a traffic sign as 120 and predicted a speed limit of 120; the counterfactual showed that reading the text as 80, even with only about fifty percent confidence, would have produced the correct classification, confirming that the model’s arithmetic and reasoning chain were sound and only one attribute was at fault.

Empirically, the team evaluated NPCs on four benchmarks: MNIST-Addition, where two digit images must be summed; the German Traffic Sign Recognition Benchmark with 39,209 images; CelebA, a large face dataset; and AwA2, containing 37,322 animal images. NPCs outperformed representative concept-based baselines, including the original concept bottleneck model, its hybrid variant, the Concept Embedding Model, and the Deep Concept Reasoner, and remained competitive with, and on two datasets superior to, fully end-to-end black-box networks, all without a single uninterpretable component. Ablation studies showed that grouping concepts into attributes preserves relational constraints and improves recognition accuracy, that removing decisive attributes such as symbol or text severely degrades performance while removing non-decisive ones like color barely matters, and that data-driven circuits excel on complex datasets while knowledge-injected circuits suffice for simpler ones. Human interventions on attribute values also improved accuracy in line with expectations, and when all attributes were corrected the models reached perfect accuracy on the two simpler benchmarks.

The authors are candid about limitations. The attribute recognition network itself remains a black box vulnerable to spurious correlations, the conditional independence assumption can introduce reasoning shortcuts when confounders are present, and the computational cost of enumerating attribute combinations grows exponentially with the number of attributes. They sketch remedies ranging from large language model generated concept annotations to tensorized circuit implementations such as Einsum Networks. Even so, the message of the work is clear: the long-assumed trade-off between accuracy and transparency is not an iron law. By composing neural perception with tractable probabilistic reasoning, Neural Probabilistic Circuits show that machines can be made to show their work, a capability that may prove essential wherever automated decisions must be trusted, audited, and defended.

Subject of Research: An interpretable machine learning architecture combining neural attribute recognition with probabilistic circuits for logical reasoning

Article Title: Neural Probabilistic Circuits: Enabling Compositional and Interpretable Predictions Through Logical Reasoning

Article References: Chen, W., Yu, S., Shao, H., Sha, L., & Zhao, H. (2026). Neural Probabilistic Circuits: Enabling Compositional and Interpretable Predictions Through Logical Reasoning. Machine Learning, 115(9), Article 207. https://doi.org/10.1007/s10994-026-07118-7

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07118-7

Keywords: machine learning, interpretability, probabilistic circuits, concept bottleneck models, neuro-symbolic AI, logical reasoning, explainable AI, deep learning, image classification, counterfactual explanations, error bounds, graphical models