Simulation-based Benfordness estimation (SBBE) quantifies how closely a dataset conforms to Benford’s law, the characteristic distribution of the first significant digits observed in many naturally occurring datasets and widely used in data auditing. Applicable across scientific domains, SBBE provides uncertainty estimates and enables meaningful comparisons across datasets of different sizes. When applied to bioactivity data, SBBE ranks databases by curation quality and flags anomalous subsets for which targeted expert review is likely to yield the greatest benefit.

Tzanakakis et al. benchmark long-form genomic sequences generated by Evo 2 and megaDNA against natural genomes across complementary compositional and organizational features. Synthetic sequences remain detectably distinct, diverge progressively from their conditioning seeds, and reveal that correcting one artifact does not ensure broader genomic realism.

(Patterns 7, 101471; April 10, 2026)

Thanks to the generous sponsorship of Patterns and Cell Press, the University of Utah Scientific Computing and Imaging (SCI) Institute is pleased to offer the Patterns and Cell Press Early Career Trav...

No two neurons in the brain are identical. In this study, Santhosh and Narayanan show that differences between neurons affect learning, task performance, activity patterns, and resistance to disruption in highly varied ways. Using many recurrent neural networks trained on the same cognitive task, they found that the effects depended on network hyperparameters, the amount and type of variation, and how performance was measured. Their results highlight the importance of studying neural networks as complete, interacting systems across many examples, rather than focusing on individual differences in isolation.

In this review, Hasib et al. introduce EXHYTE a process centric survey to unify research on AI-assisted scientific discovery. Mapping recent studies reveals rapid advances in evidence retrieval, knowledge assembly, and idea generation, but it also reveals a persistent bottleneck in translating hypotheses into executable tests and using results to refine subsequent steps. The survey underscores the need for human oversight, reproducibility, validation, and auditable end-to-end discovery workflows.

Safety for large language models and artificial intelligence agents remains fragmented; defenses against jailbreaking coexist with unresolved hallucination, fairness, and autonomous-control problems. Chen et al. argue that these failures can interact and amplify each other, making safety a system property, rather than a collection of independent problems. They identify four safety challenge classes, ground their interactions in real-world incidents and a controlled clinical audit, and propose three design principles for integrated safety architectures.

This work proposes a conceptual framework to compare and contrast statistical and mathematical modeling using the notion of nonparametric partial identification from causal inference. The perspective adopted provides a set of assumptions that allow mathematical operations to be linked to causal claims, which are illustrated in a case study of pharmacodynamical modeling for a blood pressure medication.

As generative AI tools become routine parts of research, the journal has growing concerns about how this transformation will affect the reproducibility of research. We call upon our authors to take st...

Sahil et al. introduce unsupervised random forest (URF), a self-supervised feature-selection method that identifies informative structural descriptors from molecular dynamics simulations of proteins without requiring labels. A built-in learning coefficient lets URF assess its own training quality. Across ten protein systems, from folded enzymes to intrinsically disordered proteins, URF matches or outperforms supervised and deep learning baselines, recovers known functional residues, and improves downstream Markov state model construction, offering a general, interpretable, unsupervised framework for representation learning in biomolecular simulations.

(Patterns 6, 101345; September 12, 2025)

Researchers are rapidly adopting large language models (LLMs) in their work, but this can create challenges for scientific reproducibility. In this tutorial, the authors offer an organizing question—does the LLM sit on or off the data path of the published analysis? They then explain how this question, combined with familiar goals of reproducibility, provenance, determinism, and interpretability, can help researchers deliberately reason about their AI use and preserve best practices.

Shen et al. present mBoost, a diagnostic framework for building phenotype predictive models. mBoost evaluates model structure, feature redundancy, feature validity, and applicability to new data, helping researchers decide whether to use more complex models, do feature selection, or train a new model on new dataset.

(Patterns 6, 101208; March 14, 2025)

Reproducibility is widely considered a cornerstone of the scientific method and thus a key element of research, yet researchers often experience a wide variety of issues when attempting to reproduce empirical studies in AI. This work explores the relationship between the reproducibility of publications and the standards of their corresponding venues.

(Patterns 7, 101610; October 9, 2026)

Online exams make coordinated answer sharing difficult to detect, while existing collusion-detection methods typically focus on suspicious pairs of students and are often restricted to multiple-choice responses. Yan et al. introduce Q-SID, which detects likely collusion groups, including groups larger than two, directly from graded question scores, and provides estimated false-positive rates. Across diverse real exams, Q-SID improves accuracy and speed, giving instructors a scalable screening tool for identifying cases that warrant careful follow-up review.

YongWoo Lee and colleagues re-examine raw datasets from a recent crosslinking heat-activated purification (CLAP) study on the polycomb repressive complex 2 (PRC2). Their findings reveal that non-standard computational preprocessing suppressed vital signals, and correcting these workflow steps successfully reaffirms PRC2 as an RNA-binding protein.

Can AI systems be trusted when the model behind them keeps changing? Farajiamiri et al. evaluate 34 language models on text-based radiology questions and show that giving every model the same structured evidence makes their answers more consistent and more reproducibly correct. The gains come from synthesizing evidence, not retrieval alone, yet models still sometimes agree on the same wrong answer support.

Batch effects can obscure biological differences in single-cell transcriptomic studies, particularly when experimental batches are confounded with disease, treatment, or developmental stages. Sakaguchi et al. introduce Kanade, a variational autoencoder that separates batch effects from condition-associated signals. Kanade selectively removes batch effects while preserving biologically meaningful differences between conditions and produces corrected gene-expression values suitable for downstream analyses, thereby enabling more reliable comparisons across complex single-cell datasets.

BiomarkerKB addresses the fragmentation of biomarker information by integrating diverse biomarker resources into a standardized, computable knowledgebase. The resource combines over 200,000 biomarker-disease associations with rich contextual metadata and a knowledge graph accessible through a public web portal. BiomarkerKB supports data reuse, knowledge discovery, and the development of biomarker-driven research.

Foundation models generate embeddings that capture complex biological information, but interpreting and comparing these representations remains challenging. Rissom et al. introduce EmmaEmb, a model-agnostic framework for the analysis of embedding spaces. Applied to protein language models and high-stakes molecular biology and drug-discovery tasks, EmmaEmb ranks representations by task relevance, reveals patterns underlying prediction errors, and tracks representational changes during model adaptation, providing a quantitative approach to understanding learned biological representations. All methods are available as an open-source Python library.

Explore Patterns benchmarks studies through research articles, perspectives and expert opinions

Artificial intelligence has significantly expanded candidate generation in drug discovery, but selecting optimal compounds under real-world experimental constraints remains a major bottleneck. This review examines quantum annealing as a potential downstream optimization layer within hybrid computational workflows. By focusing on constraint-dominated decision tasks formulated as quadratic unconstrained binary optimization (QUBO) problems, the authors analyze how annealing architectures complement AI-generated scores. They further discuss practical hardware limitations, reporting considerations, and the critical role of rigorous benchmarking against classical baselines.

Single-cell data often capture continuous biological processes, yet common visualization tools can introduce misleading structure. This study systematically evaluates neighbor embedding methods using the predictability-computability-stability (PCS) framework, revealing key limitations in reliability. The authors introduce NESS, a stability-guided approach that improves the recovery of smooth biological patterns, enabling more trustworthy interpretation of cell-state dynamics across diverse datasets.

Liu et al. present a physics-induced deep-learning approach that extracts tissue composition information from routine single-energy CT scans. By incorporating X-ray imaging principles into model design, this approach improves material characterization and enhances performance across diverse clinical tasks. The framework expands the diagnostic value of standard CT imaging and supports more accessible precision medicine.

Biological systems rely on coordinated interaction networks between molecules and cells that can be disrupted by physiological challenges. NeDis is an open-source framework that measures how these networks change. By identifying groups of biomarkers with characteristic disruption patterns across conditions and time points, NeDis reveals system-level adaptations that are missed by standard analyses. Applications to immune data uncover coordinated disruptions during pregnancy and preeclampsia.

Lorenz, Engel et al. characterize the limited-sampling bias in partial information decomposition calculations. The bias is substantial when computed on empirical brain datasets, but the bias-correction methods developed here are demonstrated to work highly effectively to produce unbiased estimates of information synergy and redundancy.

Zhang et al. developed an interpretable machine learning framework to analyze population-level calcium ion signals in the suprachiasmatic nucleus (SCN), the central circadian clock of mammals. They uncover a hemispheric asymmetry in the SCN, particularly during the subjective day. Time prediction experiments further show that each hemisphere encodes differential time-encoding features. The work highlights hemispheric asymmetry as a fundamental property of the brain’s circadian clock.

Many problems in engineering, the sciences, and machine learning require optimizing a system with expensive evaluations. The authors present mlr3mbo, a rich R toolbox for Bayesian optimization (BO)—the standard methodology for such problems—that pairs a modular design of BO building blocks with well-performing, empirically chosen defaults and convenient entry points. On standard benchmarks, it performs on par with or better than HEBO, SMAC3, Ax, and Optuna.

Many biological measurements are bounded: a protein site cannot be less than unmodified or more than fully modified. Cobley shows that treating such data with ordinary linear tools can hide biological structure. Using cysteine oxidation as an example, Oxi-Shapes applies tropical geometry to reveal how aging reorganizes the redox proteome despite small average changes, offering a broader framework for analyzing bounded biochemical state spaces.

OscillomeR is a master-transcription-factor-anchored framework that infers ultradian gene oscillations from bulk and single-cell RNA sequencing. By reconstructing cyclic trajectories and phase-resolved gene expression, the method recovers Hes-driven oscillations in asynchronous muscle stem cells and pancreatic progenitors. It maps widespread rhythmic programs and phase shifts, revealing dynamic rewiring of gene-regulatory networks and helping to dissect the biological functions of gene-expression oscillations across development and disease.

Shao et al. introduce MedicalAgentsBench, a benchmark designed to separate genuine multi-step clinical reasoning from pattern matching. Using this benchmark, they compare internalized reasoning models against externalized agent frameworks and find that the two approaches are complementary; layering agents onto an internalized reasoning model achieves the highest accuracy, exceeding either approach alone.

BioMaster is a knowledge-guided multi-agent framework that automates bioinformatics workflow planning, execution, debugging, and validation. Across 49 tasks and 102 tools, it improves workflow completion over existing automated systems and supports both proprietary and open-source language models.

This survey reviews efficient attention mechanisms for scalable long-context language modeling, with a focus on linear and sparse attention. It summarizes their algorithmic foundations, hardware considerations, and integration into large-scale pre-trained language models, providing a systematic reference for designing efficient transformer-based architectures.

Multimodal spatial omics technologies enable mapping of molecular features within tissues, but integrating these diverse data remains challenging. This review outlines experimental strategies and computational frameworks, highlighting how combining modalities can reveal tissue organization and advance biological discovery, while also outlining key challenges and future directions.

Aligning AI models with human preferences remains a central challenge, yet the AI alignment field is increasingly divided between behavioral and representational approaches, resulting in narrowly scoped alignment strategies that are more vulnerable to failure. Drawing lessons from immunology and cybersecurity, the authors propose a set of design principles for integrated alignment frameworks that combine the complementary strengths of diverse alignment approaches. They emphasize the role of strategic diversity and outline steps toward greater coordination within the AI alignment research field.

Large language models now exhibit human-like abilities, raising a fundamental question: are the brain and AI similar only at the surface level, or do they share deeper computational principles? This perspective argues that both systems build predictive world models through prediction-error learning and reuse those models for sensory understanding and motor generation. This indicates that the brain and AI share computational principles underlying advanced cognitive functions and intelligence, despite the long-standing view that their internal processes are fundamentally distinct.

Community challenges drive benchmarking in computational biomedicine, but more than 2,000 such challenges are scattered across numerous platforms, with no central registry. OpenChallenges introduces a FAIR-aligned, open-source hub that aggregates a growing collection of challenges into a searchable registry with standardized metadata, ontology integration, and a transparent four-tier completeness score, transforming a fragmented competition ecosystem into persistent, reusable benchmarking infrastructure.

Tracking the same neurons across days or weeks is essential for understanding learning, stability, and plasticity in the brain but remains technically challenging in chronic electrophysiology due to probe drift and unit turnover. Huang et al. present DANT, a framework using density-based clustering for longitudinal neuron tracking that improves matching yield while maintaining accuracy across Neuropixels recordings in cortex and striatum during trained motor behaviors in freely moving rats.

Disentangling true biological signals from confounding factors remains a major challenge in single-cell genomics. Here, Møller and Madsen present DeepDive, a deep generative framework that separates covariate effects in single-cell chromatin accessibility data. DeepDive improves reconstruction accuracy, enables counterfactual predictions, and enhances the detection of differential accessibility under complex confounding. Applied to human pancreatic islets, it reveals disease-associated regulatory programs and links chromatin variation to genetic risk and beta-cell functional states.

Understanding how AI models interpret pathology images remains a major challenge. Here, the authors introduce CLEAR-HPV, an interpretable framework that identifies meaningful tissue patterns linked to human papillomavirus status directly from whole-slide images without manual annotation. By transforming complex model representations into compact, biologically meaningful concepts, the approach maintains predictive accuracy while improving transparency. This work highlights how interpretable AI can bridge prediction and biological understanding in computational pathology.