A multi-modal foundation model for brain disease diagnosis and medical imaging Accurate diagnosis of brain disorders requires combining medical images with clinical knowledge. Brainfound is a large multi-modal AI model trained on millions of brain CT and MRI images paired with reports. By integrating image understanding, text reasoning, and image generation, Brainfound supports diagnosis, report writing, and clinical dialogue, demonstrating how foundation models can advance human-in-the-loop brain health care.

Online Now: A multi-modal foundation model for brain disease diagnosis and medical imaging #datascience

Are diffusion models ready for materials discovery in unexplored chemical space? In this study, the authors benchmark two leading crystal diffusion models, MatterGen and DiffCSP, to evaluate whether they can recover ground-state (or near-ground-state) structures for materials in chemical spaces that were not included in their training data. The results reveal performance limitations in under-sampled elemental groups and in structures with larger numbers of atoms. The authors develop a theory to explain the observed failures and outline a path toward developing dependable generative models for materials discovery.

Online Now: Are diffusion models ready for materials discovery in unexplored chemical space? #datascience

A framework for reproducibly managing coupled research software and data assets based on shared transformation functions This study refines DataDesc, a metadata-driven approach for coupling research software and data into reproducible workflows. Using automated data model comparison and the ioProc workflow manager to design reusable transformation functions, it identifies required data conversions and ensures transparent documentation, advancing interoperability and FAIR-compliant research across scientific domains.

Online Now: A framework for reproducibly managing coupled research software and data assets based on shared transformation functions #datascience

Check out also the short Opinion from members of CompMotifs group (www.compmotifs.com) about their efforts to tackle shared challenges and support reproducible computational research.
www.cell.com/patterns/ful...

This month's cover image, related to the study by Kuzikov et al., shows a representative cellular phenotype of drug-induced phospholipidosis. The effect is characterized by the excessive accumulation of phospholipids within cells that also can be induced by cationic amphiphilic drugs. Using the shown high-content imaging, the drug-induced phospholipidosis phenotype was characterized, and over 5,000 repurposed drugs were annotated for the ability to induce phospholipidosis. The results were then used to understand which compounds are likely to induce phospholipidosis based on a drug’s chemical structure, finally resulting in the development of a machine learning model capable of predicting the risk of phospholipidosis. Image provided by Mariya Pereverzeva.

Our April issue is now live!
www.cell.com/patterns/iss...

On the cover this month is an image of cells experiencing phospholipidosis, highlighting the paper by Kuzikov et al that develops a phospholipidosis risk prediction model to advance drug screening
www.cell.com/patterns/ful...

The things that I find most interesting about this paper are the differences between the tested models. The OpenAI and Claude models tend to perform similarly and have a common weakness in "no-evidence" contexts, i.e. they like to make up an answer even if there isn't medical evidence.

We are hiring!

Are you a researcher in #datascience, machine learning, artificial intelligence, statistics, or a related field? Are you passionate about #openscience and clear scientific communication? If yes, please check out this opportunity:

www.linkedin.com/jobs/view/43...

Dendritic nonlinearities mitigate communication costs Artificial neural networks typically ignore the complex, active branches found in biological neurons. This study challenges the assumption that these “dendritic nonlinearities” primarily boost computa...

🧠Our latest collaborative work, led by Xundong Wu and published in @cp-patterns.bsky.social, shows that adding dendritic nonlinearities to ANNs slashes communication costs. This is a key for the next generation of sustainable AI systems.
Full paper here: 🔗 www.cell.com/patterns/ful...

Helix 1.0: An open-source framework for reproducible and interpretable machine learning on tabular scientific data Helix is an open-source, FAIR-by-design analytics framework to address the growing need for transparency, interpretability, and reproducibility in data-driven scientific research. Designed for tabular data, Helix integrates the full analytical life cycle within a single modular, extensible environment. Helix places experimental provenance and human interpretability at its core, ensuring that modeling decisions, data transformations, and analytical outcomes are fully traceable and accessible. Its lightweight, browser-based interface lowers technical barriers for domain scientists while preserving methodological rigor and flexibility.

Online Now: Helix 1.0: An open-source framework for reproducible and interpretable machine learning on tabular scientific data #datascience

Zero-shot reconstruction of mutant spatial transcriptomes Measuring spatial transcriptomes in mutant tissues is costly and technically challenging. This study proposes ZENomix, a zero-shot framework that predicts mutant spatial transcriptomes without mutant-specific spatial data, using only wild-type spatial data as side information. ZENomix recovers spatial gene expression patterns across disease models and identifies previously unknown spatially disrupted genes.

Online Now: Zero-shot reconstruction of mutant spatial transcriptomes #datascience

Dendritic nonlinearities mitigate communication costs Artificial neural networks typically ignore the complex, active branches found in biological neurons. This study challenges the assumption that these “dendritic nonlinearities” primarily boost computational power. Instead, the authors show they function like efficient data filters. By processing signals locally before transmission, these nonlinear structures drastically reduce the “traffic” required between neurons. This reveals a biological blueprint for scaling AI by solving communication bottlenecks rather than just increasing calculation capacity.

Online Now: Dendritic nonlinearities mitigate communication costs #datascience

Evaluating large language models for evidence-based clinical question answering Can AI systems reliably navigate the complexity and uncertainty of real-world medical evidence? Wang and Chen evaluate leading AI models on 20,000 questions synthesized from over 8,000 systematic reviews and clinical guidelines, finding that models struggle when underlying studies show high effect-size variance or limited citation support.

Online Now: Evaluating large language models for evidence-based clinical question answering #datascience

The AI risk repository: A meta-review, database, and taxonomy of risks from artificial intelligence Slattery et al. systematically analyzed 74 AI risk frameworks containing 1,725 distinct risks to create the AI Risk Repository. They developed two complementary taxonomies—a Causal Taxonomy (classifying risks by entity, intent, and timing) and a Domain Taxonomy (organizing risks across seven societal impact areas)—providing a unified foundation for AI risk assessment, governance, and auditing.

Online Now: The AI risk repository: A meta-review, database, and taxonomy of risks from artificial intelligence #datascience

Data-driven deformation correction in X-ray spectro-tomography with implicit neural networks Wang et al. introduce CANet, a self-supervised, coordinate-based neural network designed to correct complex deformations in X-ray spectro-tomography. By implicitly modeling deformation fields without external training data, CANet achieves robust alignment across both tomographic and spectral dimensions. This framework significantly enhances the fidelity of 3D chemical mapping, enabling the precise visualization of nanoscale degradation mechanisms in battery materials.

Online Now: Data-driven deformation correction in X-ray spectro-tomography with implicit neural networks #datascience

Scalable data harmonization for single-cell image-based profiling with CytoTable High-content imaging enables large-scale phenotypic discovery, but its biological value is often limited by fragmented single-cell data organization. CytoTable addresses this challenge by providing a standardized, harmonized foundation emphasizing consistent structure, explicit data types, and fast modular integration. This approach reduces technical artifacts, improves reproducibility, and allows researchers to focus on biological questions, supporting reliable pattern discovery and collaborative analysis as microscopy datasets grow in scale and complexity.

Online Now: Scalable data harmonization for single-cell image-based profiling with CytoTable #datascience

Online Now: VoxelCoder: Classification of human cellular phenotypes via autoencoder batch alignment and hyperdimensional representation of cytometry data #datascience

Sample size calculation for training ensemble machine learning models on health data Most clinical prediction modeling studies using machine learning models do not estimate sample size. A conceptual model for thinking about power for machine learning models is presented in this study, and an estimator has been fitted for sample size calculations during study design.

Online Now: Sample size calculation for training ensemble machine learning models on health data #datascience

Just a friendly reminder: we have an open call for papers that critically reanalyze prior publications. Learn more here:

www.cell.com/patterns/spe...

#datascience #openscience

Modeling of longitudinal immune profiles reveals distinct immunogenic signatures following five COVID-19 vaccinations among people living with HIV. In the study by Korosec et al., researchers generate synthetic data trajectories (“virtual patients”) that preserve the statistical properties of real immune measurements by using data-driven models. The illustration, by Andrée Fournier (National Research Council Canada), explores the intersection of human biology and computational modeling. In the foreground, a softly rendered human hand reaches forward, symbolizing real patients and empirical data. Emerging from the background, a semi-transparent wireframe hand, constructed from a digital mesh, represents the virtual patients and the models’ aim to capture the essential features of true biological responses.

Our March issue is now live!
www.cell.com/patterns/iss...

This month's cover is a wonderful image from Andrée Fournier highlighting a paper by Korosec et al where researchers generate “virtual patients” to explore the complexity of human immune responses
Full study: www.cell.com/patterns/ful...

First page of article "Unveiling gender disparities in corporate board career paths using deep learning"

Study finds women rely on professional and social networking more than men do in order to advance their careers. spkl.io/63322AI6JM

@cribravo.bsky.social & colleagues
@cp-patterns.bsky.social

Unveiling gender disparities in corporate board career paths using deep learning Gender gaps in corporate boardrooms are shaped not only by hiring decisions but also by the structure of professional networks. Using 23 years of career histories and network data on thousands of executives, we show in this study that women must hold broader and more central positions than men to reach similar board roles. Connections to other female leaders are especially important. These findings illustrate how everyday networking practices can help reduce or reinforce inequality in corporate leadership.

Online Now: Unveiling gender disparities in corporate board career paths using deep learning #datascience

First page of Opion piece.

"A call to join a collective effort on AI evaluation" spkl.io/63329Axcjd

Irving Torres and the spkl.io/63323Axcj9 Consortium
@cp-patterns.bsky.social

A call to join a collective effort on AI evaluation AI evaluations increasingly shape deployment, governance, and trust, but expectations for how they are conducted and reported remain fragmented. We introduce a cross-sector Delphi process to develop community-endorsed guidance for AI evaluation practice and invite researchers, practitioners, ethics organizations, and institutions to participate.

Online Now: A call to join a collective effort on AI evaluation #datascience

Modeling of longitudinal immune profiles reveals distinct immunogenic signatures following five COVID-19 vaccinations among people living with HIV Using machine learning to map immune interdependencies, this study deploys random forests (RFs) to compare vaccine-induced immunity in people with HIV versus HIV-negative age-matched controls across up to five SARS-CoV-2 doses. RFs reveal distinct saliva- and blood-based humoral patterns and a subset of subjects whose responses resembled controls, signaling partial immunologic restoration. Privacy-preserving synthetic patients enable RF training on synthetic data that generalize to real individuals, supporting precision-guided vaccination and follow-up.

Online Now: Modeling of longitudinal immune profiles reveals distinct immunogenic signatures following five COVID-19 vaccinations among people living with HIV #datascience

TweetyBERT: Automated parsing of birdsong through self-supervised machine learning Parsing birdsong into behavioral units typically requires human-labeled data or pre-segmented audio. TweetyBERT, a self-supervised transformer, overcomes these limitations by learning directly from raw spectrograms. Operating at a 2.7 ms temporal resolution, the model preserves a one-to-one correspondence between input time bins and latent states. Applied to canary song, TweetyBERT autonomously discovers syllable-level representations that align closely with expert annotations, enabling large-scale automated labeling of vocal sequences with minimal human intervention.

Online Now: TweetyBERT: Automated parsing of birdsong through self-supervised machine learning #datascience

Promoting sustainable human mobility for income segregation mitigation Despite extensive studies that have addressed the quantification of income segregation, its impact on human mobility remains unclear. This study introduces a segregation visitation index and a mobility prediction model to reveal biased travel patterns, showing how segregation and mobility reinforce inequalities and guide more inclusive, sustainable urban planning.

Online Now: Promoting sustainable human mobility for income segregation mitigation #datascience

DestinyNet: A deep-learning framework for cell-fate analysis from lineage-tracing single-cell RNA sequencing data DestinyNet is a deep-learning framework for analyzing lineage-tracing single-cell sequencing data that integrates fate clustering, fate-flow inference, and fate prediction. It is robust to barcode sparsity and experimental variability, providing a unified approach for interpreting cell-fate dynamics.

Online Now: DestinyNet: A deep-learning framework for cell-fate analysis from lineage-tracing single-cell RNA sequencing data #datascience

Cell Press: A dedication to quality. www.cell.com/about?utm_so...

Creating strong predictive models in oncology Many oncology predictive models fail to improve care. Issues include risks of bias, underpowered radiomics studies, and limited clinical impact. A path forward involves an emphasis on clinically actio...

Also in this issue, check out the Opinion from Michael Gensheimer, who argues that predictive model research in oncology must focus more on generalizable performance and clinical usefulness if patient care is to be improved.
www.cell.com/patterns/ful...

On the cover: Inspired by Jia et al. in this issue, the cover visualises how incomplete or incorrect segmentation labels can be filled in with information from the image itself. The grid marks are where labels are absent, while the completed shapes illustrate how visual context helps recover those regions. The robotic hand represents an automated process that reduces the need for manual relabelling. Overall, the image shows how combining images with partial labels can produce more reliable training data for semantic segmentation. Image credit: Phillip Krzeminski.

Our February issue is now live!
www.cell.com/patterns/iss...

On the cover this month, we are highlighting a work from Jia et al that describes a method to fix defective annotations in image libraries and thereby advance computer vision research.
www.cell.com/patterns/ful...