The development of new therapeutics and the validation of pathogenetic cancer mechanisms require representative laboratory models1,2. However, existing collections represent only a fraction of the diversity observed in human cancer2-4. Recent technologies have enabled efficient in vitro model derivation (for example, tumour organoids)5. However, whether these maintain essential properties of patient tumours during long-term expansion has not been systematically investigated. Here we present results of a large-scale international programme-the Human Cancer Models Initiative-which involved the generation of a resource of 665 next-generation models from 2,780 donors with 25 cancer types and integrated tumour-model whole genome, exome, methylome and transcriptome analyses. The resource provides 522 models with comprehensive clinical data, 153 models of rare cancers and 71 models from participants with non-European ancestry. Analyses of 421 matched tumour-model pairs reveal high genetic (97.8%) and epigenetic (95%) concordance and define correlates of model discordance. Single-nucleus RNA sequencing of tumour-model pairs reveals subsets of models in which culture conditions significantly influence cell states. Finally, we characterize model preservation of extrachromosomal DNA and post-treatment mutational signatures to provide opportunities to study therapeutic resistance. This model repository is being made available to the community-including multimodal molecular profiling, clinical information and integrative software tools-thus providing a valuable resource for preclinical investigation of cancer pathogenesis and treatment response.

Modern statistical and machine learning techniques are effective at describing, testing hypotheses and making predictions from complex data. This effectiveness is strongly influenced by the volume and heterogeneity of available data. In many fields, including much of biomedicine, large centralized datasets are not available because of cost, privacy, regulatory or other restrictions. In these cases, smaller datasets are distributed across a large number of independent sites. Medical record data is a classic example of this challenge: the total number of patients may be large, but their records are distributed across many health systems and cannot easily be centralized. Federated learning (FL) is a machine learning paradigm that enables training and validation of a shared model in settings of decentralized data. FL can improve model accuracy and generalizability by increasing sample size, but has trade-offs ranging from operational complexity to data-privacy risks to the potential to introduce unexpected imbalances in model accuracy. We outline ten tips for successfully and sustainably implementing FL for Biomedical applications, ensuring both ethical data governance and improved model performance in sensitive domains.

The continuous and reliable open access to curated biological data repositories is indispensable for accelerating rigorous scientific inquiry and fostering reproducible research outcomes. However, the current paradigm, which relies heavily on centralized infrastructure for the storage and distribution of foundational biomedical datasets, inherently introduces significant vulnerabilities. This centralized model is susceptible to single points of failure, including cyberattacks, technical malfunctions, natural disasters, and even political or funding uncertainties. Such disruptions can lead to widespread data unavailability, data loss, integrity compromises, and substantial delays in critical research, ultimately impeding scientific progress. The downstream effect of such interruptions can be the widespread paralysis of diverse research activities, including computational, clinical, molecular, and climate studies. This scenario vividly illustrates the inherent dangers of consolidating essential scientific resources within a single geopolitical or institutional locus. As data generation is accelerating and the global landscape continues to fluctuate, the sustainability of centralized models must be critically re-evaluated. A shift toward federated and decentralized architectures may offer a robust and forward-looking approach to enhancing the resilience of scientific data infrastructures by reducing exposure to governance instability, infrastructural fragility, and funding volatility, while also promoting equity and global accessibility. Inspired by established models such as ELIXIR’s federated infrastructure and the policy and funding frameworks developed by CODATA and the Global Biodata Coalition (GBC), emerging Decentralized Science (DeSci) initiatives can contribute to building more resilient, fair, and incentive-aligned data ecosystems. The future of open science depends on integrating these complementary approaches to establish a globally distributed, economically sustainable, and institutionally robust infrastructure that safeguards scientific data as a public good, further ensuring continued accessibility, interoperability, and preservation for generations to come. Here, we examine the structural limitations of centralized repositories, evaluate federated and decentralized models, and propose a hybrid framework for resilient, fair, and sustainable scientific data stewardship.

RNA sequencing (RNA-seq) has emerged as an exemplary technology in biology and clinical applications, offering a crucial complement to other transcriptomic profiling protocols due to its high sensitivity, precision, and accuracy in characterizing transcriptomes. However, the rapid proliferation of RNA-seq tools necessitates the adoption of robust software development practices. Such development underscores the critical need to examine how RNA-seq tools are developed, maintained, and distributed; and whether the data they generate is reproducible as all of these factors are essential for ensuring software reliability, transparency, and trust in scientific findings. We conducted a comprehensive assessment of 434 RNA-seq tools developed between 2008 and 2024, categorizing them based on the type of analysis they perform. Our evaluation encompassed their software development and distribution methodologies, as well as the attributes contributing to their widespread adoption and dependability within the biomedical community, which were quantified by factors such as package manager availability, containerization, multithreading support, documentation quality, and inclusion of example datasets. Our findings establish the first documented positive association between rigorous software development practices and their adoption of published RNA-seq tools as measured by citations (Mann-Whitney U test, p-value = 4.9e-26). By identifying key characteristics of widely adopted software, our findings guide developing robust and user-friendly RNA-seq tools, thereby reinforcing the call for rigorous community-wide standards.

Traditional gene expression deconvolution methods assess a limited number of cell types, therefore do not capture the full complexity of the tumor microenvironment (TME). Here, we integrate nine deconvolution tools to assess 79 TME cell types in 10,592 tumors across 33 different cancer types, creating the most comprehensive analysis of the TME. In total, we found 41 patterns of immune infiltration and stroma profiles, identifying heterogeneous yet unique TME portraits for each cancer and several new findings. Our findings indicate that leukocytes play a major role in distinguishing various tumor types, and that a shared immune-rich TME cluster predicts better survival in bladder cancer for luminal and basal squamous subtypes, as well as in melanoma for RAS-hotspot subtypes. Our detailed deconvolution and mutational correlation analyses uncover 35 therapeutic target and candidate response biomarkers hypotheses (including CASP8 and RAS pathway genes).

As genes tend to be co-regulated as gene modules, feature selection in machine learning (ML) on gene expression data can be challenged by the complexity of gene regulation. Here, we present a protocol for reconciling differences in classifier features identified using different ML approaches. We describe steps for loading the PathwaySpace R package, preparing input for analysis, and creating density plots of gene sets. We then detail procedures for testing whether apparently distinct feature sets are related in pathway space. For complete details on the use and execution of this protocol, please refer to Ellrott et al.1.

We present a protocol for obtaining cancer type and subtype predictions using a machine learning method (subSCOPE). We describe steps for data preparation, subSCOPE setup, and running subSCOPE inference on prepared data. The protocol supports five -omics data types as input (DNA methylation, gene expression, microRNA [miRNA] expression, point mutations, and copy-number variants) and allows individual cancer type and data type selection. For non-The Cancer Genome Atlas (TCGA) cancer samples, it provides subtype-level classification across 26 different TCGA cancer cohorts and 106 subtypes. For complete details on the use and execution of this protocol, please refer to Ellrott et al.1.

AI based embeddings offer the possibilities of encoding complex biological data into low dimensional spaces, called embedding spaces, that maintain the relationships between entities. There is an open question about the compatibility of embedding spaces that are created without any coordination. It has been assumed that signals in these unaligned embedding spaces would be destroyed if vectors were aggregated into summed values. We trained embedding models across different data modalities and tested aggregating the values together to test this assumption. Our research shows that signal from unaligned embedded values is conserved and able to still be used for learning tasks, such as data modality and tumor of origin recognition.