GLEAM：让高质量机器学习在癌症研究中普及化

Machine learning (ML) has revolutionized biomedical research, enabling accurate prediction from large and often high-dimensional datasets through statistical modeling and pattern recognition. Examples of ML applications include predicting patient responses to cancer therapies and classifying cancer lesions. However, the lack of accessible, well-structured guidance on how to develop, validate, and communicate ML models continues to limit their routine and effective use in research. To address this gap, we introduce the second-generation ML toolkit for the Galaxy platform, a widely used bioinformatics workbench. Named Galaxy Learning and Modeling (GLEAM), it is built around four core principles that target critical barriers in cancer research: no-code accessibility through a web interface; adherence to best practices via standardized setups and reports; reproducibility by recording parameters, data versions, environments, and workflows; and scalability through native Galaxy integration. Our goal is to provide end-to-end ML tools that enable researchers to configure, run, evaluate, and share publication-ready analyses. GLEAM integrates three learner tools: (1) Tabular Learner for structured data, (2) Image Learner for pixel-based inputs, and (3) Multimodal Learner for heterogeneous data. The AutoML backends are PyCaret, Ludwig, and AutoGluon, respectively. Users supply datasets, labels, and optional metadata, and GLEAM automates preprocessing, feature extraction, algorithm selection, hyperparameter tuning, and cross-validation. As a proof of concept, we reproduced results from published studies using GLEAM tools by running their publicly available datasets and comparing model metrics to benchmark tool quality. On the HAM10000 dermoscopy dataset, the Image Learner, using a deep-learning model, achieved accuracy 0.86 (vs. 0.86 in published models), precision 0.85 (vs. 0.88), recall 0.85 (vs. 0.85), and F1 score 0.85 (vs. 0.86), matching state-of-the-art outcomes for skin lesion classification on this dataset. On the LORIS immunotherapy-response dataset, the Tabular Learner achieved accuracy 0.80 (vs. 0.70), AUC 0.76 (vs. 0.75), and PR-AUC 0.55 (vs. 0.56), demonstrating robust prediction of treatment response. Using the HANCOCK dataset to predict survival in head and neck cancer, the Multimodal Learner achieved an ROC AUC of 0.74 (vs. 0.79). This modest decrease reflects a different modeling strategy, in which two engineered structured modalities were replaced with raw image and plain-text inputs, substantially reducing preprocessing complexity while maintaining competitive predictive performance. By embedding the GLEAM toolkit within Galaxy, we shift the focus from tool development to knowledge extraction. As a result, the framework produces high-quality, transparent, and shareable models that accelerate data-driven discovery and support cancer research.