FAIROps: In-Model Explainability for Biomedical AI/ML

Recent advances in biomedical artificial intelligence (AI) and machine learning (ML) promise to greatly accelerate therapeutic discoveries and improve patient care. However, AI/ML explainability (XAI) across pre-model, in-model, and post-model dimensions is required for reproducibility and reliable deployment in research and clinical applications. Pre-model XAI establishes AI-readiness of data and its ability to be reused, as demonstrated in the FAIRSCAPE framework. In-model and post-model XAI form the basis for explaining decisions made during model training and execution. Operational practices like DevOps, DataOps, and MLOps partially address the conceptual requirements for in-model and post-model XAI. Several software libraries, such as MLFlow and Weights & Biases, allow developers to implement the practices to document in-model explainability data, such as hyperparameters used for training, model performance metrics, and optimization experiment metadata. However, these tools do not provide user-friendly methods to aggregate and publish this information. To address this gap, we have developed a Python library, FAIROps, which allows biomedical AI/ML practitioners to easily publish and share in-model explainability results and metadata with the biomedical AI/ML community. The FAIROps library allows users to aggregate in-model metadata and results, package these data using the RO-Crate standard, and publish the information to both FAIRSCAPE and common repositories including Zenodo, Figshare, and Dataverse. We will illustrate the use of this tool with data from the functional genomics grand challenge, Cell Maps for AI (CM4AI), to monitor, log, and publish in-model XAI metrics and metadata to compare classical and quantum community detection algorithms as part of the cell mapping process. By implementing this library, AI/ML developers can easily increase the explainability and reproducibility of their models and publish them in accordance with FAIR principles.