An explainable AI platform integrating diverse data for enhancing rigor, reproducibility, and translatability of animal research

The National Institutes of Health Office of the Director awarded The Regents of the University of California (operating Lawrence Berkeley National Laboratory) $1,603,286 on July 15, 2026, under the Research Infrastructure Programs (CFDA 93.351) to develop PhenomicsAI, an explainable AI platform that integrates genomic and phenomic data to enhance rigor, reproducibility, and translatability of animal research models.

The platform addresses fragmentation and standardization gaps across animal research repositories by combining disparate genomic and phenomic datasets in a single, interpretable system. PhenomicsAI will predict the efficacy of proposed animal disease models, reveal gaps in existing animal research, and facilitate creation of tailored disease models that more effectively recapitulate human disease. The system addresses a core translational research barrier: the lack of standardized reporting, interoperability of phenotyping across species, and the difficulty of scaling AI methods to complex phenomic data alongside genomic information. The work builds toward models with higher translatability, stronger scientific rigor, and more efficient development cycles for costly and time-consuming animal model creation.

Performance is based in Berkeley, California, with the period of performance running through March 31, 2030. The award is structured as a Project Grant, a categorical assistance type under the Research Infrastructure Programs administered by the Office of Research Infrastructure Programs within the NIH Division of Program Coordination, Planning, and Strategic Initiatives.

Abstract

An explainable AI platform integrating diverse data for enhancing rigor, reproducibility, and translatability of animal research animal models provide essential insights into human disease mechanisms and treatment strategies, making them valuable tools for translational research. For example, cystic fibrosis (CF) used to be a fatal disease with no effective treatment; advances using mice as a model system led to the development of drugs that have doubled life expectancy. However, there are biological and technical barriers to maximizing the translational effectiveness of animal models of disease. These challenges include a lack of adherence to standards in reporting studies, a lack of interoperability of phenotyping within and across species, and the complex relationship between a model system and a disease. There is a need for animal models that effectively recapitulate human disease, that are discoverable, and where the data can be maximally utilized, to optimize the often costly and time-consuming development of such models. Artificial intelligence (AI) holds great potential to improve the translatability of animal research, but several issues need to be addressed: (1) much of the relevant data is spread across different repositories, and is not standardized or labeled consistently; (2) while AI methods work well on sequence data, they are harder to adapt to complex phenomic data, and using genomic and phenomic data together is even harder; (3) AI is frequently unreliable, and is a "black box", which is an obstacle to trusting the results. We plan to address these limitations with an explainable AI (XAI) platform, phenomicsai, that will enhance translational research, increase rigor, predict the efficacy of proposed animal models, and reveal gaps in animal research, ultimately facilitating the creation of disease models tailored to individual phenotypes, diseases and mechanisms. Building on the monarch initiative knowledge graph, phenomicsai will be trained on a wide range of public animal model data, including structured knowledge about genes, diseases and phenotypes from curated repositories, atlases, and ontologies. IT will semi-automatically harmonize diverse genomic and multi-scale phenomics datasets across a diversity of animal disease models and extract relevant metadata from surrounding literature. Our collected genotype-phenotype-disease assertions will form the largest and most comprehensive corpus of animal model-to-disease knowledge to date. This corpus will be a useful resource in its own right, as well as serving as a gold standard/benchmark for AI training and validation. Phenomicsai will support model-disease matchmaking (given a disease, propose existing or new animal models recapitulating mechanism, outcomes, and potential interventions) and explainability (e.g. outlining why an animal model is a good or poor model for a disease). IT will help researchers design and validate studies reporting animal research, ensuring conformance to standardized reporting checklists. We will build a community of practice around phenomicsai, assisting a broad range of users in applying its capabilities to a wide range of disease model use cases and making their data standardized, interoperable, and AI-ready.