Cell Maps for AI: Revolutionizing Biomedical Research with AI-Ready Data

The Cell Maps for Artificial Intelligence (CM4AI) project, a Functional Genomics Data Generation Project within the NIH’s Bridge2AI program, aims to revolutionize biomedical AI research. Its primary mission is to generate ethical, AI-ready datasets of cell architecture, derived from multimodal data collected from human cell lines. This initiative seeks to provide researchers with the tools and data necessary to develop transformative AI applications in biomedicine. CM4AI focuses on three main pillars: Data, People, and Ethics, organized into six modules covering data acquisition, tools, standards, skills development, teamwork, and ethical considerations. By creating machine-readable hierarchical maps of cell architecture, CM4AI enables a deeper understanding of cellular processes and their implications for human health.

Cell maps are hierarchical directed acyclic graphs (DAGs) that represent the organization of proteins within a cell at various scales. Each node in the graph represents an assembly of proteins in proximity, ranging from large cell compartments like the nucleus and mitochondria to smaller protein complexes. These maps are constructed using data from perturbed and unperturbed cell lines, including cancer cell lines and induced pluripotent stem cells (iPSCs). Techniques such as affinity purification-mass spectrometry (AP-MS) and immunofluorescence (IF) staining are used to generate protein interaction networks and reveal protein localization. By integrating this data, cell maps provide a foundation for interpreting genetic variants and mutations, and they can be used in AI tools for visible machine learning to understand how protein assemblies affect cell-level phenotype predictions.

CM4AI defines AI-Ready biomedical data as fully characterized FAIR data with known provenance, ethically and reliably processed for AI applications. This includes ensuring that the models and software used are available, well-described, and validated, and that the predictions made can be explained and interpreted. Key principles include: FAIRness (Findable, Accessible, Interoperable, Reusable), Provenance (availability of computation graphs), Characterization (complete schemas and data sheets), Explainability (statistical characterization and limitations), and Ethical considerations (ethical treatment of subjects and responsible data analysis). CM4AI uses an expanded version of the FAIRSCAPE framework to establish a basis for AI-readiness, focusing on rich metadata, persistent identifiers, and validation procedures.

CM4AI utilizes specific cell lines, including the MDA-MB-468 breast cancer cell line and the KOLF2.1J iPSC line, both ethically sourced. Data acquisition involves protein-protein interaction (PPI) mapping using AP-MS and SEC-MS, spatial proteomics mapping using immunofluorescence, and genetic perturbation mapping using single-cell CRISPR screens. For PPI mapping, chromatin regulators are tagged, and their interactions are analyzed under different conditions. Spatial proteomics mapping involves automated fixation and permeabilization protocols to map the subcellular organization of key proteins. Genetic perturbation mapping uses CRISPR screens to perturb chromatin regulators and analyze the resulting data.

The Multi-Scale Integrated Cell (MuSIC) pipeline is a key tool for integrating data and producing cell maps from multiple input data streams. The pipeline includes segments for downloading PPI and image data, generating embeddings using deep learning models, co-embedding to integrate PPI and image information, protein community detection, hierarchy creation, and hierarchy evaluation. The pipeline interfaces with the FAIRSCAPE infrastructure to validate inputs and create RO-Crate packages. Integrative structure modeling is also explored to increase understanding of the MuSIC communities.

CM4AI emphasizes AI-readiness packaging through the development of standards for data integration and metadata management. This includes creating data dictionaries, formatting standards, and a FAIRSCAPE metadata and provenance API. The goal is to ensure that the data is easily accessible, interoperable, and reusable for AI applications. The project also focuses on mapping data elements to public ontology vocabularies and using JSON-Schema mini-data-dictionary descriptions.

Cell maps generated by CM4AI have numerous applications in AI research. They can be used to interpret genetic variants and mutations, understand how protein assemblies affect cell-level phenotypes, and develop AI tools for visible machine learning. By providing a comprehensive view of cellular architecture, cell maps enable researchers to build more accurate and effective AI models for biomedical applications. These models can be used to predict disease outcomes, identify potential drug targets, and develop personalized treatment strategies.

The CM4AI project is continually evolving, with future directions including enhancing AI-readiness features, expanding the range of cell lines and conditions studied, and developing more sophisticated data integration and analysis tools. The project aims to have a significant impact on biomedical research by providing the data and tools necessary to develop transformative AI applications. By adhering to ethical principles and promoting FAIR data practices, CM4AI ensures that its resources are used responsibly and for the benefit of humanity.