Researchers from the Broad Institute of MIT and Harvard, the Massachusetts Institute of Technology (MIT), and ETH Zurich, in collaboration with the Paul Scherrer Institute (PSI), have introduced APOLLO – an innovative artificial intelligence framework designed to interpret complex, multilayered cellular data. This method empowers scientists to distinguish biological signals that are common across various measurement techniques from those unique to specific assays, enhancing precision in disease research and experimental planning.
In modern cell biology, multimodal strategies are essential for capturing diverse aspects of cellular behavior. Techniques such as transcriptomics (for gene expression), chromatin accessibility assays, protein quantification, and cell morphology imaging each reveal distinct dimensions. However, integrating these data streams has been challenging, as traditional machine learning models often fuse them into a single latent representation, losing track of signal origins.
APOLLO overcomes this by structuring data into shared and modality-specific latent spaces, akin to a Venn diagram. Overlapping biological information is encoded in a common space, while exclusive features are isolated in separate compartments. This preserves traceability and enables granular analysis.
At its core, APOLLO employs a redesigned multimodal autoencoder with a two-stage optimization process. The first stage trains decoders to reconstruct inputs from latent spaces, establishing stable feature extraction per modality. The second refines encoders for alignment, separating shared from unique signals. Once trained, APOLLO analyzes unseen datasets, classifying information as cross-modal or modality-specific.
Validation on synthetic datasets confirmed APOLLO’s accuracy in recovering predefined signals. In real-world applications, it excelled with paired single-cell data.
Practically, APOLLO identifies assay-responsible biomarkers, such as DNA damage markers in cancer cells, guiding assay selection for monitoring disease or therapy responses. It also supports decisions on direct measurements versus computational inference, optimizing costs in multimodal profiling.
Complementing such advanced frameworks are specialized AI tools focused on early detection, like QuData’s AI-powered computer-aided detection system for breast cancer. This solution uses deep learning to automatically analyze and classify mammography images according to the BI-RADS system, marking suspicious lesions with bounding boxes, enhancing diagnostic accuracy, reducing missed diagnoses and false positives, and supporting radiologists in achieving earlier and more consistent breast cancer detection.
Beyond cancer, APOLLO holds promise for neurodegenerative diseases like Alzheimer’s, metabolic disorders such as diabetes, and other conditions involving multilayered cellular regulation. By elucidating interactions across components, it fosters a systems-level grasp of disease mechanisms.
Future enhancements aim to boost interpretability, extend to unpaired data (e.g., via distribution-matching losses), and scale to biobanks for precision medicine.
