AAIC Conference (2025)
Sungjoon Park, Kyungwook Lee, Soorin Yim, Doyeong Hwang, Kiyoung Kim, Danniel Gatti(Jackson Laboratory), Elissa Chesler(Jackson Laboratory), Amy Dunn(Jackson Laboratory), Kristen O'Conell(Jackson Laboratory)
Abstract
Background
Multi-omics data from large-scale consortium databases, such as ROSMAP and ADNI, combined with advanced AI technologies, holds significant promise for identifying biological mechanisms and biomarkers for diagnosis and treatment. However, the incompleteness of multi-omics data presents a challenge for machine learning approaches in fully utilizing these rich resources. Additionally, early integration of multi-omics data often leads to modality collapse. To address these challenges, we designed an Alzheimer's disease (AD) prediction method that utilizes incomplete modalities and identifies key biomarkers through feature importance analysis.
Methods
Our model consists of Encoder-Aggregator-Predictor modules. Each omics data type is independently *encoded* into an embedding vector. The embedding vectors are then *aggregated* into a single combined vector, allowing complementary contributions to AD prediction. The aggregated vector, along with other individual embedding vectors, is fed to a shared *predictor*. The outputs are used to compute a collective loss, ensuring alignment across the heterogeneous omics embeddings. In cases of missing modalities, the embedding vectors of available modalities are amplified to compensate for the loss of information.
Results
Multiple heterogeneous omics data types were used for AD prediction without requiring domain-specific knowledge for each omics modality. Our model was trained by samples with complete and incomplete modalities, achieving state-of-the-art accuracy with a score of 0.890 in classifying CogDX labels on the ROSMAP dataset. Ablation studies and an examination of the aggregator confirmed that all omics data contributed to the prediction, effectively preventing modality collapse. We also evaluated the prediction performance across varying numbers of available modalities, demonstrating its practicality. Feature importance analysis identified biomarkers, which align with findings from existing literature.
Conclusions
The suggested method successfully integrates multi-omics data, particularly in scenarios involving incomplete modalities. The model rediscovered biomarkers consistent with other studies, presenting the potential of deep learning approaches in multi-omics research for AD.