AI in Healthcare and Bioenergy

This project will develop Geometry-Induced Collective Inference (GICI), a decentralized computational framework for integrating heterogeneous biological omics data without sharing raw data. The study will treat each omics modality, such as scRNA-seq, scATAC-seq, spatial transcriptomics, and proteomics, as an independent computational agent. Each agent will independently encode its data into a latent representation, construct a local biological graph, and extract a compressed geometric signature using Graph Laplacian spectral information.

I will then formally develop the Geometry-Induced Consensus Operator (GCO), which will allow agents to iteratively align their geometric representations using only compressed signatures rather than raw measurements. The project will mathematically analyze GCO's convergence, communication requirements, robustness, and relationship to existing geometric alignment methods such as Gromov-Wasserstein, Procrustes, and Graph Laplacian alignment. After validating the framework on synthetic data, I will test it using public multimodal biological datasets and evaluate its ability to reconstruct biological interaction and gene regulatory networks.

Performance will be compared with established methods such as SCENIC, GENIE3, GRNBoost2, and CellOracle using Precision-Recall, ROC-AUC, network similarity, runtime, memory usage, communication cost, and robustness to noise and missing modalities. The final goal is to determine whether decentralized geometric consensus can integrate heterogeneous biological data while maintaining biological accuracy and substantially reducing the need for raw-data sharing.