PreSiBOGNN: Multi‐Modal Graph Neural Network for Prediction of Cognitive Improvement in Medication Usage in UK Biobank

Dhawal Priyadarshi, Heng Huang, Nathan Sahelijo, Reza Shirkavand, Gyungah JunAlzheimer 2024 · DOI

TL;DR

This paper introduces PreSiBOGNN, a novel multi-modal graph neural network designed to predict cognitive improvement in Alzheimer's Disease (AD) patients using medication. By integrating diverse biological and clinical data, it aims to accelerate in silico drug repositioning and pave the way for personalized AD treatments.

What Problem Does This Paper Address?

Developing new drugs for Alzheimer's Disease (AD) is notoriously expensive and time-consuming. This paper tackles the challenge of finding effective treatments more efficiently by focusing on drug repositioning – repurposing existing drugs for AD. The core problem is how to accurately predict which medications might lead to cognitive improvement in AD patients, especially at an individual level, by leveraging the complex interplay of biological and clinical factors. Traditional methods often struggle to integrate the vast and heterogeneous data available, hindering the discovery of novel therapeutic strategies.

Methods

The authors propose PreSiBOGNN, a novel machine learning framework built upon Graph Neural Networks (GNNs). The core innovation lies in its ability to integrate multi-modal data through a hierarchical, multi-layer approach.

Here's a breakdown of the key techniques:

  • Multi-Modal Integration: PreSiBOGNN combines four types of data: transcriptome (gene expression), proteome (protein interactions), drug information, and subject-level clinical data.
  • Graph Construction: This data is structured into bipartite graphs. Specifically, it uses:
  • Gene-gene coexpression networks.
  • Protein-protein interaction networks.
  • Gene-protein relationships.
  • Protein-drug relationships.
  • Drug-subject relationships.
  • Sequential Graph Attention Networks (GATs): These are employed to learn embeddings by processing the graph layers sequentially. The attention mechanism allows the model to weigh the importance of different connections within and between layers. The sequential processing follows a hierarchical order: gene-gene, gene-protein, protein-protein, protein-drug, and finally, drug-subject.
  • Two Data Aggregation Models:
    1. Strict Hierarchy: Data is aggregated in a single, predefined direction (gene -> protein -> drug -> subject).
  • Fluid Message Passing: This model allows for more flexible information flow between layers, potentially capturing more complex interactions.
  • Decoder: A Multi-Layer Perceptron (MLP) is used to decode the learned feature embeddings and predict cognitive improvement between two cognitive exams.

Datasets

The study utilizes data from the United Kingdom Biobank (UKBB). Specifically, it includes:

  • Medication usage data.
  • Clinical data.
  • Genome-Wide Association Study (GWAS) data.
  • First and second cognitive exam scores for 48,187 subjects.

The protein and transcriptome layers were constructed using:

  • The String database for protein-protein interactions.
  • Single nuclei RNA data (as described in Sahelijo et al. 2022) for gene coexpression networks.

Drug information was sourced from:

  • UniProt database.
  • DrugCentral database.

The UKBB data was split into training (60%), validation (30%), and testing (10%) sets.

Key Insights and Ideas

  • The Power of Multi-Modal Integration: The study demonstrates that combining diverse biological (omics) and clinical data within a graph framework can significantly improve predictions for drug efficacy in AD. This moves beyond single-modality approaches.
  • Hierarchical Graph Representation: Structuring the data into a hierarchy of interconnected biological entities (genes, proteins, drugs) and then linking them to subjects allows the GNN to learn complex, context-aware representations.
  • Fluid Message Passing Advantage: The "fluid message passing" model, which allows for more dynamic information flow between graph layers, outperformed the strictly hierarchical model. This suggests that biological and drug interactions are not always linear and can benefit from more flexible information exchange.
  • Potential for Drug Repositioning: The framework's ability to predict cognitive improvement based on medication usage highlights its potential as a tool for identifying existing drugs that could be repurposed for AD treatment, thereby reducing development costs and timelines.

Key Results

The PreSiBOGNN framework was evaluated on its ability to predict cognitive improvement. The key quantitative results are:

  • Strict Hierarchy Model: Achieved validation accuracy of 57.7% and test accuracy of 51.2%.
  • Fluid Message Passing Model: Showed improved performance with validation accuracy of 61.3% and test accuracy of 58.1%.

These results indicate that the fluid message passing approach offers a more accurate prediction of cognitive improvement compared to the strict hierarchical aggregation.

Limitations and Future Work

The authors acknowledge the following limitations and suggest future directions:

  • Model Optimizations: There is scope for further refinement and optimization of the PreSiBOGNN model to enhance its predictive power.
  • Integration of Additional Modalities: The authors plan to incorporate more data types, specifically mentioning compound-specific fingerprint data. This could provide richer information about drug properties and their interactions.
  • Broader Applicability: While focused on AD, the framework's multi-modal graph approach could potentially be adapted for predicting treatment outcomes in other complex diseases.

Summary updated

Abstract

Abstract Background The prohibitive costs of drug development for Alzheimer’s Disease (AD) emphasize the need for alternative in silico drug repositioning strategies. Graph learning algorithms, capable of learning intrinsic features from complex network structures, can leverage existing databases of biological interactions to improve predictions in drug efficacy. We developed a novel machine learning framework, the PreSiBOGNN, that integrates muti‐modal information to predict cognitive improvement at the subject level for precision medicine in AD. Method The graph neural network framework integrates four layers of input data including transcriptome, proteome, drug, and subject to connect to bipartite graphs in the United Kingdom Biobank (UKBB). Medication usage, clinical, and GWAS data were downloaded for 48187 subjects with first and second cognitive exams from the UKBB. The protein and transcriptome layers were constructed using the String database and gene coexpression networks generated from single nuclei RNA data (Sahelijo et al. 2022). Layers were connected by binary bipartite graphs constructed using drug information from the UniProt and DrugCentral databases. Sequential Graph Attention Networks convoluted embedded features generated by each layer in a hierarchical order: 1. gene‐gene, 2. gene‐protein, 3. protein‐protein, 4. protein‐drug, 5. drug‐subj. Feature embeddings were decoded using a multilayer perceptron to predict cognitive improvement between the first and second cognitive exams. Two models compared methods of data aggregation. The first model follows a strict hierarchy, aggregating inter‐layer data in a single direction (gene > protein > drug > subj). The second model allows for fluid message passing between inter‐layers. We used 60% of the UKBB subjects for training and 30% for validation. We assessed the model with the greatest training accuracy using the remaining 10% of subjects. Result We observed that the performance of the strict message‐passing model attained validation and test accuracy of 57.7% and 51.2%, respectively. Performance of the fluid message passing model improved prediction accuracy with 61.3% and 58.1% in the test set. Conclusion Our investigation suggests the feasibility of the PreSiBOGNN framework to infer cognitive improvement of existing drugs by integrating medication, multi‐omics, and clinical data. Future work will focus on model optimizations and the integration of additional modalities including compound‐specific fingerprint data.
BibTeX
@inproceedings{priyadarshi2024presibognn,
  author = {Dhawal Priyadarshi and Heng Huang and Nathan Sahelijo and Reza Shirkavand and Gyungah Jun},
  title = {{PreSiBOGNN: Multi‐Modal Graph Neural Network for Prediction of Cognitive Improvement in Medication Usage in UK Biobank}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.089290},
}
Figure