GptGet

AGATHA: Automatic Graph-mining And Transformer based Hypothesis generation Approach

TL;DR

AGATHA uses graph-mining and Transformer models for hypothesis generation in medical research, achieving a recommendation accuracy of 0.901.

cs.LG πŸ”΄ Advanced 2020-02-14 2 views
Justin Sybrandt Ilya Tyagin Michael Shtutman Ilya Safro
AI Reader Arxiv Page Download PDF
medical research graph mining Transformer hypothesis generation deep learning

Key Findings

Methodology

AGATHA employs deep learning techniques by constructing semantic graphs and using Transformer encoders for hypothesis generation. The system converts biomedical literature from the MEDLINE database into multi-layer semantic graphs, processes text with SciBERT and ScispaCy, and performs graph embedding using PyTorch-BigGraph.

Key Results

  • AGATHA achieved a ROC AUC of 0.901 in predicting post-2015 data, significantly outperforming the previous 0.718.
  • In 'Gene - Cell Function' relationships, the system's top-10 average precision is 0.83, with a mean reciprocal rank of 0.61.
  • The system demonstrated high recommendation scores across twenty popular relationship types.

Significance

AGATHA provides early data-driven insights in medical research, reducing experimental costs and time. By automating hypothesis generation, it aids researchers in quickly identifying potential research directions, advancing drug discovery and biomedical research.

Technical Contribution

AGATHA introduces the combination of large-scale semantic graph construction and Transformer models, significantly improving the efficiency and accuracy of hypothesis generation. Compared to existing heuristic-based methods, AGATHA offers superior prediction performance and broad cross-domain discovery capabilities.

Novelty

AGATHA is the first system to apply Transformers to large-scale semantic graphs for hypothesis generation, overcoming limitations of traditional 'ABC' models and enabling more efficient many-to-many recommendation queries.

Limitations

  • The system relies on UMLS terms, limiting discovery capabilities for uncoded entities.
  • Noise in training data may affect prediction accuracy.
  • High computational resources are required for graph embedding and model training.

Future Work

Future work includes expanding the system to support more term types, optimizing the model to reduce computational resource demands, and exploring applications in more biomedical subdomains.

AI Executive Summary

AGATHA uses automated graph-mining and Transformer models for hypothesis generation in medical research, helping researchers identify potential research directions more quickly. Current medical research faces challenges with large data volumes and difficult-to-mine information; AGATHA significantly improves the efficiency and accuracy of hypothesis generation through semantic graph construction and deep learning processing.

The system leverages biomedical literature from the MEDLINE database, constructing a multi-layer semantic graph and processing text with SciBERT and ScispaCy. Through PyTorch-BigGraph graph embedding, AGATHA can perform efficient many-to-many recommendation queries in large-scale data, demonstrating high predictive capabilities across twenty popular relationship types.

AGATHA's applications extend beyond drug discovery to more biomedical subdomains. Although the system relies on UMLS terms, limiting discovery capabilities for uncoded entities, its innovation and efficiency in hypothesis generation offer new possibilities for future medical research.

Deep Dive

Abstract

Medical research is risky and expensive. Drug discovery, as an example, requires that researchers efficiently winnow thousands of potential targets to a small candidate set for more thorough evaluation. However, research groups spend significant time and money to perform the experiments necessary to determine this candidate set long before seeing intermediate results. Hypothesis generation systems address this challenge by mining the wealth of publicly available scientific information to predict plausible research directions. We present AGATHA, a deep-learning hypothesis generation system that can introduce data-driven insights earlier in the discovery process. Through a learned ranking criteria, this system quickly prioritizes plausible term-pairs among entity sets, allowing us to recommend new research directions. We massively validate our system with a temporal holdout wherein we predict connections first introduced after 2015 using data published beforehand. We additionally explore biomedical sub-domains, and demonstrate AGATHA's predictive capacity across the twenty most popular relationship types. This system achieves best-in-class performance on an established benchmark, and demonstrates high recommendation scores across subdomains. Reproducibility: All code, experimental data, and pre-trained models are available online: sybrandt.com/2020/agatha

cs.LG stat.ML