Self organising hypothesis networks: a new approach for representing and structuring SAR knowledge

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Integrating diverse knowledge sources for structure-activity relationship (SAR) models is challenging due to varied data formats and lack of interoperability.
• Current methods often rely on consensus models at the prediction level, limiting early-stage synergy.
• This work explores combining knowledge at the knowledge level for enhanced synergy and discovery.

Purpose of the Study:

• To develop a general methodology for facilitating knowledge discovery.
• To create accurate and interpretable SAR models by integrating diverse knowledge.
• To enable earlier synergy by combining knowledge at the knowledge level.

Main Methods:

• Proposes a pivot representation (lingua franca) based on 'hypotheses' to decouple learning and knowledge application.
• Unifies disparate knowledge sources by breaking them into interpretable knowledge units (hypotheses).
• Organizes hypotheses into a hierarchical network, forming the Self Organising Hypothesis Network (SOHN).

Main Results:

• Demonstrates the feasibility of representing knowledge in a unified hypothesis network.
• Achieves interpretable predictions with performance comparable to mainstream machine learning techniques.
• Illustrates the approach with an application to predicting mutagenicity.

Conclusions:

• The Self Organising Hypothesis Network (SOHN) approach enables unified knowledge representation for interpretable predictions.
• This method offers potential for combining diverse knowledge sources within a common framework.
• Future work will explore high-level reasoning and meta-learning applications on the unified model.