Chebifier 2: An Ensemble for Chemistry

Accurately classifying chemical structures is a central task in cheminformatics. Different approaches to classification exist, but they all come with drawbacks. Manual classification is time-intensive and hard to scale for large databases. Rule-based methods only work for some classes and require a high level of maintenance, while deep learning models lack reliability and explainability. In this work, we present Chebifier 2, an ensemble learning framework that combines deep learning and symbolic approaches to classify chemicals into the Chemical Entities of Biological Interest (ChEBI) ontology, covering a total of 2,234 classes. The framework consists of two stages: an aggregation stage, in which base learner predictions are combined per class, and an inconsistency resolution stage, in which predictions are aligned with the axiomatisation of ChEBI. For the aggregation stage, this work evaluates five methods: Majority Voting, two Weighted Majority Voting schemes, and the meta-learning methods Learning To Rank and Dynamic Ensemble Selection. For the inconsistency resolution stage, three methods are evaluated: Hierarchy and Exclusion graphs, Iterative Local Refinement and a novel 3-step inconsistency resolution algorithm. The results show that ensemble learning can beat the best non-ensemble method by 5.9 percentage points in the macro F1 score. Particularly on small and chemically difficult classes, the ensemble outperforms the baseline by a substantial margin (≥10 percentage points on classes with fewer than 30 members). Scientific Contribution: This work introduces Chebifier 2, an ensemble framework for the classification of chemical structures in the ChEBI ontology. It covers a larger part of ChEBI than previous approaches and outperforms existing methods, especially on small classes. Besides adapting existing aggregation and inconsistency resolution methods to ChEBI, this work also proposes a novel 3-step resolution method specialised in disjointness and subsumption relations.