RESEARCH GROUP

THEORETICAL

COMPUTER SCIENCE

Connecting chemical structural representations with meaningful categories and semantic annotations representing existing knowledge enables data-driven digital discovery from chemistry data. Ontologies are semantic annotation resources that provide definitions and a classification hierarchy for a domain. They are widely used throughout the life sciences. ChEBI is a large-scale ontology for the domain of biologically interesting chemistry that connects representations of chemical structures with meaningful chemical and biological categories. Classifying novel molecular structures into ontologies such as ChEBI has been a longstanding objective for data scientific methods, but the approaches that have been developed to date are limited in several ways: they are not able to expand as the ontology expands without manual intervention, and they are not able to learn from continuously expanding data. We have developed an approach for automated classification of chemicals in the ChEBI ontology based on a neuro-symbolic AI technique that harnesses the ontology itself to create the learning system. We provide this system as a publicly available tool, Chebifier, and as an API, ChEB-AI. We here evaluate our approach and show how it constitutes an advance towards a continuously learning semantic system for chemical knowledge discovery.

Integrating human knowledge into neural networks has the potential to improve their robustness and interpretability. We have developed a novel approach to integrate knowledge from ontologies into the structure of a Transformer network which we call ontology pre-training: we train the network to predict membership in ontology classes as a way to embed the structure of the ontology into the network, and subsequently fine-tune the network for the particular prediction task. We apply this approach to a case study in predicting the potential toxicity of a small molecule based on its molecular structure, a challenging task for machine learning in life sciences chemistry. Our approach improves on the state of the art, and moreover has several additional benefits. First, we are able to show that the model learns to focus attention on more meaningful chemical groups when making predictions with ontology pre-training than without, paving a path towards greater robustness and interpretability. Second, the training time is reduced after ontology pre-training, indicating that the model is better placed to learn what matters for toxicity prediction with the ontology pre-training than without. This strategy has general applicability as a neuro-symbolic approach to embed meaningful semantics into neural networks.

Chemical data is increasingly openly available in databases such as PubChem, which contains approximately 110 million compound entries as of February 2021. With the availability of data at such scale, the burden has shifted to organisation, analysis and interpretation. Chemical ontologies provide structured classifications of chemical entities that can be used for navigation and filtering of the large chemical space. ChEBI is a prominent example of a chemical ontology, widely used in life science contexts. However, ChEBI is manually maintained and as such cannot easily scale to the full scope of public chemical data. There is a need for tools that are able to automatically classify chemical data into chemical ontologies, which can be framed as a hierarchical multi-class classification problem. In this paper we evaluate machine learning approaches for this task, comparing different learning frameworks including logistic regression, decision trees and long short-term memory artificial neural networks, and different encoding approaches for the chemical structures, including cheminformatics fingerprints and character-based encoding from chemical line notation representations. We find that classical learning approaches such as logistic regression perform well with sets of relatively specific, disjoint chemical classes, while the neural network is able to handle larger sets of overlapping classes but needs more examples per class to learn from, and is not able to make a class prediction for every molecule. Future work will explore hybrid and ensemble approaches, as well as alternative network architectures including neuro-symbolic approaches.

Abstract In cognitive science, image schemas are identified as fundamental patterns of cognition. They are schematic prelinguistic conceptualisations of events and serve as conceptual building blocks for concepts. This paper proposes that image schemas can play an important role in computational concept invention, namely within the computational realisation of conceptual blending. We propose to build a library of formalised image schemas, and illustrate how they can guide the search for a base space in the concept invention work flow. Their schematic nature is captured by the idea of organising image schemas into families. Formally, they are represented as heterogeneous, interlinked theories.

Conceptual blending has been employed very successfully to understand the process of concept invention, studied particularly within cognitive psychology and linguistics. However, despite this influential research, within computational creativity little effort has been devoted to fully formalise these ideas and to make them amenable to computational techniques. Unlike other combination techniques, blending aims at creatively generating (new) concepts on the basis of input theories whose domains are thematically distinct but whose specifications share structural similarity based on a relation of analogy, identified in a generic space, called the base ontology. We here introduce the basic formalisation of conceptual blending, as sketched by the late Joseph Goguen, and discuss some of its variations. We illustrate the vast array of conceptual blends that may be covered by this approach and discuss the theoretical and conceptual challenges that ensue. Moreover, we show how the Distributed Ontology Language DOL can be used to declaratively specify blending diagrams of various shapes, and discuss in detail how the workflow and creative act of generating and evaluating a new, blended concept can be managed and computationally supported within Ontohub, a DOL-enabled theory repository with support for a large number of logical languages and formal linking constructs.

Conceptual blending has been employed very successfully to understand the process of concept invention, studied particularly within cognitive psychology and linguistics. However, despite this influential research, within computational creativity little effort has been devoted to fully formalise these ideas and to make them amenable to computational techniques. We here present the basic formalisation of conceptual blending, as sketched by the late Joseph Goguen, and show how the Distributed Ontology Language DOL can be used to declaratively specify blending diagrams. Moreover, we discuss in detail how the workflow and creative act of generating and evaluating a new, blended concept can be managed and computationally supported within Ontohub, a DOL-enabled theory repository with support for a large number of logical languages and formal linking constructs.

Conceptual blending has been employed very successfully to understand the process of concept invention, studied particularly within cognitive psychology and linguistics. However, despite this influential research, within computational creativity little effort has been devoted to fully formalise these ideas and to make them amenable to computational techniques. We here present the basic formalisation of conceptual blending, as sketched by the late Joseph Goguen, and show how the Distributed Ontology Language DOL can be used to declaratively specify blending diagrams. Moreover, we discuss in detail how the workflow and creative act of generating and evaluating a new, blended concept can be managed and computationally supported within Ontohub, a DOL-enabled theory repository with support for a large number of logical languages and formal linking constructs.

The theory of conceptual blending has been applied very successfully in cognitive science to explain the process of concept generation. In previous work, we have shown a formalisation and implementation of conceptual blending that is borrowing techniques from logic, ontological engineering, and algebraic specification. However, so far our attempts (and similar approaches in the literature) have been reconstructive; that is they show how a particular concept (e.g., houseboat) can be `discovered' by computationally blending two carefully selected input spaces (e.g, a house and a boat) and a suitable base. This paper describes an attempt to take the next step and automate the concept generation based on a database of input spaces. Particularly, we attempt to create monsters from a library of animals formalised as OWL ontologies.

Motivation: Advancing our understanding of how nervous systems work will require the ability to store and annotate 3D anatomical datasets, recording morphology, partonomy and connectivity at multiple levels of granularity from subcellular to gross anatomy. It will also require the ability to integrate this data with other data-types including functional, genetic and electrophysiological data. The web ontology language OWL2 provides the means to solve many of these problems. Using it, one can rigorously define and relate classes of anatomical structure using multiple criteria. The resulting classes can be used to annotate datasets recording, for example, gene expression or electrophysiology. Reasoning software can be used to automate classification and error checking and to construct and answer sophisticated combinatorial queries. But for such queries to give consistent and biologically meaningful results, it is important that both classes and the terms (relations) used to relate them are carefully defined. Results: We formally define a set of relations for recording the spatial and connectivity relationships of neuron classes and brain regions in a broad range of species, from vertebrates to arthropods. We illustrate the utility of our approach via its application in the ontology that drives the Virtual Fly Brain web resource.