Neuro-symbolic AI

Neuro-symbolic methods in bioinformatics aim to combine the deductive guarantees of symbolic knowledge with the inductive power of statistical learning. Our group develops methods that map entities described in formal ontologies into vector spaces while preserving the semantic relations expressed by their axioms, so that downstream models can use background knowledge directly in similarity search, link prediction, and classification. The distinctive angle at KAUST is a focus on description logics as the source of structure: we design embedding constructions for languages such as EL++ and ALC that come with mathematical guarantees about the logical theories they approximate, rather than treating ontologies as plain graphs.

Our early work in this area established that knowledge graphs derived from biomedical ontologies can be used as substrates for representation learning. In Neuro-symbolic representation learning on biological knowledge graphs we introduced feature-learning methods that operate over RDF and OWL data, and Onto2Vec showed that treating logical axioms as sentences over class names yields embeddings that capture both formal structure and annotation context. OPA2Vec extended this idea by combining axioms with annotation properties such as labels, synonyms, and natural-language definitions, while Vec2SPARQL: integrating SPARQL queries and knowledge graph embeddings demonstrated how such vectors can be queried alongside symbolic data through standard Semantic Web interfaces. The broader rationale for this programme is set out in Data science and symbolic AI: Synergies, challenges and opportunities and developed further in the recent overview Neuro-Symbolic AI in Life Sciences.

Geometric embeddings for description logics

A core technical line of work is the construction of geometric models of description-logic theories. EL Embeddings: Geometric construction of models for the Description Logic EL++ introduced a family of embeddings in which classes are represented as n-balls and axioms become geometric constraints, so that satisfying the constraints corresponds to constructing a model of the theory. More recent work, including Enhancing Geometric Ontology Embeddings for EL++ with Negative Sampling and Deductive Closure Filtering and Lattice-Preserving ALC Ontology Embeddings, extends these ideas to more expressive logics and tightens the relationship between embedding geometry, logical entailment, and the lattice of concepts. From Axioms over Graphs to Vectors, and Back Again: Evaluating the Properties of Graph-based Ontology Embeddings systematically compares projection-based and axiom-aware approaches; the Ontology Embedding: A Survey of Methods, Applications and Resources consolidates the now-substantial literature on this question.

Applications to function and variant prediction

Neuro-symbolic models translate naturally into biomedical prediction tasks where the label space itself is an ontology. The DeepGO: predicting protein functions from sequence and interactions using a deep ontology-aware classifier and DeepGOPlus: improved protein function prediction from sequence families showed how Gene Ontology structure can be exploited inside neural classifiers, while DeepGOZero: improving protein function prediction from sequence and zero-shot learning based on ontology axioms used EL embeddings to predict functions for classes that were never observed during training. Predicting protein functions using positive-unlabeled ranking with ontology-based priors addresses the partial-annotation problem with a learning theory that respects the open-world semantics of GO. In parallel, DeepPheno: Predicting single gene loss-of-function phenotypes using an ontology-aware hierarchical classifier, DeepPVP: phenotype-based prioritization of causative variants using deep learning, and Prioritizing genomic variants through neuro-symbolic, knowledge-enhanced learning apply the same philosophy to phenotype and variant prioritization, and Combining biomedical knowledge graphs and text to improve predictions for drug-target interactions and drug-indications illustrates the approach for drug discovery.

Most of these methods are released through the mOWL: Python library for machine learning with biomedical ontologies framework, which unifies projection-, axiom-, and geometric-embedding methods behind a single API. The programme is actively supported through KAUST projects on sound, complete, and explainable machine learning with biomedical ontologies, on variant prioritization in complex disease, on personalized cancer treatment prediction, and through participation in the KAUST Center of Excellence for Generative AI and the Smart Health Initiative.