Work

Explainable and Knowledge-Enhanced Machine Learning for Biomedical Representation Learning

Year: 2026
Type: dissertation
Abstract: Machine learning has achieved remarkable success in biomedical data analysis, enabling significant advances in disease prediction, protein function characterization, and molecular interaction modeling. However, many state-of-the-art models operate as black boxes, limiting their interpretability, trustworthiness, and adoption in high-stakes biomedical applications. In addition, biomedical data are inherently heterogeneous and governed by rich domain knowledge, which is often underutilized in purely data-driven approaches. These challenges highlight the need for machine learning frameworks that are both interpretable and capable of integrating domain knowledge while effectively modeling complex biomedical data. This dissertation investigates the development of explainable and knowledge-enhanced machine learning methods for biomedical representation learning. The goal is to improve both predictive performance and interpretability by combining explainable artificial intelligence, domain knowledge integration, advanced representation learning, and multimodal modeling. First, we propose an interpretable representation learning framework that uncovers task-specific dimensions in word embeddings through automatic rule learning. This approach reveals how embedding dimensions contribute to downstream tasks, providing a novel perspective on the internal structure of distributed representations. Second, we introduce a knowledge-enhanced neurosymbolic framework for transcription factor prediction. By integrating Logic Tensor Networks with deep neural architectures, the proposed method incorporates biological constraints as logical rules, improving both predictive accuracy and interpretability while ensuring consistency with domain knowledge. Third, we develop ComplexWaveGAT, a complex-valued graph neural network for protein–protein interaction prediction. By extending graph attention mechanisms into the complex domain and introducing wave-inspired interaction modeling, the model captures richer interaction patterns beyond conventional real-valued representations. Finally, an Evidence-Calibrated Mixture-of-Experts framework is developed for diabetes risk prediction using heterogeneous real-world biomedical data. This approach integrates multimodal representation learning, uncertainty calibration, and evidence-guided decision refinement to address label inconsistency and clinical ambiguity in real-world biomedical data. Together, these contributions establish a principled paradigm for integrating interpretability, domain knowledge, and advanced representation learning in biomedical machine learning. The results demonstrate that knowledge-guided and explainable models not only improve predictive performance but also produce more trustworthy and biologically grounded insights, paving the way toward reliable AI systems for healthcare and computational biology.
Author: Gao, Liyuan
Language: English
Open Access status: green
OpenAlex ID: W7214141714
Cites: 108
Cited by:
Related to: 10