STonKGs

STonKGs integrates a multimodal Transformer with natural language processing and knowledge graph embedding models to learn joint representations of biomedical text and knowledge graphs for downstream biological classification and prediction tasks.

Key Features:

• Multimodal Integration: Combines structured information from biomedical knowledge graphs with unstructured textual data extracted from scientific literature.
• Joint Representation Learning: Processes combined input sequences that include both text and graph-based data to produce unified representations.
• Pre-training on INDRA Knowledge Base: Pre-trained on millions of text–triple pairs assembled by the Integrated Network and Dynamical Reasoning Assembler (INDRA) and curated through multiple NLP systems.
• Benchmarking and Superior Performance: Demonstrated higher performance than baselines trained on either modality alone (text or KG) across eight biological classification tasks, improving the best baseline F1-score by up to 0.083 and particularly outperforming on tasks with many classes.
• Transfer Learning Adaptability: Provides pre-trained models and an architecture intended for transfer learning to a variety of downstream biomedical applications.

Scientific Applications:

• Disease classification: Supports classification of diseases using integrated evidence from text and knowledge graphs.
• Drug discovery: Supports tasks in drug discovery by combining literature-derived and knowledge-graph-derived information.
• Gene function prediction: Supports prediction of gene function through multimodal representations.
• Other complex biological inquiries: Applicable to classification and prediction tasks that benefit from comprehensive integration of textual and graph-based biomedical knowledge.

Methodology:

Pre-trains a multimodal Transformer using NLP-derived text and knowledge graph embedding models (KGEMs) on millions of text–triple pairs from INDRA, processes combined text-and-graph input sequences to learn joint representations, and fine-tunes the pre-trained model on downstream classification tasks.