An Interpretable Complex Knowledge Multi-Hop Reasoning Model for Predicting Synthetic Lethality in Human Cancers.

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Synthetic lethality (SL) is a key cancer treatment strategy, but experimental validation is costly and slow.
• Machine learning (ML) models enhance SL prediction but lack interpretability and struggle with complex, multi-factor reasoning.
• Current ML models often focus on simple gene pairs, limiting their applicability to real-world clinical scenarios.

Purpose of the Study:

• To develop an explainable, multi-hop reasoning model for synthetic lethality (SL) prediction.
• To address the interpretability limitations and scope restrictions of existing ML-based SL prediction methods.
• To integrate diverse medical entities into SL prediction through a first-order logic query framework.

Main Methods:

• Constructed query graphs using triplet transformations for various SL prediction tasks.
• Employed a sparse Transformer encoder for node embeddings and a graph attention decoder for multi-hop logical reasoning chains.
• Incorporated node masking in intermediate steps to enable explicit prediction and observation of the reasoning process.

Main Results:

• Achieved superior performance over state-of-the-art methods on complex SL prediction benchmarks.
• Demonstrated the model's capability to handle diverse medical entities and multi-factor reasoning.
• Successfully generated specific, multi-hop logical reasoning chains for model predictions.

Conclusions:

• The proposed EFOL-SL model offers a significant advancement in explainable AI for synthetic lethality prediction.
• EFOL-SL provides interpretable insights into the reasoning behind SL predictions, facilitating clinical understanding and trust.
• The model's ability to perform multi-hop reasoning with diverse medical entities enhances its potential for real-world applications in cancer medicine.