Knowledge Inclusive Machine Learning for Disease Gene Prioritisation

This paper introduces Knowledge Inclusive Machine Learning (KIML), a novel paradigm that integrates experimental data with literature-derived and structured biomedical knowledge to significantly improve the accuracy, interpretability, and generalizability of disease gene prioritization compared to existing methods.

The Gist

Imagine you are trying to find a specific missing person in a massive, crowded city. To do this, you have two very different types of help available, but neither is perfect on its own.

The Two Types of Help

1. The "Live Camera Feed" (Experimental Data): This is like watching a live security camera feed of the city right now. It shows you exactly who is where at this specific moment. However, the camera is glitchy; sometimes the image is blurry, sometimes it's too dark, and it only shows you what's happening right now without telling you who these people are or what they usually do. If you rely only on this, you might mistake a stranger for the person you're looking for because they happened to be wearing the same red hat.
2. The "Encyclopedia of the City" (Curated Knowledge): This is like having a giant, well-written encyclopedia that lists every person in the city, their family trees, their jobs, and their known habits. It's accurate and reliable, but it's too general. It tells you that "John Smith is a doctor," but it doesn't tell you which specific "John Smith" is currently standing in the park looking for help. It lacks the fine detail needed to pick out one specific individual from a crowd.

The Problem
Most scientists trying to find disease-causing genes (the "missing people") have been using only the "Live Camera Feed." Because the data is noisy and specific to just one experiment, their computer models often get tricked. They start guessing based on random patterns (like "everyone in this photo is wearing a red hat") rather than understanding the real biology.

The Solution: Knowledge Inclusive Machine Learning (KIML)
The authors of this paper introduced a new method called KIML. Think of KIML as a super-intelligent detective who refuses to rely on just one source. Instead, this detective:

• Watches the live camera feed (the experimental data).
• Cross-references it with the encyclopedia (curated knowledge).
• Even checks the local newspaper archives (literature from PubMed) and the city's official database (biomedical knowledge graphs).

By combining the "now" with the "known history," the detective can ignore the camera glitches and focus on the real story.

What They Found
The researchers tested this new detective (KIML) on a specific condition called Developmental and Epileptic Encephalopathy. They compared it against other methods that only used the "camera feed."

• Better Accuracy: KIML was much better at correctly identifying the right genes.
• Real Understanding: When the model made a guess, it could explain why it made that choice using biological facts, not just random math.
• Versatility: The method wasn't a one-trick pony; it worked just as well when tested on six other different diseases.

The Bottom Line
This paper argues that to truly understand complex diseases, you can't just look at the raw data from a single experiment. You need to wrap that data in the context of everything we already know about biology. By teaching machines to read the "encyclopedia" while they watch the "camera," we get smarter, more reliable answers about which genes are causing diseases.

Technical Summary

Based on the abstract provided, here is a detailed technical summary of the paper "Knowledge Inclusive Machine Learning for Disease Gene Prioritisation":

1. Problem Statement

The core challenge addressed is the limitation of current machine learning (ML) approaches in disease gene prioritisation. Existing models typically rely heavily on specific context derived from sample-level experimental data (e.g., gene expression profiles, protein-protein interaction networks). However, this approach suffers from two critical flaws:

• Noise and Overfitting: Experimental data is often sensitive to dataset-specific noise, leading models to learn spurious correlations rather than true biological mechanisms.
• Lack of General Grounding: Relying solely on experimental data ignores broader biological context, limiting the model's ability to generalize.

Conversely, general context from curated biological knowledge (established relationships between genes, diseases, and pathways) lacks the resolution necessary for fine-grained, gene-level discrimination. The paper posits that neither data source is sufficient in isolation, necessitating a unified approach.

2. Methodology: Knowledge Inclusive Machine Learning (KIML)

The authors propose KIML, a novel paradigm designed to unify specific and general contexts within a single analytical pipeline. The methodology integrates three distinct data modalities:

1. Experimental Data: Sample-level data providing specific context (e.g., gene expression, PPI networks).
2. Literature-Derived Representations: Unstructured knowledge extracted from PubMed, capturing semantic relationships and recent findings.
3. Structured Biomedical Knowledge Graphs: Curated, structured data encoding established biological relationships.

By fusing these heterogeneous data sources, KIML aims to ground experimental observations in established biological theory while maintaining the resolution required for specific gene prioritisation.

3. Key Contributions

• Unified Framework: The introduction of KIML as a paradigm that successfully bridges the gap between noisy experimental data and high-level curated knowledge.
• Multi-Modal Integration: The specific architectural integration of literature-derived embeddings and structured knowledge graphs alongside traditional experimental features.
• Interpretability: Unlike "black-box" models, the framework generates interpretable explanations for why specific genes are prioritized, enhancing trust and biological utility.
• Robust Evaluation Protocol: The adoption of temporal-split evaluation, a rigorous testing method that ensures models are evaluated on future data relative to their training set, preventing data leakage and providing a realistic assessment of generalisability.

4. Results

The approach was rigorously evaluated using Developmental and Epileptic Encephalopathy (DEE) as a primary case study, with benchmarks against recent state-of-the-art methods using publicly available datasets.

• Predictive Performance: KIML consistently outperformed existing approaches, demonstrating superior predictive accuracy.
• Biological Validity: Beyond raw metrics, the model's outputs were validated through ontology enrichment analysis, confirming that the prioritized genes were biologically meaningful and relevant to the disease mechanisms.
• Generalisability: The framework demonstrated strong transferability, maintaining high performance when applied to six additional diseases beyond the primary case study.

5. Significance

This work represents a significant shift in computational biology by moving away from data-hungry, purely experimental ML models toward knowledge-inclusive systems.

• Scientific Impact: It mitigates the risk of learning spurious correlations, ensuring that gene prioritisation is grounded in established biological reality.
• Clinical Utility: The ability to provide interpretable explanations makes the tool more actionable for researchers and clinicians seeking to identify novel therapeutic targets.
• Future Direction: KIML establishes a new standard for integrating heterogeneous biological data sources, suggesting that future ML models in genomics must explicitly incorporate general knowledge to achieve robust, generalisable, and biologically sound results.