Identifying genes associated with phenotypes using machine and deep learning

This study proposes a machine and deep learning pipeline that classifies individuals based on genotype data to identify phenotype-associated genes, demonstrating that SNPs selected by high-performing models effectively prioritize disease-related genes with a mean gene identification ratio of 0.84.

The Gist

The Big Picture: Finding the "Smoking Gun" in a DNA Crime Scene

Imagine your DNA is a massive library containing billions of books (genes) and millions of tiny typos (mutations called SNPs). Sometimes, a specific typo in a book causes a person to have a certain trait, like being tall, having blue eyes, or getting a specific disease.

For a long time, scientists have tried to find these "typos" by reading the library one book at a time. This is called GWAS (Genome-Wide Association Study). It's like trying to find a needle in a haystack by picking up one piece of hay at a time and checking if it's a needle. It works, but it's slow and often misses the bigger picture.

This paper proposes a new strategy: Instead of reading one book at a time, let's hire a team of super-smart detectives (Machine Learning and Deep Learning) to look at the entire library at once, find patterns, and tell us which specific typos are the real culprits.

---

The Cast of Characters

1. The Data (openSNP): The researchers used a public database called "openSNP." Think of this as a giant, crowdsourced diary where thousands of people have uploaded their DNA results and answered questions about their lives (e.g., "Do you have allergies?" "Are you anxious?" "Do you crave sugar?").
2. The Detectives (Machine Learning & Deep Learning):
   • Machine Learning (ML): Think of these as experienced detectives who follow strict rules. They look at the data and say, "If the person has this set of typos, they are likely a 'Case' (has the trait). If they have that set, they are a 'Control' (don't have the trait)."
   • Deep Learning (DL): These are the "super-detectives" with neural networks that mimic the human brain. They are better at spotting complex, hidden patterns that the rule-following detectives might miss.
3. The Goal: To figure out which specific typos (SNPs) are the most important for predicting a trait, and then map those typos back to the specific genes (books) they live in.

---

How the Investigation Worked (The Workflow)

The researchers set up a massive experiment with 30 different "cases" (phenotypes), ranging from ADHD and Asthma to "Craving Sugar" and "Sensitivity to Mosquito Bites."

Step 1: The Lineup
They took the DNA data and cleaned it up, removing the messy or incomplete pages. They split the people into two groups: those who had the trait (Cases) and those who didn't (Controls).

Step 2: The Training Camp
They trained 21 different Machine Learning algorithms and 80 different Deep Learning models on this data.

• Analogy: Imagine training 100 different detectives on a mock crime scene. Some are good at spotting small details, others are good at seeing the big picture. They all try to guess who committed the "crime" (has the trait) based on the DNA clues.

Step 3: The Scoreboard
They tested the detectives on a new set of people they hadn't seen before. They scored them on how well they could distinguish between the "Cases" and "Controls."

• The Result: The Deep Learning detectives (specifically Artificial Neural Networks) were generally better at finding the subtle, complex patterns, while some Machine Learning detectives were great at specific tasks.

Step 4: The "Feature Importance" (The Interrogation)
This is the most crucial part. Once a detective solved the case, the researchers asked: "Which specific clues (SNPs) did you use to make that decision?"

• Analogy: Imagine a detective says, "I knew he was the thief because he had a muddy shoe, a red hat, and was holding a bag." The researchers then focus on those three items. In this study, they looked at which DNA typos the models relied on most heavily to make their predictions.

Step 5: The Cross-Reference
Finally, they took the list of "top clues" found by their AI detectives and compared it to the official police records (the GWAS Catalog, which is a database of previously confirmed gene-trait links).

• The Question: Did our AI detectives find the same culprits that the old-school methods found? Or did they find new ones?

---

The Findings: What Did They Discover?

1. The AI Detectives Were Good
On average, the AI models successfully identified 84% of the genes that were already known to be associated with these traits. This is a huge success rate! It proves that AI can effectively sift through the noise to find the signal.

2. Different Detectives, Different Strengths

• Some models were great at finding the "obvious" genes (high accuracy).
• Some models found genes that the others missed.
• Key Insight: The researchers found that by combining the results from different models (an "ensemble" approach), they could get a more complete picture than using just one. It's like having a team of detectives where one spots the shoe print, another spots the hat, and together they solve the whole case.

3. The "Missing" Clues
For some traits, the AI didn't find the known genes. Why?

• Data Quality: Sometimes the DNA data was too "fuzzy" (missing pieces).
• Population Differences: The "police records" (GWAS Catalog) might be based on people from different backgrounds than the people in this study.
• Complexity: Some traits are so complicated that a single typo isn't the cause; it's a combination of hundreds of tiny typos working together, which is hard to pin down.

4. The "Sugar Craving" Mystery
Interestingly, for some traits like "Craving Sugar," the AI couldn't find any known genes. This suggests that either the genetic link is very weak, or our current understanding of the biology is incomplete.

---

Why Does This Matter? (The "So What?")

Imagine you are a doctor trying to treat a patient.

• Old Way: You guess which gene is causing the problem based on general statistics.
• New Way (This Paper): You use an AI pipeline to scan the patient's DNA, identify the specific "typos" that are most likely causing their specific condition, and then target those exact genes with a therapy.

The Takeaway:
This paper shows that we can use Machine Learning and Deep Learning not just to predict if someone will get sick, but to actually identify the specific biological causes (genes) behind it. It's a powerful tool for Precision Medicine—moving away from "one size fits all" treatments to therapies tailored to your unique genetic blueprint.

In short: The AI didn't just guess the answer; it pointed the finger at the right genes, helping us understand the "why" behind our traits and diseases.

Technical Summary

1. Problem Statement

The identification of genes associated with specific phenotypes (observable traits or diseases) is critical for precision medicine and understanding biological mechanisms. Traditional methods, such as Genome-Wide Association Studies (GWAS), rely on scanning individual Single-Nucleotide Polymorphisms (SNPs) to find significant associations. However, GWAS often struggles with:

• Limited Predictive Value: Identified variants often explain only a small fraction of heritability.
• Linear Assumptions: GWAS typically assumes linear additive effects, potentially missing complex non-linear interactions between SNPs.
• Data Integration: Difficulty in integrating diverse data sources to refine causal gene identification.

The authors propose a pipeline using Machine Learning (ML) and Deep Learning (DL) to classify individuals based on genotype data and utilize feature importance techniques to prioritize SNPs and genes associated with specific phenotypes, aiming to overcome the limitations of traditional GWAS.

2. Methodology

The study employs a two-stage pipeline: Phenotype Classification and Feature Importance Extraction.

A. Data Source and Pre-processing

• Dataset: Data was sourced from openSNP, a crowdsourced personal genomics repository.
• Scope: Initially, 6,401 genotype files and 668 phenotypes were available. The study focused on binary phenotypes (Case vs. Control).
• Quality Control (QC):
  • Duplicate files/SNPs removed.
  • Filters applied: Hardy–Weinberg equilibrium ($p > 10^{-6}$), genotype missingness ($<0.01$), minor allele frequency ($>0.01$), and individual missingness ($<0.7$).
  • Phenotype Cleaning: Ambiguous phenotype values (e.g., "Right-handed" vs. "R") were standardized into binary classes (Case/Control/Unknown).
• SNP Selection: Fisher's exact test was performed on training data to generate GWAS summary statistics. Top SNPs were extracted based on p-value thresholds (ranging from top 50 to 10,000 SNPs) to create reduced datasets for model training.
• Final Cohort: After filtering for phenotypes with overlapping SNPs between the dataset and the GWAS Catalog, 30 phenotypes were selected for analysis.

B. Modeling Approach

The pipeline tested 21 Machine Learning algorithms and 80 Deep Learning variants.

1. Machine Learning (ML):

   • Implemented via scikit-learn.
   • Algorithms included: XGBoost, Random Forest, Gradient Boosting, SGD, SVM (SVC), and others.
   • Hyperparameters were largely kept at defaults.
   • Feature Importance:
     • Tree-based: Gini impurity reduction or feature usage count.
     • Linear (SVC/SGD): Absolute values of learned hyperplane coefficients.

2. Deep Learning (DL):

   • Architectures: Artificial Neural Networks (ANN), Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM).
   • Dynamic Architecture: The number of neurons in layers was dynamically adjusted based on the input SNP count ($S$) using formulas like $128 + 2\sqrt{S}$, allowing the same architecture to handle datasets of varying dimensionality.
   • Hyperparameter Tuning: 80 models were generated by varying 4 hyperparameters: Dropout (0.2, 0.5), Optimizer (Adam), Batch Size (1, 5), and Epochs (50, 200).
   • Feature Importance: Calculated using Feature Dropout. The model's performance was evaluated after dropping each input feature individually; the magnitude of performance drop indicated the feature's importance.

C. Evaluation and Validation

• Metrics: Models were evaluated using Area Under the Curve (AUC), F1 Score, and Matthews Correlation Coefficient (MCC).
• Validation: Stratified 5-fold cross-validation.
• Gene Identification: The top-ranked SNPs from the best-performing models (optimized for AUC, F1, or MCC) were mapped to genes. These were compared against known phenotype-associated SNPs/genes from the GWAS Catalog.
• Metric: Gene Identification Ratio (GIR) was calculated as:
  $$\text{GIR} = \frac{\text{Number of Genes Identified by ML/DL}}{\text{Number of Genes in GWAS Catalog}}$$

3. Key Results

• Classification Performance:
  • ML: XGBoost variants achieved the highest AUC for 18 out of 30 phenotypes.
  • DL: ANN performed best across most phenotypes for all metrics.
  • Comparison: DL models generally outperformed ML in MCC and F1 Score, while ML (specifically XGBoost) achieved slightly higher AUC on average.
• Gene Identification:
  • The mean Gene Identification Ratio (GIR) across all phenotypes was 0.84.
  • Correlation: There was a positive correlation between model performance (specifically MCC-optimized DL models) and the number of genes identified.
  • Three Scenarios Observed:
    1. High Performance, No Genes: 11 phenotypes had high classification accuracy but identified no overlapping genes. Reasons included low SNP coverage, linkage disequilibrium removing causal SNPs, non-linear model weighting of unrelated SNPs, or population structure mismatches.
    2. High Performance, High Identification: 9 phenotypes showed a strong correlation where better models identified more causal genes.
    3. Performance Independent: Some genes were identified regardless of the specific performance metric used.
• Impact of P-value Thresholding: Reducing the number of SNPs via p-value thresholds improved classification performance but reduced the absolute number of common SNPs found between the dataset and GWAS Catalog. However, the resulting GIR remained robust (0.84 mean).
• Cross-Phenotype Overlap: The study identified shared SNPs and genes across related conditions (e.g., Depression, Mental Disease, and ADHD shared risk variants), suggesting the pipeline can detect pleiotropic effects.

4. Key Contributions

1. Novel Pipeline: Proposed a unified ML/DL pipeline that combines variant prioritization (via p-value filtering) with feature importance extraction to identify causal genes, offering an alternative or complementary approach to standard GWAS.
2. Comprehensive Algorithm Benchmarking: Systematically evaluated 21 ML and 80 DL variants, providing empirical evidence that DL models (optimized for MCC/F1) and specific ML models (XGBoost for AUC) offer distinct advantages in gene identification.
3. Dynamic DL Architecture: Introduced a method to scale neural network layer sizes based on input SNP count ($\sqrt{S}$), enabling a single model architecture to handle diverse genomic datasets without re-engineering.
4. Feature Importance via Dropout: Applied feature dropout in DL models to rank SNPs, demonstrating that maximizing classification performance correlates with identifying biologically relevant features.
5. Resource Availability: All code, processed datasets, and detailed results (including specific gene lists for 30 phenotypes) are made publicly available on GitHub.

5. Significance and Conclusion

This study demonstrates that Machine Learning and Deep Learning are viable tools for prioritizing SNPs and identifying phenotype-associated genes, potentially outperforming or complementing traditional GWAS in specific contexts.

• Precision Medicine: By identifying genes that maximize the separation between cases and controls, the pipeline helps prioritize candidate therapeutic targets and understand disease mechanisms.
• Handling Complexity: The success of DL models suggests that non-linear interactions between SNPs (which GWAS often misses) are significant in phenotype determination.
• Limitations & Future Work: The study notes that genotype data quality, population structure, and the specific p-value thresholds used significantly impact results. The authors suggest this pipeline could serve as a pre-processing step for GWAS to narrow down genomic regions for deeper investigation.

In summary, the paper validates that optimizing ML/DL models for classification performance effectively highlights the genetic variants driving phenotypic variation, providing a robust framework for modern genomic analysis.