MIDLPred: Ensemble Learning for Phenotype-Oriented Classification of Male Infertility Candidate Proteins from Primary Amino Acid Sequences

Technical Summary: MIDLPred – Ensemble Learning for Phenotype-Oriented Classification of Male Infertility Candidate Proteins

Problem Statement
Male infertility is a multifactorial condition involving complex genetic, molecular, and environmental interactions. While advances in reproductive genetics have identified numerous genes and proteins associated with spermatogenesis and testicular function, linking candidate proteins to specific clinical infertility phenotypes remains a significant challenge. Traditional diagnostic methods, primarily relying on semen analysis and World Health Organization (WHO) criteria, often lack the sensitivity and mechanistic insight required for precision medicine. Furthermore, the pleiotropic nature of infertility-related genes complicates the interpretation of phenotypes. There is a critical need for computational tools capable of extracting biologically meaningful molecular signatures directly from primary amino acid sequences to assist in prioritizing candidate proteins involved in male infertility mechanisms.

Methodology
The study introduces MIDLPred, a computational framework designed to classify male infertility candidate proteins into phenotype-oriented categories based solely on their primary amino acid sequences. The methodology follows a structured three-phase approach:

1. Data Collection and Preprocessing:

   • Source: A dedicated dataset was constructed by integrating data from the Male Infertility Knowledgebase (MIK) and the UniProt database.
   • Curation: The initial collection of 31,774 protein entries underwent rigorous cleaning, including the removal of duplicates, isoforms, and sequence variants. Sequences were filtered to retain lengths between 100 and 1,000 amino acids.
   • Final Dataset: The curated dataset comprised 10,678 unique protein sequences.
   • Classification Strategy: To address class sparsity and biological heterogeneity, proteins were grouped into four high-level phenotype categories: MOTILITY (sperm movement defects), MORPH (morphological abnormalities), PHYSIO (testicular physiology/spermatogenic dysfunction), and OTHER (heterogeneous conditions like cryptorchidism or endocrine disorders).
   • Preprocessing: Sequences were tokenized at the character level (amino acid), converted to integers, and padded to a fixed length of 1,000 residues.

2. Model Architecture:

   • Core Model: A 1D Convolutional Neural Network (1D-CNN) was developed to capture local dependencies and motifs within protein sequences.
   • Structure: The architecture includes an embedding layer (128 dimensions), three stacked convolutional blocks (128 filters, kernel size 3, ReLU activation) followed by MaxPooling1D, and a series of fully connected dense layers (256, 128, 64, 64 units) with ReLU activation and Dropout (0.5) to prevent overfitting.
   • Output: A final dense layer with 4 neurons and Softmax activation performs multi-class classification.
   • Training: The model was trained using the Adamax optimizer and sparse categorical crossentropy loss.

3. Ensemble Learning Strategy:

   • To enhance generalization and reduce variance, an ensemble of ten independently trained 1D-CNN models was implemented.
   • Aggregation: Predictions were aggregated using probabilistic averaging (summing and averaging predicted class probabilities) to determine the final class label.
   • Validation: Performance was assessed using stratified 5-fold cross-validation on the training set and a held-out 10% test set. Metrics included accuracy, precision, recall, F1-score, confusion matrices, and ROC curves.

Key Results

• Performance: The MIDLPred ensemble achieved an overall classification accuracy of 96% and a weighted F1-score of 96% across the four phenotype categories.
• Class-Specific Metrics: The model demonstrated high consistency across classes:
  • MORPH: Precision 0.98, Recall 0.99, F1 0.98.
  • MOTILITY: Precision 0.98, Recall 0.94, F1 0.96.
  • OTHER: Precision 0.94, Recall 0.97, F1 0.95.
  • PHYSIO: Precision 0.94, Recall 0.93, F1 0.94.
• Ensemble Efficacy: The ensemble approach yielded a systematic performance gain over individual models, with the ensemble achieving an accuracy of 0.9757 compared to the best single model's 0.974. The low dispersion in metrics across models (standard deviation < 0.005) confirmed the stability of the architecture.
• Generalization: Loss and precision curves indicated effective learning without significant overfitting, as training and validation metrics remained closely aligned.

Significance and Claims
The authors position MIDLPred not as a standalone diagnostic tool, but as a bioinformatics decision-support framework. Its primary significance lies in:

1. Prioritization: Assisting clinicians, reproductive biologists, and geneticists in prioritizing candidate proteins for further experimental investigation by identifying phenotype-associated sequence signatures.
2. Molecular Insight: Demonstrating that biologically relevant information regarding infertility phenotypes is embedded within primary protein sequences, which can be decoded by deep learning without explicit structural annotations.
3. Hypothesis Generation: Providing a platform to generate biologically relevant hypotheses regarding the molecular mechanisms of male infertility, potentially guiding the discovery of novel biomarkers.
4. Integration: Highlighting the value of integrating artificial intelligence with reproductive genetics. The authors note that future iterations could incorporate explainable AI (XAI), protein language models, and multi-omics data to further improve biological interpretability and clinical applicability.

The study concludes that MIDLPred represents a significant step toward precision medicine in male reproduction by offering a computationally efficient and accurate method to explore the molecular landscape of infertility through protein sequence analysis.