Mid-infrared spectroscopy and supervised machine learning can determine age class of arboviral vector Culex pipiens at high age resolution

Technical Summary: Mid-infrared Spectroscopy and Supervised Machine Learning for Age Determination in Culex pipiens

Problem Statement
Culex pipiens is a primary vector for arboviruses such as West Nile virus (WNV) and Usutu virus (USUV) in temperate regions. The efficiency of pathogen transmission is heavily dependent on vector survival, as only mosquitoes surviving the extrinsic incubation period (EIP) can transmit disease. Consequently, determining the age structure of field populations is critical for assessing transmission risk and evaluating vector control strategies. However, existing methods for age-grading Cx. pipiens, such as ovarian dissection, are labor-intensive, difficult to standardize, and provide only crude estimates of chronological age. While Mid-infrared spectroscopy (MIRS) combined with machine learning (ML) has been validated for tropical vectors like Anopheles and Aedes, its applicability to temperate Culex species—which exhibit longer lifespans and different physiological characteristics—remains untested.

Methodology
The study utilized a recently established laboratory colony of Cx. pipiens pipiens from Scotland. The experimental design involved:

• Sample Collection: Over 1,500 female and 500 male mosquitoes were sampled across three independent replicates. Females were monitored from 1–3 days post-emergence up to 79 days old, sampled at 11-day intervals. Males were sampled at 1–3, 11–13, and 22–24 days. To ensure physiological variability, mosquitoes were allowed to blood-feed and lay eggs, resulting in a mix of unfed, blood-fed, and gravid states.
• Spectroscopy: Individual mosquitoes were scanned using a Bruker Alpha II FT-IR Spectrometer with a diamond ATR crystal. Spectra were collected from the head and thorax across the 4000–400 cm⁻¹ range.
• Data Processing: Low-quality spectra were discarded. The remaining data were split into training (80%) and test (20%) sets.
• Machine Learning: Six algorithms (Logistic Regression, SVC, Random Forest, KNN, Gradient Boosting, and Multi-Layer Perceptron) were compared. The Multi-Layer Perceptron (MLP) classifier demonstrated the highest baseline performance and was selected for hyperparameter optimization via GridSearch.
• Resolution Levels: Models were trained to classify age at three resolutions:
  • Low: 2 age classes (1–35 days vs. 44–79 days).
  • Medium: 4 age classes.
  • High: 8 age classes (11-day intervals).
• Validation: An independent "unseen" dataset of 202 females from a separate rearing batch was used to test model generalizability and prevent overfitting. Permutation importance was calculated to identify key biochemical wavenumbers.

Key Results

• Model Performance (Seen Data): The MLP model successfully classified age classes, with accuracy inversely proportional to resolution.
  • Females: Balanced accuracy was 92.8% (low), 88.9% (medium), and 79.0% (high).
  • Males: Balanced accuracy was 82.5% for three age classes.
  • Misclassifications predominantly occurred between adjacent age classes, suggesting that biochemical changes over short intervals are subtle.
• Model Performance (Unseen Data): When applied to the independent dataset, accuracy decreased but remained significantly above random chance.
  • Females: Accuracy dropped to 80.9% (low), 46.9% (medium), and 34.0% (high).
  • Notably, the high-resolution model retained 83% accuracy in identifying newly emerged mosquitoes (1–3 days old).
• Biochemical Features: Permutation importance analysis indicated that model learning relied heavily on the fingerprint region (1300–600 cm⁻¹). Specific regions of interest included C–H and C–O stretches, amide I and III regions, and a distinct concentration of importance between 2162–2100 cm⁻¹ for low-resolution models.

Significance and Claims
The paper claims to present the first successful application of MIRS for age-grading within the genus Culex. The study demonstrates that MIRS can distinguish age classes in Cx. pipiens with sufficient precision to support epidemiological risk assessment, particularly for distinguishing newly emerged mosquitoes from older, potentially infectious vectors.

The authors emphasize that while the technique shows promise for evaluating vector control interventions (e.g., monitoring population age structure post-intervention) and understanding vectorial capacity, the current models are derived from laboratory colonies. The paper modestly notes that accuracy declines with higher resolution and when transitioning to unseen data, highlighting the need for future validation in field populations and the inclusion of genetically diverse samples to improve transferability. The work establishes a proof-of-principle that MIRS-ML can be adapted from tropical to temperate vector settings, offering a non-destructive (for the abdomen, though destructive for the head/thorax) and high-throughput alternative to traditional dissection methods.