Rapid host species identification from dried bloodspots of Culex mosquito blood meals using Mid-Infrared Spectroscopy

Technical Summary: Rapid Host Species Identification from Culex Mosquito Blood Meals Using Mid-Infrared Spectroscopy

Problem Statement
Understanding mosquito blood-feeding preferences is critical for inferring vector-borne pathogen transmission dynamics and optimizing intervention strategies. While Culex mosquitoes are primary vectors for high-priority multi-host pathogens (e.g., Japanese encephalitis virus, West Nile virus, filariasis), their feeding patterns remain understudied compared to genera like Anopheles. The current "gold-standard" methods for identifying blood meal sources—DNA-based techniques (PCR, sequencing) and antigen-based assays (ELISA)—are often expensive, time-consuming, require specialized laboratory capacity, and suffer from reduced accuracy as blood meals are digested or degraded. There is a specific need for a rapid, cost-effective, and scalable alternative that functions effectively in low-resource settings and can handle samples from Culex species, which have not previously been analyzed using this approach.

Methodology
The study employed a two-phase approach to develop and validate a Mid-Infrared Spectroscopy coupled with Machine Learning (MIRS-ML) pipeline using Dried Blood Spots (DBS).

1. Laboratory Model Development:

   • Sample Generation: Laboratory-reared Culex quinquefasciatus were fed blood from six host species: human, pig, chicken, cattle, dog, and horse. Blood was collected via a membrane feeding system using anticoagulated blood.
   • DBS Creation: Abdominal contents were rolled onto Whatman™ 903 Proteinsaver cards at 6, 12, 24, and 30 hours post-feeding to simulate digestion.
   • Spectroscopy: Samples were scanned using a Bruker ALPHA Fourier-transform infrared (FTIR) spectrometer (400–4000 cm⁻¹ range).
   • Data Processing: Spectra were pre-processed to remove regions affected by the card substrate (900–1100 cm⁻¹) and atmospheric CO₂ (1800–2200 cm⁻¹).
   • Machine Learning: Seven supervised ML algorithms were tested. A Multilayer Perceptron (MLP) was selected as the optimal model after hyperparameter tuning via grid search and 5-fold cross-validation. The model was trained on 80% of the data and validated on a held-out 20% test set.

2. Field Validation:

   • Collection: Blood-fed mosquitoes were collected from field sites in the Philippines (Central Luzon and Cordillera) and Australia (Queensland) using CO₂-baited traps.
   • Reference Standard: Host identification for field samples was performed using DNA extraction, universal vertebrate PCR (targeting COI), and Sanger sequencing.
   • Application: The laboratory-trained MLP model was applied to the FTIR spectra of field-collected DBS to predict host species.
   • Probability Thresholding: The study evaluated the use of model probability scores to filter low-confidence predictions, aiming to improve accuracy by excluding uncertain samples.

Key Results

• Laboratory Performance: The optimized MLP model achieved 93.6% accuracy in predicting host species from laboratory-reared Cx. quinquefasciatus DBS, significantly outperforming random classification (16.7%). Accuracy varied slightly by host (89% for dog to 97% for pig) and declined modestly at the 30-hour digestion mark (OR = 0.32 for correct classification relative to earlier timepoints).
• Field Performance: When applied to field-collected samples with known hosts (validated by PCR), the model achieved 67.9% accuracy. This lower performance is attributed to differences between lab and field conditions, including the absence of anticoagulants in field samples, direct feeding from live hosts rather than membranes, and unknown digestion times.
• Probability Thresholding: Applying a probability threshold of 0.73 to field samples improved classification accuracy to 78% for known hosts, though this required filtering out 21% of the samples (40 samples).
• Generalizability: The model, trained on Cx. quinquefasciatus, successfully predicted host species for other genera (e.g., Aedes and Armigeres) collected in the field, correctly identifying 6 out of 10 non-target species.
• Comparison to PCR: In the Philippines field dataset, 29.6% of samples failed to amplify via PCR, whereas MIRS-ML provided predictions for all samples, suggesting higher utility for degraded or late-stage digestion samples.

Key Contributions

• First Application to Culex: This study represents the first application of MIRS-ML to identify blood meal sources in Culex mosquitoes, a genus of increasing concern for vector-borne diseases.
• DBS Utility: It demonstrates the viability of using Dried Blood Spots on filter paper as a stable, room-temperature storage medium for MIRS analysis, overcoming preservation challenges associated with dried mosquitoes.
• Scalability: The method offers a rapid alternative to sequencing; scanning 203 DBS took one workday, compared to 15 workdays for the gene sequencing workflow.
• Cost-Effectiveness: The approach utilizes a single instrument (FTIR spectrometer) with no ongoing reagent costs, making it suitable for low-resource settings.

Significance and Claims
The paper claims that MIRS-ML with DBS is a practical, scalable tool for blood meal analysis that can improve the understanding of host-vector interactions for pathogens like Japanese encephalitis virus. The authors posit that while field accuracy is currently lower than laboratory accuracy, the method's ability to function without specialized molecular infrastructure and its success with degraded samples make it a valuable addition to surveillance programs. The study suggests that further model development, specifically incorporating field-collected samples into training sets and expanding the range of identifiable hosts, could enhance classification accuracy. The authors conclude that this method holds promise as an economical alternative to PCR for large-scale vector-borne disease surveillance, particularly in regions where pathogen transmission dynamics are shifting due to climate change and globalization.