The significance of the advance lies partly in the mosquito genus involved. Culex species have received far less attention from spectroscopy-based methods than Anopheles mosquitoes, which dominate malaria research, despite the fact that Culex vectors are increasingly important players in the transmission of orthoflaviviruses and filarial parasites. Because these pathogens typically infect multiple host species, including birds, mammals and humans, mapping the feeding patterns of Culex mosquitoes is essential for predicting where and when spillover into human populations is likely to occur. Traditional blood meal identification relies on techniques such as PCR-based assays or sequencing, which are accurate but costly, time-consuming and progressively less reliable as the blood meal is digested inside the mosquito. A method that is rapid, inexpensive and robust to digestion time could transform the scale at which blood meal surveillance is feasible, particularly in the low-resource settings where mosquito-borne disease burdens are often highest.

The experimental design was straightforward but rigorous. The team reared Culex quinquefasciatus mosquitoes in the laboratory and fed them blood from six different host species: human, pig, chicken, cattle, dog and horse. This panel deliberately spans the range of hosts that matter most for multi-host pathogens in many endemic regions, from domestic animals that amplify viruses to humans who suffer the consequences. At four time points after feeding, namely 6, 12, 24 and 30 hours, the researchers rolled the abdominal contents of the mosquitoes onto filter paper to create dried bloodspots, a simple and field-friendly way of preserving samples without refrigeration or elaborate cold chains. Each bloodspot was then scanned using a Bruker ALPHA Fourier-transform infrared spectrometer, an instrument that measures how the sample absorbs mid-infrared light across a spectrum of wavelengths.

The underlying physics is what gives the method its power. Mid-infrared light interacts with the vibrational modes of chemical bonds in proteins, lipids, carbohydrates and nucleic acids, producing a complex absorption spectrum that acts as a molecular fingerprint of the sample. Because the blood of different host species differs subtly in its biochemical composition, those fingerprints carry enough information to distinguish a chicken blood meal from a dog blood meal, or a human meal from a horse meal. The raw spectra, however, are far too intricate for visual interpretation, which is where machine learning enters. The researchers trained a multilayer perceptron, a type of artificial neural network, on the spectra from laboratory mosquitoes with known blood meal origins, allowing the algorithm to learn the spectral patterns that discriminate between host species and then apply that learned knowledge to unknown samples.

The results were striking. On laboratory-reared mosquitoes, the trained model predicted host species with 93.6 percent accuracy, a dramatic improvement over the 16.7 percent expected from random classification among six host species. Even more importantly for real-world application, accuracy declined only modestly with digestion time, falling by just 6 percent at the 30-hour mark. This resilience matters because mosquitoes caught in surveillance traps are rarely fresh; they may have fed a day or more earlier, and conventional molecular methods lose accuracy as digestive enzymes break down the host DNA they rely on. A technique that remains reliable across a full day of digestion therefore captures a much larger and more representative fraction of the blood meals collected in the field, reducing the bias that plagues traditional approaches.

The true test came with field-collected mosquitoes. When the laboratory-trained model was applied to bloodspots from wild-caught Culex mosquitoes, accuracy dropped to 67.9 percent. The authors attribute this gap to differences between laboratory and field conditions rather than any fundamental flaw in the approach. Wild mosquitoes feed directly on living hosts, whereas the laboratory mosquitoes were fed through artificial feeders, and the blood used in the laboratory was treated with anticoagulants that wild blood meals naturally lack. Wild blood meals also reflect the true biochemical diversity of free-ranging animals, whose diet, health and physiology vary far more than those of the standardized blood sources used in the laboratory. In other words, the model was trained on a simplified version of the world and then asked to interpret a messier one, and the performance drop is an expected consequence of that mismatch.

Crucially, the team found a practical way to manage this uncertainty. By applying a probability threshold to the model’s predictions, accepting only classifications in which the algorithm expressed high confidence, accuracy on field samples rose to 78 percent, at the cost of discarding 21 percent of samples that the model could not confidently assign. This trade-off is a familiar one in applied machine learning, and for surveillance purposes it is often the right one. A public health programme that needs to know which hosts are driving transmission can afford to leave a fifth of samples unclassified if the remaining classifications are substantially more trustworthy. The threshold approach effectively lets users dial between coverage and confidence depending on the question at hand, a flexibility that rigid molecular assays do not offer.

The economic and logistical implications are considerable. Once the spectrometer is in place, the marginal cost of scanning a bloodspot is minimal, and the analysis requires no reagents, no cold chain and no highly specialized molecular biology laboratory. Filter paper bloodspots can be collected in remote villages, shipped at ambient temperature and analysed centrally, which makes the method particularly attractive for large-scale surveillance programmes in low-resource settings. The study’s field collections in the Philippines, supported by the Department of Health and local partners, alongside work in Australia’s Northern Peninsula Area, demonstrate that the sampling workflow is feasible in exactly the kinds of environments where Culex-borne pathogens impose their greatest burden. The work was funded by the Wellcome Trust, the European Research Council and several university and institutional grants, reflecting a collaboration spanning the University of Glasgow, the University of Sydney, NSW Health Pathology and the Ifakara Health Institute.

There are, of course, limits to what the current model can do. The laboratory training set covered only six host species, whereas real mosquitoes bite a far wider menagerie, including wild birds, rodents, bats and marsupials, and the model cannot classify hosts it was never taught to recognize. Expanding the library of host species, incorporating spectra from wild-fed mosquitoes into the training data, and refining the preprocessing of spectra to correct for the chemical differences introduced by anticoagulants and artificial feeders are all obvious next steps. The authors note that further development using samples from field-collected mosquitoes could enhance classification accuracy and broaden the range of identifiable hosts, and the open-access publication of the work should accelerate those efforts across the research community.

Even in its present form, however, the study marks a meaningful shift in how blood meal analysis could be conducted. By demonstrating that a benchtop infrared spectrometer and a neural network can recover host identities from a day-old blood meal with near-laboratory accuracy, the researchers have opened a path toward blood meal surveillance at a scale that PCR-based methods could never economically achieve. For diseases that move between animals and humans along cryptic transmission chains, that scale is precisely what has been missing. If subsequent studies confirm and extend these results across more host species and more field settings, the humble dried bloodspot, read by a beam of infrared light, may become one of the standard tools of vector-borne disease intelligence, helping health authorities see not just which mosquitoes are present, but whose blood they are carrying and which transmission pathways need to be interrupted first.

Subject of Research: Mid-infrared spectroscopy and machine learning for identifying host species of Culex mosquito blood meals

Article Title: Rapid host species identification from dried bloodspots of Culex mosquito blood meals using mid-infrared spectroscopy

Article References: Rapid host species identification from dried bloodspots of Culex mosquito blood meals using mid-infrared spectroscopy. (n.d.). https://doi.org/10.1186/s13071-026-07703-8

Image Credits: AI Generated

DOI: 10.1186/s13071-026-07703-8

Keywords: Culex mosquitoes, blood meal analysis, mid-infrared spectroscopy, machine learning, multilayer perceptron, vector-borne disease, orthoflaviviruses, lymphatic filariasis, dried bloodspots, disease surveillance, one health, entomology