Colorectal cancer (CRC) is one of the leading causes of cancer death worldwide. In Japan alone, approximately 50,000 people die from CRC annually, and the number is rising. The disease is most treatable when caught early, but it often produces no noticeable symptoms in its initial stages, making timely diagnosis challenging.
Current standard diagnostic tools each have significant drawbacks. Colonoscopy is highly accurate but invasive, expensive, and resource-intensive. The fecal occult blood test (FOBT), while non-invasive and cheap, has limited sensitivity (50% to 75%) and relies on manual implementation, introducing variability. These limitations have driven growing interest in liquid biopsy approaches that could screen for cancer using easily obtained body fluids.
Liquid biopsy analyzes tumor-derived information from fluids such as blood, urine, or saliva, without requiring direct tissue sampling. The ideal CRC liquid biopsy would be non-invasive, inexpensive, automatable, and sensitive enough to detect cancer at early stages when treatment is most effective. This study explores a novel approach using biological sensors modeled on the insect sense of smell.
Cancer cells have distinct metabolic programs compared to normal cells. One consequence of these altered metabolic pathways is the production of specific volatile organic compounds (VOCs), small molecules that can evaporate and enter bodily fluids. VOCs from cancer metabolism have been detected in blood, urine, and exhaled breath, and different cancer types appear to produce different VOC signatures, suggesting they could be used for cancer-type-specific diagnosis.
Traditionally, VOCs are measured using instruments such as gas chromatography-mass spectrometry (GC-MS), which can identify and quantify specific molecules with high precision. However, GC-MS instruments are expensive, technically complex to operate, and poorly suited for routine clinical screening. There is a need for simpler, more scalable technologies that can detect relevant VOC patterns without requiring specialized analytical chemistry infrastructure.
An alternative approach inspired by biology is to use olfactory receptors (ORs) as VOC sensors. Insects have evolved an extraordinary diversity of olfactory receptors tuned to detect a wide range of volatile molecules, often at very low concentrations. By expressing these receptors in laboratory cells, researchers can create living biosensors that respond to VOCs in complex biological samples, even without knowing in advance exactly which molecules are relevant.
The biosensor in this study uses HEK293FT human cell lines engineered to express insect-derived olfactory receptors on their surface. When a VOC in the urine sample binds to the receptor, it triggers an influx of calcium ions into the cell. The system also includes a co-receptor called Orco and a bioluminescent protein called aequorin, so when calcium enters the cell, it triggers a luminescence signal that can be measured in real time.
The researchers assembled a library of 483 olfactory receptor genes derived from six insect species: Anopheles gambiae (a malaria mosquito, 130 ORs), Aedes aegypti (a dengue mosquito, 124 ORs), Drosophila melanogaster (the common fruit fly, 60 ORs), Bombyx mori (the silkworm, 36 ORs), Locusta migratoria (the migratory locust, 86 ORs), and Cimex lectularius (the bed bug, 47 ORs). Each OR was tested for its response to urine samples from CRC patients and healthy controls.
Luminescence was measured continuously for 4 minutes after each urine sample was added to the sensor cells, generating a time-series data profile for each OR-sample combination. This time-course data was then processed and used as input for machine learning models. Three technical replicates per OR-sample combination were performed to ensure reproducibility, with the median value used in analysis.
The study enrolled 75 CRC patients and 75 non-cancer controls, all of whom provided urine samples. Among the CRC patients, the cancer stages represented were: 1 patient at Stage 0, 22 at Stage I, 29 at Stage II, 14 at Stage III, and 9 at Stage IV. Both colon and rectal cancer patients were included in the CRC group. Patients ranged in age from their 30s to 80s, while controls ranged from their 20s to 50s.
Urine samples were frozen within 2 hours of collection and stored at minus 80 degrees Celsius. Before analysis, each sample was thawed, mixed with calcium and a chemical activator, centrifuged, and aliquoted into a 384-well plate format. Each sample underwent between five and seven freeze-thaw cycles, and stability data confirmed that OR responses were not substantially affected by this handling. All 150 samples were analyzed without exclusion.
To reduce variability between experimental batches, a critical design choice was made: all samples were assayed on the same plate for a given OR. This minimized the plate-to-plate batch effects that are common in high-throughput biological assays and helped ensure that differences in luminescence response truly reflected biological variation between patient samples rather than experimental artifacts.
From the initial library of 483 ORs, the researchers first identified a subset that responded strongly to human urine in general. A further analytical step divided these responding ORs into two categories: corrective ORs, which responded equally to cancer and non-cancer urine and were used to normalize for inter-sample variation, and CRC-discriminative ORs, which showed differential responses between the two groups.
This normalization strategy was an important methodological contribution of the study. Because individual urine samples naturally vary in composition due to diet, hydration, and other factors, raw luminescence signals can differ substantially between samples for reasons unrelated to cancer. By using the corrective ORs as internal references to calculate a correction coefficient for each sample, the researchers were able to reduce this noise and improve the signal-to-noise ratio for genuine CRC-related VOC differences.
After normalization, 65 ORs were classified as CRC-discriminative and used for machine learning analysis. For each of these 65 ORs, four types of features were generated from the time-series luminescence data, including raw smoothed signal, first-order derivative, second-order derivative, and time-series features extracted by a software tool called Tsfresh. This produced a total of 260 OR-feature datasets as input candidates for the diagnostic model.
The 260 OR-feature datasets were first ranked by their individual diagnostic performance using automated machine learning (AutoML) via Microsoft Azure, which tested multiple algorithms and identified the best-performing approach for each dataset. The top-ranked single OR-feature combination was AgOR47_variant6 with smoothed signal data (AUC 0.87). Multiple OR-feature combinations from Anopheles gambiae and Aedes aegypti mosquito receptors appeared in the top 13 rankings.
Using a forward stepwise selection process with LightGBM (a gradient boosting machine learning algorithm) and five-fold double cross-validation, the researchers identified the optimal combination of datasets for building the final diagnostic model. Performance improved as more OR-feature datasets were added, reaching a peak with 8 datasets combining responses from six different olfactory receptors. The final model used ORs from multiple mosquito species, reflecting the complementary VOC-detection capabilities of receptors from different evolutionary lineages.
The final diagnostic model achieved the following performance on held-out test data: accuracy 77%, sensitivity (recall) 80%, precision 77%, specificity 75%, F1-score 78%, and ROC-AUC 0.84. These results represent a substantial improvement over the fecal occult blood test benchmark of 50% to 75% sensitivity. The model also demonstrated good performance for early-stage CRC (Stages 0 and I), supporting its potential utility for screening applications where catching cancer early is most critical.
The choice to use insect olfactory receptors, rather than mammalian ones, is deliberate and strategically advantageous. Insects have evolved highly diverse olfactory receptor families tuned to detect a vast range of volatile molecules, including many at extremely low concentrations. With 483 receptor variants from six species tested in this study, the researchers were effectively casting a wide net to find receptors that happen to be sensitive to whatever VOCs CRC produces, without needing to know the identity of those VOCs in advance.
This agnostic VOC detection strategy is particularly powerful because the specific metabolic compounds that distinguish CRC urine from healthy urine remain incompletely characterized. Traditional analytical chemistry approaches require knowing what to look for. The biosensor approach sidesteps this limitation by using biological detectors that respond to the complex VOC mixture as a whole pattern, letting the machine learning model discover the discriminative signal empirically.
Five of the six ORs that contributed most to the final model were consistently selected across all five cross-validation folds (LmOR130, AaOR25, AgOR47_variant6, AgOR47_variant17, and AaOR33), with a high mean pairwise overlap score of 0.90. This consistency suggests that the model is relying on stable biological signal patterns rather than overfitting to random variation in the training data, which is an important indicator of the approach's genuine discriminative power.
The most important limitation is that the model was developed and tested using urine samples from a single institution. Single-institution studies are prone to institution-specific confounding factors such as diet, local environmental exposures, and patient population demographics. Before this approach can be considered for clinical use, validation in diverse multicenter datasets is needed to confirm that the discriminative ORs and the diagnostic performance generalize beyond the original study cohort.
The non-cancer control group was also younger on average (20s to 50s) than the CRC patient group (30s to 80s), which is a demographic mismatch that could introduce age-related differences in urinary VOC profiles as a potential confound. Future studies should age-match controls more carefully and also assess performance against patients with benign colorectal conditions such as polyps or inflammatory bowel disease, as these populations represent the clinically relevant differential diagnosis.
If these limitations are addressed and the approach is validated, OR-based urine biosensors could offer a genuinely transformative screening tool. Urine collection is completely non-invasive, the cellular assay format is compatible with laboratory automation, and the approach is potentially adaptable to other cancer types by screening for disease-discriminative ORs in different patient populations. The authors envision this technology as a new class of liquid biopsy tool capable of cancer-type-specific diagnosis from simple urine samples.