GPCR-Based Machine Olfaction On Urine Scent Surpasses PSA at Predicting Prostate Cancer

Objectives. To determine whether medical machine olfaction via tracking the activation of mammalian G-Protein Coupled odorant Receptors (GPCR) stabilized by proprietary co-polymers on a photonic MZI chip can be used to diagnose prostate cancer (PCa) via urine scent. Specifically, scent character is compared against the current diagnostic PCa screening gold-standard in the US: the serum level of prostate specific antigen (PSA). The device is an artificial nose sensor built on a commercial photonic platform that reads interchangeable Mach-Zehnder interferometer (MZI) chips. These chips were functionalized with a stabilised panel of mammalian olfactory G-protein-coupled receptors (GPCRs). These samples had been characterized into POSITIVE or CONTROL for PCa six to eight years prior by standard hospital diagnostic procedures and by trained medical detection dogs, then stored at -80 Celcius. A subset of 80 patients urine samples was subsequently thawed and used for training and testing the medical machine olfaction system of RealNose as an initial validation of the novel technology and methodological approach. We posed two primary research questions: (a) whether the cancer-associated odor profile would remain detectable by machine-based systems following long-term storage and with what accuracy could it be used to cluster (YES and AUC 0.79 from scent character alone), and (b) what technical and procedural requirements would be necessary to translate such a signal into a clinically useful diagnostic assay (more training samples (500 predicted to yield 0.93) and increased breadth of receptors per chip and/or more chips per device in next iteration seen as helpful). Design, setting, participants. Retrospective diagnostic-accuracy feasibility study on 80 biobanked urine samples (40 PCa, 40 non-cancer; 368 sensor runs; a subset of unknown Gleason grade) from a single UK NHS urology service, the same collection used to train canine detectors. Main outcome measures: Patient-level Receiver Operating Characteristic (ROC) area under the curve (AUC) under patient-grouped cross-validation with a fold-honest pooled-control reference (reconstructed from training-partition controls only); sensitivity, specificity and predictive values at pre-specified operating points; 1000-fold whole-procedure label-permutation significance; patient bootstrap 95% CIs; and leave-one-day-out / leave-one-chip-out generalisation. Results. An L2-regularised linear classifier when allowed to see between three and six chips outcome on a patient sample extracted within-instrument AUC 0.79 (95% CI 0.69 to 0.88; 1000-permutation p = 0.001) from urine scent alone, exceeding this cohort own serum prostate-specific antigen (PSA) discrimination (AUC 0.645; itself within the population range for PSA 0.67) and obtained without a blood draw (at the Youden point, sensitivity 0.75, specificity 0.78, PPV 0.77, NPV 0.76). Upon allowing PSA the total AUC rose to 0.82. This was not a plateau: AUC rose from chance at 30 training samples, passed the serum-PSA range at 40, and reached 0.79 at 80 patients (0.82 if PSA was included), with an inverse-power fit projecting 0.93 by n = 500 and 0.96 by n = 1000. The discriminant was a genuine multivariate receptor pattern, independent of patient age (Spearman 0.09; the cohort is not age-matched). So at least for these data, neither age, nor collection day, ambient humidity/temperature, or overall signal amplitude (sometimes thought of as intensity of smell) were predictive of prostate cancer status, yet the scent character was. Transfer to a new sensor chip fell to AUC 0.57 without calibration, meaning the remaining obstacles are hardware portability rather than signal existence: much as a detection dog acclimatizes to a new setting, the system improves with on-site calibration prior to use. Conclusions: A genuine, confound-controlled olfactory PCa signature is recoverable from 80 samples, surpasses this cohort serum PSA (0.645) and exceeds the population PSA range, and improves monotonically with training-set size. We present this as a small-sample feasibility benchmark, not yet a validated diagnostic; the dominant remaining factor is training-set size, and the path to clinical-utility and improved AUC is clearly found to be a larger, multi-site, age-matched, and ideally prospective training cohort. A transferable small-sample lesson is also reported: adaptive feature searches (evolutionary and self-calibrating-protocol handle search) artificially inflate cross-validation and collapse under whole-procedure permutation, whereas non-adaptive averaging survives, giving a robust scent signal obtainable from the headspace of urine samples and recordable by the RealNose device that keeps improving with expanding sample training set.