摘要： Genetic, lifestyle and environmental factors, along with development and aging impact molecular composition of human blood. Although many diseases leave their trace in blood, the question is whether this trace can be robustly and reproducibly measured and used for health monitoring of a given adult population. Infrared molecular spectra of blood serum can be obtained in a non-invasive, time- and cost-efficient manner, delivering molecular information from all molecular species within the highly complex samples. We demonstrate that broadband infrared spectroscopy can be used for reproducible molecular fingerprinting of human blood. To evaluate whether certain medium is sufficiently robust to facilitate detection of disease onset, the quantitative extent of variability of a person as well as a reference population needs to be evaluated. If within-person variability would exceed that of the between-personal variability in a reference population, the approach would not be suited for disease detection. To assess the extent of uniqueness of infrared molecular fingerprints as well as their biological variability, we performed a comprehensive prospective longitudinal study collecting blood samples of 27 healthy individuals donating blood at 8 consequent intervals. We apply broadband infrared molecular fingerprinting by Fourier transform infrared spectroscopy (FTIR) and analyse between-person and within-person variability based on all different molecular classes in the blood simultaneously. We report experimental evidence of the feasibility of identifying a person within a group of individuals based on her/his infrared molecular fingerprint, similarly to metabolic fingerprints [1]. In a first step, using standard methods for descriptive analysis we observe that the between-person variability is larger than the within-person variability by a factor of 3 (Fig.1 Left). This observation opens up the possibility for disease detection. In a second step, we combine standard dimensionality-reduction methods, such as principal component analysis (PCA), and several high-accuracy machine-learning algorithms [1] (random forests, extreme gradient boosting, k nearest neighbours) for deriving classification rules, which we then use for making predictions on test sets of unseen data. For a group of 7 donations that span a period of 6 weeks, we reach peak classification accuracy of above 95% (Fig.1 Right), while the accuracy of a classifier predicting in random would have been as low as 3.7%. In addition, we evaluate the underlying spectral features with respect to their importance on signalling separation and in this way identifying the human blood serum constituents associated with between-person variation. Observed robustness of infrared molecular fingerprints suggests their applicability for health and treatment monitoring.