Abstract Background Tuberculosis (TB) is an infectious disease that induces a complex response from the host immune system. Most carriers of the bacteria experience the asymptomatic, non-infectious phase, latent tuberculosis infection (LTBI), but have the potential to reactivate to active, contagious TB infection. Current diagnostics for TB are not sufficient for LTBI and are focused on using one biomarker, interferon-gamma (IFN-g). To build better diagnostics for the complex disease, multi-biomarker approaches to profile the immune system, coupled with machine learning algorithms, are necessary to diagnose LTBI and determine patients' risks of reactivation to active TB. We aim to use a multiplexed silicon photonic microring resonator sensing platform to profile thirteen biomarkers in patients with LTBI infection and use machine learning data analysis to identify biomarkers relevant for diagnosing LTBI and predicting reactivation risk. Methods We employed microring resonators coupled with a sandwich-style detection format to simultaneously detect thirteen cytokines and chemokines in plasma in under 40 minutes. Chip-integrated silicon photonic microrings are small in size and easily multiplexed with up to sixteen different capture agents. Biomolecular binding events are measured by monitoring the resonant wavelength within the sensing cavity, which shifts as binding events occur. Quantitation of each biomarker is possible with calibration curves that correlate standard protein concentrations to wavelength shifts. The multiplexed assay has been optimized for each individual biomarker, tested for cross-reactivity, and calibrated in the plasma matrix of interest. The LODs vary, depending on biomarker and matrix dilution, but range from 4.9 pg/mL to 691 pg/mL. The patient plasma samples are left over from the TB screening IFN-g release assay, QuantiFERON-TB Gold Plus (QFT), putting this method within the current clinical workflow. We use precision normalization approaches from the four QFT stimulations to account for immunologic differences among the subjects. The biomarker concentrations are evaluated using a random forest machine learning model to identify which biomarkers are important for identifying patients with LTBI and predicting their reactivation risk. Results We used this sensor method to profile 42 patients, 24 LTBI negative controls and 18 LTBI positive controls, with 13 being considered at high risk of reactivation. Using a precision normalization approach, a combination of nine normalized conditions using five of the thirteen biomarkers (CCL4, CCL8, IP-10, IL-2, IL-17) discriminated between LTBI positive and negative subjects with a ROC AUC of 0.90. Additionally, eight normalized conditions using four of the same biomarkers discriminated between high and low-risk subjects with an AUC of 0.83. Currently, we have expanded the sample cohort to include an additional 72 subjects, 25 LTBI positive and 47 LTBI negative. Preliminary analysis shows CCL8, IL-2, and IP-10 alone are able to distinguish between LTBI status (P values <0.01) using raw target concentrations prior to precision normalization. Conclusion Overall, we show the importance of multi-biomarker approaches to develop diagnostic and prognostic tools for a complex disease, such as TB. Further work to use the initial machine learning algorithm with the expanded cohort to validate biomarker importance is underway.
Read full abstract