This paper demonstrates that machine learning (ML) approaches coupled with high throughput real-time digital PCR (dPCR) can be used to increase detection accuracy of multiplex PCR assays when screening clinical isolates for the presence of carbapenemase-producing organisms (CPOs). We used a recently reported ML method called Amplification and Melting Curve Analysis (AMCA), which leverages the target-specific information encoded in each amplification event (via real-time data), to identify the nature of nucleic acid molecules (Moniri et al., 2020a). The AMCA approach is based on training supervised machine learning algorithms to extract kinetic and thermodynamic information from PCR amplification and melting curves to enhance the classification accuracy in multiplexing. Validation of this methodology using clinical isolates has never been reported before; therefore, this work represents a step forward towards the implementation of this method into clinical microbiology laboratories. Nucleic acid amplification tests (NAATs) that incorporate the AMCA classifier for multiple target detection will greatly improve their specificity, sensitivity and turn-around time to result, reducing overall resource consumptions and improving diagnostic performance.

Antimicrobial resistance (AMR) is a serious global threat and poses a challenge for modern medicine, compromising effective infectious disease management (Bush and Fisher, 2011; Tzouvelekis et al., 2012). One of the most concerning forms of AMR is the rapid spread of CPOs; bacteria producing enzymes that inactivate the potent antibiotics, carbapenems. Whilst overall United Kingdom incidence is low, there are centres nationally facing increasing rates and outbreaks, including Imperial College Healthcare NHS Trust (ICHNT), and it is endemic in many other regions worldwide (Otter et al., 2017b; Rodriguez-Manzano et al., 2020). CPO infections are associated with higher morbidity and mortality than susceptible strains, in part because their resistance can lead to ineffective empirical therapy and suboptimal treatment (Neuner et al., 2011; Eliopoulos et al., 2014). Therapeutic options are severely restricted, and in many cases clinical management relies on “last line” antibiotics that are less effective and have more side effects (Bleumin et al., 2012).

Patients infected with CPOs present significant challenges for diagnostics and infection control. There is an urgent need for accurate and timely diagnosis to improve patient outcomes and prevent the spread of AMR. Carbapenemase resistance genes are often co-localised on highly transmissible plasmids and are readily shared between bacterial species, providing the ideal conditions for multidrug resistant organisms (Johnning et al., 2018). Incorrect diagnosis delays appropriate intervention, increases financial burdens for the healthcare system, and complicates antimicrobial stewardship efforts (Charani et al., 2021). A local ICHNT economic analysis estimated the cost of a large hospital outbreak (−100 infections) of carbapenemase producing Klebsiella pneumoniae to be £1M. Some of the increased expenditure was associated with increased screening, bed closures, medication and patient bed-days (Otter et al., 2017a); better diagnostics could reduce these costs.

Diagnosis of CPOs is often too complicated and time-consuming, as it is normally based upon multiple tests which employ a wide range of instruments and diagnostic tests. Phenotypic methods typically target carbapenemase production and provide no information on the underlying resistance mechanism (Codjoe and Donkor, 2017). These tests represent a low-cost (£2–15 per sample) and robust methodology; however, they rely on pure culture which increases turnaround times (12–24 h) (Moloney et al., 2019). A variety of molecular methods, including amplification (PCR-based), microarray and sequencing assays have been developed and are frequently used in microbiology laboratories (Matsumura and Pitout, 2016; Reta et al., 2020). Microarray and sequencing are time consuming (>12–48 h), expensive (>£50K platforms and >£80 per sample), and require bioinformatic expertise. Conversely, NAATs are commonly cheaper (£15–30 per sample) and faster (1,2 h), whereas instrument price significantly ranges between tens to hundreds of thousands of pounds for conventional and digital PCR platforms, respectively (Huggett et al., 2015; Quan et al., 2018). Furthermore, the application of sophisticated data processing for its optimisation (as done with microarray and sequencing methods) has been largely unexplored (Collins and Moons, 2019; Beinhauerova et al., 2020). As a result of all aforementioned limitations, implementation of microarrays, sequencing and molecular methods for CPO diagnosis into routine practice is often limited.

Recently, our group has demonstrated that the large volume of data obtained from real-time digital PCR (dPCR) instruments can be exploited to perform data-driven multiplexing in a single fluorescent channel, reporting a 99.33 ± 0.13% classification accuracy when using synthetic DNA in a 9-plex format (Moniri et al., 2020a). This result represented an increase of 10% over using melting curve analysis, indicative of the potential benefits of this methodology for diagnostic and screening applications. The ML method used (AMCA) leverages kinetic and thermodynamic information encoded in the amplification and melting curves to perform target identification in multiplexed environments (Moniri et al., 2019; Rodriguez-Manzano et al., 2019; Moniri et al., 2020b). Here we evaluate, for the first time, the analytical performance of AMCA method compared to Xpert Carba-R Cepheid and Resist-3 O.K.N assays when tested on clinical isolates for detection of the most common types of serine-beta-lactamases (bla_KPC and bla_OXA-48) and metallo-beta-lactamases (bla_IMP, bla_VIM and bla_NDM) (Maurer et al., 2015; Lim et al., 2018). Results were compared against another ML based classifier ‘Melting Curve Analysis’ (MCA), which uses the thermodynamic information contained in PCR melting curves for identification of multiple targets in a single well reaction (Athamanolap et al., 2014; Moniri et al., 2020a). A 5-plex PCR assay was developed in-silico and validated with synthetic DNA templates. The performance of the AMCA method, using this 5-plex, was further assessed with 253 clinical isolates provided by the microbiology department at Charing Cross Hospital, ICHNT. All samples were analysed in real-time dPCR, using an intercalating dye (EvaGreen) in a single-fluorescent channel. This work demonstrates that the AMCA method can be integrated with conventional clinical diagnostic workflows in combination with real-time dPCR platforms, as it does not require any hardware modification. Increasing multiplexing capabilities enables improved workflow efficiency while reducing per sample cost, and it is beneficial to a number of application fields beyond clinical diagnostics, such as veterinary and environmental fields, where multiple targets need to be analysed simultaneously (e.g., SNP genotyping, forensic studies and gene deletion analysis). Figure 1 illustrates the concept of data-driven multiplexing, where tailored PCR-based amplification chemistries combined with advance data analytics can be seamlessly integrated into existing diagnostics pipelines which utilize real-time platforms.

In the last decade, novel pandemic outbreaks and the continued threats of emerging multi-drug resistant microorganisms have significantly increased the demand for molecular tests, in particular PCR-based methods (Nishizawa and Suzuki, 2014; Vasala et al., 2020). To respond to this need, the AMCA technology has been designed to increase the throughput of real-time molecular platforms. Seamlessly integrated with conventional diagnostic workflows, this machine learning based approach can enhance multiplexing capabilities of traditional qPCR and state-of-the art dPCR instruments, increasing the number of nucleic acid targets that can be identified in a single fluorescent channel without hardware modifications. Individual primer sets produce amplification products at a sequence-specific amplification rate and efficiency, which generate unique amplification and melting curves for different target concentrations. Such curves can be capture as time-series data by real-time instruments, feed into machine learning models and used to identify multidimensional patterns (or signatures) specific to each primer set. Therefore, enabling the identification of multiple DNA targets per fluorescent channel using only real-time data (i.e., data-driven multiplexing). In this paper, we performed a clinical validation on diagnostic accuracy of the AMCA methodology by targeting the “big five” carbapenem-resistant genes (bla_VIM, bla_OXA-48, bla_NDM, bla_IMP and bla_KPC) in multiplex PCR. A 5-plex PCR assay was developed and characterised in both real-time qPCR and dPCR instruments, and the AMCA performance investigated through the identification of 253 clinical isolates from patients’ samples. The MCA was used as a reference method to compare results.

We successfully show a 99.2% accuracy for identifying the five carbapenem-resistant genes in the clinical isolates. The AMCA method was shown to enhance the classification performance by 7.9% compared to MCA. The AMCA takes advantage of the volume of raw data extracted from amplification and melting curves, whereas the MCA only considers melting curves. It is interesting to observe that the overlapping melting curve distribution in Figure 3B (e.g., bla_NDM and bla_KPC) represents a misclassification of 1,303 reactions (509 bla_KPC as bla_NDM, and 804 bla_NDM as bla_KPC) and 21 clinical isolates (20 bla_NDM and 1 bla_KPC as coinfections) when using the MCA, but it only represents a misclassification of 99 reactions and 1 clinical isolates for the AMCA method. As described in previous publications (Moniri et al., 2020a), these results support the hypothesis that the underlaying biological factors driving these methods for target identification are fundamentally different. As observed in Supplementary Figure S10, machine learning methods can be used to exploit the distinctive information contained on the amplification and melting curves by weighting the predictions from the ACA and MCA to optimally combine them and maximize the AMCA performance.

Although dPCR is not likely to replace all qPCR assays in the clinical laboratory due to associated instrument costs and greater complexity, it has several specific advantages over qPCR. The vast number of partitions reduce the likelihood of coamplification and inhibitors in a single reaction, facilitating accurate detection of multiple analytes; and the large amount of data enables the use of advance machine learning algorithms to detect subtle kinetic and thermodynamic differences encoded in the real-time amplification data. On the other hand, real-time dPCR platforms enable the use of digital bulk standards and offer a valuable solution for absolute quantification of clinical isolates (equivalently to conventional qPCR standards) even when the panels are saturated, expanding the dynamic range of quantification of the microfluidic chips and eliminating the need of testing the samples at multiple dilutions to ensure that at least one of them falls within the conventional dPCR range (i.e. panels at occupancy <85%). As shown in Figure 2, it is possible to create a standard curve in real-time dPCR by extracting C_q values as a function of the target concentration because there is a clear separation between the single-molecule and the bulk regions. We envision that coupling real-time dPCR instruments with data-driven multiplexing will expand the use of these platforms in clinical microbiology laboratories.

The results presented in this study represent a step forward in the use of PCR-based data-driven diagnostics for clinical applications. However, there are several aspects that need to be further investigated. Firstly, in this paper we evaluated the performance of AMCA method in clinical isolates using pure bacterial cultures, therefore a follow-up study needs to be conducted to evaluate the performance of the method directly from clinical samples. Secondly, it is important to identify co-presence of infections for patient treatment, however in this paper we address only one sample with a double infection; a larger study will be required to test the effectiveness of the AMCA in double pathogen identification. Depending on the sample concentration, this might not limit multiplexing capabilities in dPCR, but it could represent a challenge when qPCR instruments are used.

This work suggests that the AMCA approach provides a versatile solution for the accurate detection of AMR genes, representing a cost-effective interaction as it does not require hardware modifications. This study highlights the importance of integrating artificial intelligence for diagnosis and how effectively it increases result reliability of state-of-the-art dPCR instruments. Moreover, the AMCA methodology has the potential for further application in point-of-care devices and isothermal chemistries, as a solution to leverage identification accuracy and enable faster detection of multiple pathogens.