Design

1. Design Overview

Respiratory infections remain a major global health burden. The GBD 2023 analysis identified lower respiratory infections (LRIs) as the world's leading infectious cause of death, accounting for an estimated 2.50 million deaths in 2023, with the highest burden among children younger than 5 years and adults aged 70 years and older ^[1]. Meanwhile, the co-circulation of respiratory viruses such as SARS-CoV-2, influenza A/B, and respiratory syncytial virus (RSV) highlights the need for rapid and reliable multiplex testing ^[2,3]. Although conventional RT-qPCR-based workflows can provide sensitive and specific detection, their dependence on nucleic-acid extraction, complex instrumentation, and trained personnel can limit their application in decentralized screening and rapid outbreak response ^[2].

To address these limitations, we designed an integrated point-of-care detection platform combining recombinase polymerase amplification (RPA) with CRISPR-Cas12a. In our design, isothermal nucleic-acid amplification is intended to enrich low-abundance target sequences without the need for thermal cycling, while the CRISPR-Cas12a system provides programmable sequence recognition and collateral-cleavage-based signal generation ^[4]. By integrating these two processes, we aim to develop a workflow that combines rapid target amplification with sequence-specific detection while reducing dependence on complex laboratory equipment. Recent developments in one-pot isothermal amplification–CRISPR systems further support this strategy and provide a basis for simplifying the workflow toward rapid and sensitive point-of-care diagnostics ^[5,6].

However, simply combining RPA and CRISPR-Cas12a does not necessarily produce an optimal detection system. A major consideration in our design is therefore how to systematically optimize the molecular components and reaction conditions. One-pot integration of isothermal amplification with CRISPR can be affected by incompatibilities in reaction conditions and interactions between amplification and cleavage processes ^[5,6]. In addition, the selection of effective crRNAs for CRISPR-Cas12a-based diagnostics remains challenging and time-consuming. Recent deep-learning approaches provide an opportunity to prioritize candidate crRNAs and integrate crRNA selection with RPA-primer design, potentially reducing the experimental screening burden ^[7]. Based on these considerations, we designed our development strategy around a model-guided and iterative workflow rather than relying primarily on empirical trial-and-error.

Our overall design therefore integrates model-assisted sequence screening, enzyme-kinetic parameter estimation, and iterative Design–Build–Test–Learn (DBTL) cycles connecting dry-lab prediction with wet-lab validation. We selected five representative respiratory pathogens as iterative development targets. For each target, candidate molecular components and initial reaction parameters are designed computationally, evaluated experimentally, and subsequently refined according to experimental feedback. By repeating this process across multiple pathogen-specific systems, we aim to identify both target-specific requirements and generalizable design principles, thereby improving the transferability and robustness of the overall workflow.

Ultimately, our design is intended to establish a standardized development framework rather than a single pathogen-specific assay. The framework is designed to support three interconnected stages: target-specific molecular component design, rational initialization and iterative optimization of reaction parameters, and integration of the optimized assay with a portable fluorescence-detection device. Through this design, we aim to provide a systematic route from sequence selection and computational prediction to experimental optimization and point-of-care implementation.

2. Design Requirements

3. Biological Module Design

3.1 Target Selection

Our detection panel includes respiratory syncytial virus (RSV), human metapneumovirus-B (hMPV-B), Streptococcus pneumoniae (SPN), Bordetella pertussis (BP), and Haemophilus influenzae (HI). For each pathogen, assay design begins with a nucleic-acid region that is representative of the target species and relatively conserved within the population of interest. Candidate target regions must satisfy two key constraints: they should reduce the effect of within-species sequence variation on assay performance, while while being different enough from non-target respiratory pathogens. This provides the sequence basis for two layers of specificity: RPA amplification and crRNA-mediated Cas12a recognition.

3.2 Deep Learning–Assisted Ranking of RPA Primer–crRNA Combinations Using DNABERT-6 Sequence Representations

We use RPA as the upstream amplification module and couple it directly to CRISPR-Cas12a recognition. Because RPA primers and crRNAs together affect amplification efficiency, target specificity, and Cas12a activation, we do not design them as two separate parts. Instead, we incorporate DNABERT-6-derived sequence representations into a ranking model to prioritize candidate RPA primer–crRNA combinations.

Candidate RPA primers and matching crRNA target sites are first generated from the conserved, pathogen-specific regions identified in Section 3.1. RPA is selected because it amplifies nucleic acids under isothermal conditions, avoiding the denaturation, annealing, and extension cycles required for conventional PCR. During candidate generation, primer length, GC content, template matching, potential primer-dimer formation, intramolecular secondary structure, and off-target matching are checked. At the same time, the amplicon defined by the primer pair must contain a complete Cas12a-recognizable target sequence and satisfy the related PAM and local sequence requirements.

The candidate primer sequences, amplicon target regions, and matching crRNA spacer sequences are then encoded with DNABERT-6. In our workflow, DNABERT-6 does not replace standard sequence-design rules or wet-lab validation. Rather, it serves as a sequence-context representation module, whose embeddings are integrated with conventional design features and further processed by a ranking model to prioritize candidate primer–crRNA combinations.

This strategy reduces a large candidate set into a smaller set of top-ranked designs for experimental validation. Importantly, RPA amplification and CRISPR recognition are therefore designed together from the sequence-design stage. The primer pair defines the amplicon boundaries, while the crRNA spacer recognizes a pathogen-specific sequence within that amplicon. A candidate is kept only when the primer pair can support target amplification and the resulting amplicon contains an appropriate Cas12a recognition site.

During detection, RPA first enriches the target sequence under isothermal conditions. The resulting amplicon is then tested by the Cas12a-crRNA ribonucleoprotein complex. Only when the amplified sequence matches the matching crRNA and satisfies the Cas12a activation requirements is Cas12a trans-cleavage activated. RPA therefore provides target increase and signal amplification, whereas Cas12a–crRNA provides sequence-specific recognition of the amplified product, adding an additional verification step after amplification.

3.3 Fluorescence Reporting Module

The reporting layer uses a single-stranded DNA fluorescence reporter that can be cleaved by activated Cas12a. In the intact reporter, the fluorophore and quencher remain close together and fluorescence stays low. Once Cas12a is activated, trans-cleavage separates the fluorophore from the quencher, producing a fluorescence signal that can be captured by the optical module.

4. Dry-Lab Model Design

4.1 DNABERT-6-Assisted Candidate Prioritization

DNABERT-6 serves as the sequence representation component in our dry-lab workflow. For each pathogen, candidate primer–crRNA combinations are generated around conserved, species-specific target regions. The four sequence components, including forward primer, reverse primer, crRNA spacer, and amplicon region, are encoded by DNABERT-6 and integrated with conventional sequence-design features for candidate ranking.

This increases experimental efficiency while allowing different pathogen targets to be evaluated using a consistent sequence-prioritization framework. Only a limited number of high-priority primer-crRNA combinations are advanced to wet-lab testing. This increases the efficiency of experimental screening while allowing all five pathogens to be evaluated using a consistent sequence-prioritization framework.

4.2 Enzyme-Kinetic Modeling

After sequence-level candidate prioritization, we further construct an enzyme-kinetic model for the RPA–CRISPR/Cas12a system to explore how reaction-level parameters influence amplification dynamics and downstream signal generation. While the sequence model prioritizes suitable primer–crRNA combinations, the kinetic model focuses on the optimization of biochemical reaction conditions after target sequence selection.

The model describes the effects of adjustable reaction components and their relative ratios on target amplification, Cas12a-mediated recognition, and signal accumulation. The modeled parameters primarily include experimentally controllable factors associated with the RPA and CRISPR/Cas12a reaction system, such as primer concentrations, enzyme concentrations, and other key component ratios. By translating these parameters into a computationally searchable design space, the model identifies theoretically favorable regions for subsequent experimental evaluation.

Importantly, the kinetic model is not intended to fully reproduce all biochemical processes or directly determine the final assay conditions. The actual RPA–CRISPR/Cas12a reaction is affected by multiple interacting factors that may not be completely captured in a simplified kinetic framework, including buffer composition, ionic conditions, temperature, reaction time, effective enzyme activity, and non-ideal interactions among reaction components.

Therefore, we adopted a model-guided optimization strategy combining computational exploration with experimental refinement. The kinetic model first provides rational parameter ranges and relative component ratios as informed starting points for wet-lab experiments. Experimental optimization is then performed around these predicted regions by evaluating additional assay variables, including reaction temperature, RNP concentration, Mg²⁺ concentration, and other experimentally coupled conditions.

In this framework, dry-lab kinetic modeling does not replace experimental optimization but reduces unnecessary exploration of the parameter space. The final RPA–CRISPR/Cas12a reaction conditions are established through iterative refinement between computational modeling and experimental validation, allowing the model to provide mechanistic guidance while accounting for the complexity of the real biochemical system.

5. Portable Detection Device Design

5.1 Modular Hardware Architecture

To enable integrated on-site detection of the five respiratory pathogens, the portable detection device adopts a compact rectangular structure of approximately 15 cm × 30 cm × 20 cm, integrating multiple functional modules—including the main control and AI core, temperature control, optical acquisition, power management, and mechanical positioning—to support simultaneous reaction and signal readout across multiple samples and channels. The table below summarizes the hardware modules and their design responsibilities.

5.2 Optical Design

Portable fluorescence detection should not rely solely on visual inspection. We therefore design the optical path as a controlled “excitation–spectral filtering–fixed-geometry acquisition” chain. On the excitation side, a high-power LED matched to the excitation spectrum of the reporter probe (e.g., a 470 nm blue LED) illuminates the sample well, and its position relative to the sample is fixed mechanically to ensure reproducibility; on the emission side, an emission filter (e.g., an orange filter with a center wavelength of 600 nm and a bandwidth of ±20 nm) selectively transmits the target fluorescence band while rejecting the excitation light and stray light, after which a high-resolution macro camera (e.g., 48 megapixels, supporting 1080P) records only the target fluorescence signal. In this way, each of the five pathogen channels corresponds to an independent excitation–acquisition path and is mapped one-to-one to the fluorescence color of its reporter probe.

The entire optical assembly is enclosed in a closed dark chamber to reduce ambient-light interference from indoor lighting, sunlight, and phone screens, and to fix the geometric relationships among the sample, the excitation source, and the detector. This architecture standardizes signal acquisition across pathogen channels and allows the fluorescence images to serve directly as inputs to a downstream AI classification model, providing a clear interface for algorithm-based automatic interpretation—the built-in convolutional neural network (CNN) model classifies the acquired multicolor fluorescence images, thereby replacing manual interpretation and reducing subjective error.

5.3 Temperature Control Design

Because the RPA–CRISPR system does not require rapid thermal cycling, the thermal design can focus on three goals: reaching the target temperature quickly, maintaining temperature stability over time, and ensuring consistency across reaction sites. The heating element uses PTC/ceramic heating cores that transfer heat uniformly to the sample wells; temperature sensors (such as PT1000 or 18B20, with an accuracy of about ±0.3 ℃) provide real-time feedback, and the main controller regulates the heating power through a PID closed-loop algorithm to hold the reaction temperature within the preset isothermal range, with control accuracy of up to ±0.1–±0.5 ℃. To also support integrated sample pretreatment, the device reserves a separate high-temperature lysis zone (about 90 ℃) that is thermally isolated from the low-temperature reaction zone, with rapid cooling by a DC fan to avoid thermal-inertia interference, thereby accomplishing the full “high-temperature lysis–low-temperature reaction” temperature control within a single device.

5.4 Mobile Interface and User Workflow

The mobile interface is organized around the needs of on-site operators rather than around the underlying electronics. The mobile app (developed with Android Studio) connects to the device's main controller over Wi-Fi via the TCP protocol and transmits data in JSON, integrating functions such as temperature control, signal detection, protocol configuration, wireless communication, image recognition, model inference, data visualization, and result determination. The core pages include three major modules: device connection (Home), reaction-protocol selection (Protocol), and wireless connection (Wi-Fi). During detection, the user only needs to load the sample and reagents, select an appropriate detection protocol, and start the run; temperature control, fluorescence-image acquisition, CNN model inference, and data organization are all coordinated by the device.

For the five-pathogen detection panel, the result interface should clearly distinguish the fluorescence channels of RSV, hMPV-B, SPN, BP, and HI, while also including negative or invalid signal states, and should present the interpretation in the form of confidence scores or fluorescence levels to avoid ambiguity and to ensure that all displayed labels correspond directly to the targets covered by this project. In addition, the app should display real-time curves of reaction temperature and fluorescence intensity over time, support user-defined reaction parameters, and maintain consistent compatibility and usability across mainstream phone operating systems (such as HarmonyOS, HyperOS, ColorOS, and OriginOS) to meet the on-site detection needs of different end users.

6. Planned Evaluation

Before system testing, we define the types of measurements required to evaluate whether the design meets its design goals. These include discrimination between target and non-target sequences, response across target-concentration gradients, reproducibility between replicate experiments, tolerance to related components in the sample matrix, time-resolved fluorescence behavior, and compatibility with the portable readout platform.

7. Design Scope and Limitations

To maintain a clear engineering scope, we clearly define the limitations of the current workflow. First, extreme physical and chemical properties of a target sequence, such as very high or very low GC content or highly stable secondary structures, may create basic limits on assay performance. Our workflow can improve system compatibility, but it cannot remove limits caused by the underlying biology of the target itself. Second, the current iterative training and validation framework focuses on respiratory pathogens. Expansion to very different pathogen classes will require additional experimental data and further calibration of model performance.