A knowledge-guided deep learning framework for quantitative nucleic acid testing Authors: Jiayu Yang, Yulin Huang, Zhuolun Li, Fa Zhang, Dongxu Zhang Published: 2026 ABSTRACT AbstractNucleic acid testing is widely used in tumor screening, genetic disease detection as well as prenatal diagnosis, and it especially plays a significant and irreplaceable role in detecting and controlling outbreaks or new infectious diseases. Currently, nucleic acid detection instruments measure target concentrations primarily through optical detection techniques. The characterization of the relationship between optical signals and biochemical reactions is a key technology of current nucleic acid detection instruments. Usually, the number of target gene fragments in a sample is very small, and gene amplification by fluorescence quantitative polymerase chain reaction is currently the most widely used detection method for clinical nucleic acid detection instruments. As the sequence and length of target gene fragments, amplification bio-reagent systems, sample types and amplified light detection systems vary, fluorescence value assessment needs to meet individual testing requirements. In addition, large-scale universal screening efforts pose an even greater challenge to the generality of assessment methods when targeting different detection scenarios for the same virus. The fluorescence value-target sequence replication rate relationship model proposed in this paper is both personalized and universal, and can accurately assess the results of nucleic acid testing based on fluorescence values.Keywords: nucleic acid testing, deep learning, BiGRU-BWI, quantitative framework, energy migration Introduction Nucleic acid testing (NAT) is a vital molecular diagnostic technique for multiplex detection and absolute quantification [1]. In clinical settings, target sequences often exist below the detection limit, making quantitative PCR (qPCR) a preferred method due to its ability to exponentially amplify minimal targets using specific primers [2–5]. Fluorescent probes bind to target sequences, generating signals proportional to amplified products. Fluorescence values collected at each cycle form amplification curves, reflecting amplification dynamics and efficiency [6–8]. However, establishing a robust correlation between fluorescence values, target quantities, and biochemical kinetics remains challenging. Current methods, such as double sigmoidal curve fitting with fluorescence thresholds or derivative methods [9, 10], face issues with data drift from diverse pathogens, varying concentrations, and experimental variability. Additionally, maintaining sensitivity and robustness in large-scale testing is complicated by rapid viral mutations and regional variations, impacting assay consistency [11–13]. Current research on quantifying the initial copy number of target genes in qPCR assays primarily focuses on amplification curve analysis, with the threshold method and derivative method being the most widely employed. Both approaches rely on curve fitting techniques. Roche developed a double sigmoidal curve model combined with the threshold method to determine the fluorescence threshold, which is then used to calculate relative relationships through root analysis [10]. The derivative method, which is less sensitive to minor data fluctuations and local errors, offers distinct advantages. However, Rebrikov et al. [9] noted that relying on the maximum first-order derivative can distort amplification efficiency estimation, complicating the identification of key points on the amplification curve. In contrast, the second-order derivative accurately identifies regions with the fastest fluorescence value growth rates. Additionally, numerous researchers have proposed innovative models based on the kinetic characteristics of PCR reactions and amplification curve features. For instance, Kurnik et al. [14] introduced a cluster analysis approach for partitioned datasets, enabling robust evaluation of detection results by identifying cluster endpoints without requiring parameter fitting. Shain et al. developed the maxratio method, which calculates the ratio between adjacent cycles, plots the ratio curve, and attributes key point distributions to provide effective feedback on sample amplification [15, 16]. Despite these advancements, most current methods are tailored for standardized amplification curves and impose stringent requirements on pathogen types, target sequence selection, reagent formulation, and instrument performance. These limitations hinder their adaptability to the diverse conditions encountered in clinical testing, often resulting in calculation inaccuracies and inconsistent outcomes. In summary, we developed a relationship model between fluorescence values and target sequence replication rates, where the latter specifically refers to the amplification efficiency employed for absolute quantification. The resulting model is both personalized and broad-spectrum. This work makes two key contributions: We proposed the BiGRU-BWI algorithm (Bidirectional GRU for Baseline Window Identification), significantly enhancing background value processing accuracy. We proposed the BiGRU-BWI algorithm (Bidirectional GRU for Baseline Window Identification), significantly enhancing background value processing accuracy. We established a quantitative energy migration algorithm, greatly improving the assessment accuracy of personalized amplification curves. This method addresses the universality challenge in large-scale nucleic acid testing applications. We established a quantitative energy migration algorithm, greatly improving the assessment accuracy of personalized amplification curves. This method addresses the universality challenge in large-scale nucleic acid testing applications. Methods Data preprocessing Smooth. In the actual experimental operation, the amplification curve was inevitably disturbed by noise, which usually existed in the form of sudden spikes or small fluctuations. The above noises were attenuated by median filtering and mean filtering respectively. Background value processing. In the early stage of PCR amplification, the number of target genes was small, and the collected fluorescence signals were almost submerged in the background fluorescence signals and stray noises; the background fluorescence values consisted of a constant part and a variable part, and the variable part varies similarly to a linear function [9]. Therefore, the amplification curves are processed by subtracting the fitted background fluorescence signals from the primary function. Amplitude normalization. For a more visual comparative analysis of the amplification process of the samples, the amplitudes were adjusted to the same level according to the fluorescence maxima, using equal scaling of the amplitudes. Identification of the window of fluorescence background values The accurate rejection of fluorescence background values helps to achieve effective quantification. Based on this model, our study designed a fluorescent background value window identification method, which can effectively identify and locate the start and end points of the background value domain. BiGRU is a lightweight, high-precision recurrent neural network, which is ideal for balancing accuracy and computational complexity as well as widely used in the field of time series signal correlation [17, 18]. As shown in Fig. 1, the reset gate controls the flow of hidden states from the previous time step into candidate hidden states at the current time step, thus discarding historical information that is irrelevant to the prediction, while the update gate controls how the hidden states are updated by information from the current time step [19]. These two gate units help the GRU control data flow and forgetting, reduce the number of parameters, and speed up training. For a time step t, (1) outlines how to compute the state of the GRU unit. BiGRU-BWI model structure and calculation flow. , and are weight matrices connected to the input vector , , and are the weight matrices connected to the previous unit state vectors , , and are bias vectors. represents logistic sigmoid () and tanh represents hyperbolic tangent (). The elementwise product of two vectors is denoted by . To overcome the disadvantage that GRUs can only obtain previous information, BiGRU utilizes both the front and back information of the sequence by propagating two types of GRU units, forward and backward [20]. In nucleic acid detection, this architecture enhances background value window identification by leveraging backward signal analysis to resolve terminal boundary ambiguities in fluorescence data, enabling more accurate cycle-dependent trend modeling. The computational flow is. Taking actual mandate requirements into account, we modified the fully connected layer before the final model output so that the model output was within the actual range as well as consistent with the time sequence data node. Excessive fluctuations in the fluorescence signal often indicate abnormalities in the detection system or samples, significantly impacting the interpretation of NAT results. Therefore, we additionally designed the slope threshold rule to monitor the slope fluctuations within the window of the baseline region recognized by the model evaluation metrics, like (4). Where or denotes the relative change in the slope difference between a sub-interval in the sequence from position i to end position N or from start position 0 to position j and the entire time series data sequence (from start position 0 to end position N), and denotes a threshold value. By calculating the normalized slope difference, it is possible to identify intervals in the sequence whose slopes differ significantly from the overall trend, thereby identifying anomalies in the time series data. After the fluctuation exceeds the threshold value, the system will prompt the introduction of manual processing to ensure the reliability of the nucleic acid detection results. Dynamic differential determination Nucleic acid testing results are typically presented as qualitative judgments or quantitative calculations. For instance, the threshold method directly identifies key points by defining fluorescence thresholds, enabling indirect negative and positive classifications. However, when the qPCR process is severely inhibited, the amplification curve gain becomes significantly smaller than normal, increasing the likelihood of false negatives. Simultaneously, non-specific amplification curves may also raise the probability of false positives. To address this, we integrated the dynamic features of the amplification curve to achieve highly accurate determination of the target sequence’s presence or absence in the sample. The fluorescence values of two consecutively collected cycles were calculated by differentiation, and the dynamic differential equation (D value) was Taking the negative curve in Fig. 2 as an example, the corresponding dynamic differentiation is plotted. Clearly, with a vertical coordinate value of 1 as the boundary, the D value of the negative curve exhibits an irregular distribution. In contrast, the dynamic differential value of positive data shows a peak in the region where the fluorescence value increases significantly. The transition of dynamic differential values before and after the peak is smooth, and the ratio remains near 1 during the early amplification stage. By calculating the dynamic differential values of the dataset, fluctuation trends are monitored, and anomalous negative datasets are eliminated. Different types of amplification curves with dynamic differential plots. (a) Negative amplification curve 1. (b) Negative amplification curve 2. (c) Negative amplification curve 3. (d) Negative amplification curve 4. (e) Positive amplification curve. Quantitative algorithms for energy migration During the PCR reaction, the primer binds specifically to the target sequence, and the reaction occurs in the presence of polymerase. Under the function of polymerase, the probe is degraded, as a result, the donor and acceptor fall out, the fluorescent probe releases fluorescence, and some of the chemical energy is converted into light energy in a process called fluorescence resonance energy transfer (FRET). When the donor is excited, the efficiency of FRET is measured in terms of the distance between the donor and the acceptor where represents the distance between donor and acceptor when the FRET efficiency is 50%. As the PCR amplification reaction proceeds, the number of fluorescent probes in the system decreases with the increase of target genes, thus, the FRET delivery efficiency decreases accordingly. In the PCR amplification curve, when the target gene is replicated in large quantities and the fluorescent signal surpasses background noise, the fluorescence value increases dramatically, showing a stable linear phase where chemical energy is converted into light energy. The energy transferred from the FRET donor to the acceptor is emitted as light after the donor dissociates. In the energy transfer quantification algorithm, we use the midpoint between the fluorescence background and final value as the reference point. This choice is based on two reasons: (i) both FRET and PCR curves exhibit stable energy transfer efficiency during the linear phase, making the midpoint a reliable central tendency indicator; (ii) the midpoint is less sensitive to outliers, reducing noise interference and improving algorithm robustness and accuracy. As shown in Fig. 3, the median of the final fluorescence value and the fluorescence background value is used as the cut-off point , and the stable linear fluorescence growth function near this point is used as the ideal target sequence replication rate of the target sequence. The ratio of positive energy values achieved by the pre- and post-amplification curves of is inversely proportional to the ratio of energy transfers performed by FRET from the donor to the acceptor before and after the Quantitative algorithms for energy migration. The upper horizontal axis represents the cycle number, corresponding to the fluorescence value curve; the lower horizontal axis represents the normalized distance, corresponding to the energy transfer efficiency curve. The left vertical axis indicates the energy transfer efficiency, and the right vertical axis indicates the fluorescence signal intensity. The point x_Mid is defined as the median of the final fluorescence value and the fluorescence background value, and is used to determine the fitting interval for the linear function the parameter denotes the characteristic distance at which the energy transfer efficiency reaches half of its maximum value. These features collectively enable precise localization of the key point within the exponential amplification phase, marking the onset of exponential amplification and providing the basis for back-calculating the initial absolute copy number. where is a linear function fitted with as the central value, and is the last cycle of the PCR amplification assay. By solving , the key point where the fluorescence value of the amplification curve breaks through the background noise can be obtained, and then the relationship between the fluorescence value and the copy rate of the target sequence can be obtained. Finally, the initial copy number of the target gene is calculated according to the correspondence between the key point and the initial concentration of the gradient sample. In summary, the FVTSR model establishes a quantitative relationship between fluorescence values and amplification efficiency. Amplification efficiency is not directly treated as an independent parameter. Instead, it is implicitly inferred through the energy migration algorithm, which accurately identifies the key point in the exponential amplification phase. The key point reflects the amplification efficiency . By comparing with a standard curve constructed from serially diluted standards, the initial absolute copy number of the target gene can be inversely calculated. This technical pathway begins with fluorescence signal analysis, proceeds to key point identification and amplification efficiency characterization, and ultimately enables absolute copy number quantification without requiring prior assumptions about amplification kinetics. Results Correctness verification Verification testing of BiGRU-BWI The dataset used for the experiments consisted of a 1500 cyclic-valued time series of length 45, as well as information on the start and end positions of the window corresponding to the manually labeled regions of fluorescent background values. In order to scale the data closer to a standard normal distribution, we normalized the dataset. To prevent numerical instability problems such as gradient explosion or disappearance during the training process [7]. In this paper, 1200 data are randomly selected as the training set, and the remaining 300 data are used as the test set. Python 3.11.7, torch 2.3.0, and cuda 12.4 versions are chosen as the experimental environment, and the training set is cross-validated five times and L2 regularization is introduced, in which the weight_decay to the parameter is set to 1e-4 to minimize the effect of overfitting. In addition, a pruning removal function is added. The number of hidden layer units of the BiGRU model is set to the local optimal solution obtained by training on a smaller dataset (about 500 samples), and the number of tested hidden layer units are 32, 50, 64, 128, and 256, with the learning rate set to 1e-3 by default and the batch size set to 128 by default. We conducted evaluation tests on both fluorescence background value window identification accuracy and slope fitting accuracy. By evaluating the accuracy of the method in predicting the starting and stopping positions of the window and the correspondence between the window slope and actual window slope, we can ensure the reliability and accuracy of the method in practical applications. A comparison of some of the performance indexes of BiGRU-BWI with the traditional fixed window is shown in Table 1. Table 1Model performance comparison. Model BiGRU-BWI 3–15 fixed window MSE6.08666154.41499Pearson's Correlation0.966290.59977 P-value4.4e-1771.1e-30Absolute Difference Mean0.35781.4362 Model performance comparison. The mean absolute difference between the key points computed by us using the baseline zone window identification model BiGRU-BWI and those computed by manual labeling is 0.3578, while the mean absolute difference between the key points computed by the fixed window algorithm and those computed by manual labeling is 1.4362, which proves experimentally that the introduction of the baseline zone window identification model significantly improves the accuracy of the key point computation. The distribution of the predicted and labeled values of the BiGRU-BWI model is shown in Fig. 4, with the vertical and horizontal axes representing the model predicted values and the corresponding labeled values, respectively. The distribution of BiGRU-BWI predicted data (green) and labeled data is close to a nearly 45-degree straight line. This indicates that there is an obvious positive correlation between the predicted and labeled values, there are no obvious outliers in the scatter plot, and the data are uniformly distributed, which contrasts with the data of the traditional fixed-window method (red), indicating that our prediction method has good data consistency and reliability. Scatterplot of the predictive effects of the BiGRU-BWI model and traditional methods. The scatterplot compares prediction slope values from the BiGRU-BWI model and the traditional method against labeled slope values. The 45° dotted line indicates perfect prediction, and points closer to this reference line indicate more reliable predictions, whereas points farther from the line indicate lower prediction reliability. Gradient testing The quantitative test determines the initial copy number of target samples by distinguishing negative and positive results. Gradient-diluted samples are amplified and analysed, establishing a linear relationship between critical points and the logarithm of initial copy numbers. This relationship, represented by a fitted primary function, serves as a standard curve for quantifying unknown samples, making gradient concentration detection essential for validating the fluorescence detection system's accuracy. Five gradient concentrations of CMV virus samples were prepared, each with 5 replicates. The average critical point cycle value was calculated from the replicates to ensure reliability. The FVTSR model was used to determine the amplification data's critical points, and the results were compared with those from the CFX96 Touch instrument. The linear fit goodness-of-fit (R2) was calculated for the gradient concentration samples. As shown in Fig. 5, the experimental results demonstrate the superior accuracy of the FVTSR model in determining critical points compared to commercial PCR instruments. The high R2 value of the linear fit underscores the model's robustness in establishing a reliable relationship between critical points and initial copy numbers. Comparison of CMV sample amplification gradient test. The dotted line with triangle markers represents the evaluation results of gradient dilution samples obtained using the Bio-Rad commercial PCR instrument, and the solid line with square markers represents the evaluation results obtained using the FVTSR model. The plot is based on the cycle numbers of key points corresponding to the serially diluted gradient samples. Repeatability testing Differences in sample tubes or random operational errors may affect fluorescence excitation and acquisition, posing challenges for key point quantification. Therefore, the stability of key point positioning is crucial for accurately detecting samples of the same concentration. To evaluate this, CMV samples were diluted in a 4-fold gradient into three concentrations, with each concentration tested 10 times. The mean, standard deviation (SD), and coefficient of variation (CV) of the key points were calculated for each gradient. As show in Table 2, the experimental results demonstrate that the FVTSR model exhibits stability and precision comparable to commercial PCR instruments. With consistently low CV values across multiple tests, the model proves its robustness in maintaining reliable key point positioning under varying conditions. This performance highlights the FVTSR model's ability to match the repeatability and accuracy of established commercial systems. Table 2The results of CMV sample amplification stability test. Gradient FVTSR Bio-Rad 1 3 5 1 3 5 133.5334.8637.1133.9035.5837.59233.5034.8836.9233.9435.3637.28333.3435.5936.9233.5435.7437.32433.4034.9837.1333.7535.7137.73533.5535.0337.1834.0135.6538.07633.5735.6136.9634.0235.8537.31733.3434.7937.1433.5135.4238.07833.2834.8537.0633.4435.3637.75933.5934.7337.0134.0535.0837.401033.5834.6636.8834.0335.0037.09Average33.46834.99837.03233.81935.47537.561SD0.1165050.3350890.1090160.2275280.2826170.338475CV0.3481%0.9575%0.2944%0.6728%0.7967%0.9011%Data are presented as mean ± standard deviation (SD). CV refers to the coefficient of variation, calculated as the ratio of the standard deviation to the mean and expressed as a percentage. Values in parentheses indicate the 95% confidence interval (95% CI). Difference in CV values was calculated as FVTSR minus Bio-Rad; negative values indicate lower variability and thus better amplification stability of FVTSR compared with Bio-Rad under the corresponding gradient condition. The results of CMV sample amplification stability test. Data are presented as mean ± standard deviation (SD). CV refers to the coefficient of variation, calculated as the ratio of the standard deviation to the mean and expressed as a percentage. Values in parentheses indicate the 95% confidence interval (95% CI). Difference in CV values was calculated as FVTSR minus Bio-Rad; negative values indicate lower variability and thus better amplification stability of FVTSR compared with Bio-Rad under the corresponding gradient condition. Personalized verification Robustness testing Current nucleic acid detection systems frequently encounter atypical fluorescence curves, including background-enhanced, V-shaped, U-shaped, and wavy high-concentration curves with plateau phases. Conventional key-point localization methods often fail to accurately analyse such atypical data due to limitations in curve interpretation. The FVTSR model overcomes these challenges through advanced fluorescence data preprocessing for noise reduction and an energy migration algorithm, enabling more precise and reliable key-point identification. These atypical curves typically result from non-specific amplification, hardware noise, or random errors introduced by chips and optical components, as illustrated in Fig. 6. The integration of BiGRU-BWI enhances background value delineation accuracy and enables effective data preprocessing, ensuring reliable key point positioning. Different types of amplification curves in robustness testing. (a) Background-enhanced. (b) V-shaped. (c) U-shaped. (d) Wave-shaped. (e) Higher initial concentration. Fluorescence curves were independently evaluated as benchmarks by 3 senior diagnostic experts (Supplementary file III.2 for evaluation rules). The results showed that FVTSR was comparable to the conventional algorithm for background-enhanced/high-concentration curves, but performed significantly better for complex curves such as V/U/fluctuating curves (Table 3 and accompanying Tables S1–S5), effectively avoiding the false-negative problem commonly associated with conventional methods. Table 3Comparison and statistics of computational results of five types of atypical amplification curves. Type Correctness of qualitative judgments Average of the difference between traditional and visual valuation Average of the difference between FVTSR and visual valuation Background-Enhanced100%4.87%1.22%V-Shaped80%13.05%1.29%Wave-Shaped93.33%10.33%1.36%U-Shaped66.67%22.49%1.42%Higher initial concentration100%5.10%4.26%Correctness of qualitative judgments represents the proportion of samples correctly classified by the proposed method in each type of atypical amplification curve. Lower percentage differences indicate better agreement with visual valuation. The five atypical amplification curve types include background-enhanced, V-shaped, wave-shaped, U-shaped, and higher initial concentration curves. Comparison and statistics of computational results of five types of atypical amplification curves. Correctness of qualitative judgments represents the proportion of samples correctly classified by the proposed method in each type of atypical amplification curve. Lower percentage differences indicate better agreement with visual valuation. The five atypical amplification curve types include background-enhanced, V-shaped, wave-shaped, U-shaped, and higher initial concentration curves. Table 4Critical point calculations for non-specific amplification of negative samples versus positive control. Bio-Rad FVTSR Negative sample1.319.01NCNC3.686.01NCNC4.398.99NCNC6.654.64NCNC4.8811.08NCNC6.077.9NCNC3.025.75NCNC8.0310.41NCNC7.799.38NCNC2.885.34NCNC12.0413.29NCNC2.59NCPositive control29.7632.142929.432.0349 Critical point calculations for non-specific amplification of negative samples versus positive control. Evaluation of test results for non-specific amplification data Non-specific amplification, a common issue in PCR, occurs when non-target sequences are amplified alongside the target sequence, leading to increased fluorescence signals. Effective nucleic acid testing result assessment methods must identify non-specific amplification curves to ensure data interpretation accuracy and result reliability. As illustrated in Fig. 7, non-specific amplification in negative samples (23 negatives and 2 positive controls) produced fluorescence profiles with slow, linear increases. Traditional algorithms, such as those used by the Burroughs PCR instrument, often misclassify such curves as positive due to their lower fluorescence gain. In contrast, the FVTSR model employs a dynamic differential determination method to effectively identify and exclude non-specific data, ensuring unbiased and highly reliable PCR analysis results. This approach demonstrates the model's potential to improve diagnostic accuracy and reduce the likelihood of misinterpretation in complex amplification scenarios. Amplification curves for non-specific amplification of negative samples versus positive control. (a) Presence of positive control. (b) Positive control is not present. Quantitative assessment of low-gain specificity data Low-gain amplification curves, characterized by low amplification efficiency or insignificant fluorescence signal enhancement, often result from suboptimal PCR conditions, low template DNA concentration, or reaction system inhibition. Accurately interpreting such curves is crucial for reducing false-negative results in nucleic acid testing. To assess contamination removal efficacy, a PCR laboratory decontamination test was performed. Eight samples and two positive controls were processed simultaneously using a Bio-Rad PCR instrument, and the detection results are shown in Table 5. As shown in Fig. 8(a), while no positive control amplification curve was observed (Fig. 8(b)), the test samples displayed clear specific amplification with distinct linear phases, confirming positivity. Conventional algorithms (e.g. Bio-Rad’s default method) frequently yielded false-negative or erroneous results (e.g. 44.34, 46) due to substantial fluorescence gain disparities between controls and test samples. In contrast, the FVTSR model demonstrated superior accuracy, closely matching visual assessments. These findings underscore FVTSR’s enhanced capability in accurately interpreting low-gain amplification curves, even in challenging scenarios. Table 5Key point calculation results for low gain sample specific amplification versus positive control. Bio-Rad FVTSR NC033.68358033.7879344.3433.44636037.27822037.95296038.35775037.28171033.650024634.33698PC34.3634.0795834.3633.95216 Key point calculation results for low gain sample specific amplification versus positive control. Amplification curve for low-gain sample-specific amplification versus positive control. (a) Presence of positive control. (b) Positive control not present. Universal verification Universality testing of different target sequences of the same pathogen species To evaluate the versatility of FVTSR in detecting diverse target sequences of the same pathogen, amplification fluorescence data from two distinct targets (MP1 and MP2) were analysed. Despite variations in reagent standardization across regions, FVTSR demonstrated consistent performance in 75 (MP1) and 104 (MP2) amplification assays. As shown in Fig. 9 and Fig. 10 (Table S6 and S7), the model's quantitative key-point calculations exhibited high agreement with visual estimations, achieving minimal average deviations of 0.5823 and 0.3463, respectively. These results validate FVTSR's robustness and adaptability across heterogeneous clinical testing conditions, highlighting its broad applicability in nucleic acid detection. Gradient concentration detection results for MP1 target sequence. (a) Gradient concentration amplification curve of MP1 target sequence sequences. (b) Critical point assay results for MP1 gradient concentrations. Gradient concentration detection results for MP2 target sequence. (a) Gradient concentration amplification curve of MP2 target sequence sequences. (b) Critical point assay results for MP2 gradient concentrations. Universal testing of different primers and probes for the detection of the same pathogen species Multiple primers and probes can be designed to detect the same target sequence from a single pathogen. To evaluate FVTSR's computational capability, PCR amplification of the same target sequence was performed using different primer and probe systems. Specifically, nine systems for the NS1 target sequence and three for the M1 sequence of influenza B virus were tested. The evaluation of nine NS1 and three M1 systems, using 4-fold gradient dilutions and serum controls. As shown in Fig. 11, the results demonstrate that the model consistently achieves accurate qualitative judgments and maintains high precision in quantitative key point calculations, with deviations from visual estimations averaging 0.4361 and never exceeding 1.23 (Table S8). This performance underscores the model's reliability across diverse conditions, including different target sequences of the same pathogen and varied primer-probe systems for identical targets, further validating its versatility in clinical applications. Key point detection results for gradient concentrations of M1–2(a) and NS1–6(b). Clinical data detection To validate the practicality of FVSTR, we collected test data from 52 clinical samples of gonococcus, including 29 negative and 23 positive cases. FVTSR was applied to analyse these data, which were then compared with the automatically interpreted results from clinical diagnostic instruments. As shown in Table 6 and Table S9, the key value ranges obtained by the HS instrument and FVTSR were 25.62 ~ 27.7 and 24.75 ~ 27.15, respectively. The average relative difference between the two methods was 4.07% (range: 0% ~ 6.37%). These experimental results indicate that FVTSR demonstrates good consistency with commercial instruments in clinical data analysis. Table 6Statistical results of clinical tests. Statistical indicators Result Accuracy of qualitative judgments100%Scope of key points25.62–27.7Average relative difference4.07%Maximum difference6.37% Statistical results of clinical tests. Discussion Ensuring qPCR accuracy and consistency is increasingly challenging due to global mass screening demands. Current PCR evaluation methods rely on curve fitting and threshold determination, offering stability but often overlooking data characteristics, leading to potential distortions. Existing quantitative approaches are also limited to positive curves with clear linear features. Recent studies have shown that qPCR analysis is moving beyond conventional threshold-based quantification toward data-driven interpretation of full amplification curves. Machine learning-based amplification curve analysis has demonstrated that kinetic information in real-time fluorescence trajectories can support target discrimination, multiplex identification, and assay optimization, as illustrated by the Smart-Plexer framework, Smart-Plexer 2.0, and related deep domain adaptation strategies [21–23]. These advances highlight the growing role of artificial intelligence in extracting nonlinear information from qPCR data under complex testing conditions. At the same time, accurate quantification still depends on mechanistically meaningful descriptions of amplification efficiency and reaction kinetics. Studies on single-reaction efficiency estimation and amplification efficiency-aware kinetic modeling have shown that ignoring efficiency variation during amplification may introduce bias into quantitative analysis [24, 25]. In this context, our FVTSR model provides a complementary route to current AI-driven approaches. Rather than relying solely on data-driven pattern recognition, it retains physical interpretability through an energy-conversion perspective while adapting to personalized amplification behaviors, thereby offering a balanced strategy for robust qPCR quantification. However, the FVTSR model also has several limitations under specific conditions. When the fluorescence growth efficiency of positive samples is low, the transition from the background phase to the exponential growth phase may be prolonged, making it more difficult to accurately define the baseline region and potentially causing fluctuations in the evaluation results. In such low signal-to-noise ratio scenarios, the localization of the key point may become unstable, as the model's robustness heavily depends on the accurate identification of the background region and the linear growth phase. Similarly, regular fluctuations in the fluorescence background may interfere with baseline identification. In addition, the FVTSR model is built primarily on the kinetic characteristics of the exponential amplification phase, using the linear region around as the quantitative basis; consequently, its performance may be compromised when certain samples enter the plateau phase prematurely. Future work will therefore explore a segmented modeling strategy, retaining the current linear model for the exponential phase while introducing a decay function for the plateau phase, to better characterize the full amplification dynamics and improve model applicability in such cases. Another limitation is that the current model does not include a dedicated decision rule for gray-zone samples near the boundary between positive and negative results, which may restrict its direct application in some clinical reporting scenarios. Future improvements will focus on enhancing model robustness under non-ideal amplification conditions, including the introduction of a confidence assessment module to provide uncertainty indications for low signal-to-noise ratio samples, the integration of multi-feature fusion strategies such as curve slope and peak fluorescence, and the development of an abnormal curve screening function to automatically exclude samples that do not meet the model's application criteria. Additionally, segmented modeling of different amplification phases and the adoption of retesting-oriented strategies for borderline results will be pursued. NAT is essentially a dynamic biochemical reaction. It involves a combination of enzyme activity, substrate concentration, temperature, pH, and other factors. Small changes in these conditions may affect the intensity of the fluorescence signal and its rate of change over time. While our model is based on energy conversion, the positioning of the key point is closely related to the amplification efficiency. Therefore, the deviation of the key points under some specific conditions does not represent its overall failure but reflects the reaction process of complex biological systems. With the accumulation of more experimental data and in-depth understanding of the biochemical properties of the amplification process, the model will be continuously updated and improved. Key Points Proposal to quantify optical signal-biochemical reaction associations via energy conversion. Proposal to quantify optical signal-biochemical reaction associations via energy conversion. Resolves individualization of data characteristics due to sample, reagent and instrument differences. Resolves individualization of data characteristics due to sample, reagent and instrument differences. Creates a standardized analytical framework for different reagents and instrumental conditions. Creates a standardized analytical framework for different reagents and instrumental conditions. Supplementary Material Acknowledgements The authors thank the BEI Virus Resource Bank for providing the virus strains. The authors thank Dr Wangheng Hou for his help in providing samples. The authors thank the Prof Shengxiang Ge, Dr Taichun Chen, Dr Shiyin Zhang for their help in independently reviewing and interpreting the fluorescence data. The authors would like to thank Guohua Zhong for his help in the method of analysing the experimental results. Contributor Information Jiayu Yang, School of Public Health, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; National Institute of Diagnostics and Vaccine Development in Infectious Diseases, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; Discipline of Intelligent Instrument and Equipment, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China. Yulin Huang, School of Public Health, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; National Institute of Diagnostics and Vaccine Development in Infectious Diseases, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; Discipline of Intelligent Instrument and Equipment, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China. Zhuolun Li, School of Public Health, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; National Institute of Diagnostics and Vaccine Development in Infectious Diseases, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China. Fa Zhang, School of Public Health, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; School of Medical Technology, Beijing Institute of Technology, 5 South Zhongguancun Street, Haidian District, Beijing Municipality 100081, China. Dongxu Zhang, School of Public Health, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China; National Institute of Diagnostics and Vaccine Development in Infectious Diseases, Xiamen University, No. 4221 Xiang'an South Road, Xiang'an District, Xiamen, Fujian 361102, China. Conflicts of interest None declared. Funding This work was supported in part by National Natural Science Foundation of China [62472365], in part by Natural Science Foundation of Xiamen Municipality [3502Z202473003], in part by Department of Science and Technology of Inner Mongolia Autonomous Region [2023YFDZ0060], in part by Department of Science and Technology of Xinjiang Uygur Autonomous Region [ZYYD2026QY25]. Data availability The raw data for this paper can be found in this paper and its online supplement. The code can be accessed through GitHub at the following URL: https://github.com/NIDVD-MRA/FVTSR. References Associated Data Supplementary Materials Data Availability Statement The raw data for this paper can be found in this paper and its online supplement. The code can be accessed through GitHub at the following URL: https://github.com/NIDVD-MRA/FVTSR.