Ultra‐Fast Label‐Free Serum Metabolic Diagnosis of Coronary Heart Disease via a Deep Stabilizer

Abstract Although mass spectrometry (MS) of metabolites has the potential to provide real‐time monitoring of patient status for diagnostic purposes, the diagnostic application of MS is limited due to sample treatment and data quality/reproducibility. Here, the generation of a deep stabilizer for ultra‐fast, label‐free MS detection and the application of this method for serum metabolic diagnosis of coronary heart disease (CHD) are reported. Nanoparticle‐assisted laser desorption/ionization‐MS is used to achieve direct metabolic analysis of trace unprocessed serum in seconds. Furthermore, a deep stabilizer is constructed to map native MS results to high‐quality results obtained by established methods. Finally, using the newly developed protocol and diagnosis variation characteristic surface to characterize sensitivity/specificity and variation, CHD is diagnosed with advanced accuracy in a high‐throughput/speed manner. This work advances design of metabolic analysis tools for disease detection as it provides a direct label‐free, ultra‐fast, and stabilized platform for future protocol development in clinics.


Preparation and characterization of ferrous nanoparticles
Ferrous nanoparticles were synthesized by the following modified co-precipitation method. [1] Briefly, sodium citrate, sodium acetate, and ferric chloride hexahydrate were dissolved in 100 mL of ethylene glycol under vigorous stirring for 30 min. The mixture was then poured into a Teflon-lined stainless-steel autoclave and heated at 200°C for 10 h. After the reaction, the samples were carefully removed from the autoclave and cooled to room temperature. The as-prepared products were thoroughly and alternately washed with ethanol and water, and finally dried at 60°C before use.
To characterize the ferrous nanoparticles, scanning electron microscopy (SEM) images and energy-dispersive X-ray (EDX) spectra were recorded on a Hitachi S-4800 scanning electron microscope (Hitachi, Ltd., Tokyo, Japan), with ~10 μL of water-suspended nanoparticles on aluminium foil. Transmission electron microscopy (TEM), high-resolution transmission electron microscopy (HRTEM), selected area electron diffraction (SAED), and elemental mapping images were collected using a JEOL JEM-2100F instrument (JEOL Ltd., Tokyo, Japan), with ~10 μL of water-suspended nanoparticles deposited onto a copper grid with mesh diameter of 100 μm (Beijing XXBR Technology Co.，Ltd, Beijing, China). Dynamic light scattering (DLS) and zeta potential measurements were conducted using a Nano-ZS90 instrument (Malvern, Worcestershire, UK) with nanoparticles dispersed in water at 25°C.
Ultraviolet-visible (UV-vis) absorption and Fourier transform infrared (FTIR) spectra of the materials were obtained using a UV1900 spectrophotometer (Shimadzu Corporation, Kyoto, Japan) and a Nicolet 6700 FT-IR spectrometer (Thermo Fisher Scientific Inc., Massachusetts,USA), respectively. Digital images were taken using a Huawei Wheat 5 phone. (cTnI) levels were recorded by high-sensitivity immunoassay as previously described, [2] for which magnetic particles with capture antibodies were used to form immune complexes and the fluorescence intensity was converted into the cTnI concentrations in chemiluminescent reaction. For exclusion criterion, patients were excluded from the work if they had evidence of drugs or autoimmune syndromes. The blood was drawn at initial diagnosis without anaesthesia or surgery. Serum samples were collected based on a well-established protocolfrom the 261 HCs, [3] who had no clinical evidence of cardiovascular disease or other major disease and served as controls. Briefly, ~2 mL of blood was collected by venepuncture and centrifuged at 5100xg for 10 min. Then, the serum was transferred to a microtube and stored at -80°C.

Cohort characteristics and serum collection
There was no significant difference in age or sex between HCs and CHD patients.
All the investigation protocols in this study were approved by the institutional ethical committees of the Shanghai Chest Hospital and the School of Biomedical Engineering, Shanghai Jiao Tong University (KS(P)1703 and KS1736). Written informed consent was provided from all individuals participating in the study, and the use of their biological samples for analysis was approved for analysis in accordance with the Declaration of Helsinki.

Metabolic analysis by nanoparticle-assisted LDI-MS
Metabolic analyses of serum samples, metabolite standards, and prepared mixtures were performed using LDI-MS with nanoparticles as a matrix. Typically, 0.5 μL of analyte (serum sample, metabolite standard, or a prepared mixture) was mixed with 0.5 μL of deionized water-suspended nanoparticles or DHB/CHCAin a customized microarray (initially designed by Applied Biosystems, MDS SCIEX, Foster City, CA,USA) for direct LDI-MS detection. For each sample, a given number

Machine learning for the diagnosis of CHD
The machine learning algorithms (sparse machine learning (elastic net analysis) and orthogonal projections to latent structures discriminant analysis (OPLS-DA)) were applied to the previous collected SMPs. For sparse machine learning, [5] elastic net analysis was linearly combined with the least absolute shrinkage and selection operator (LASSO) and ridge regularization. The parameters λ1 and λ2 were tuned during the training process to obtain the optimized model based on its area under the curve (AUC) performance. The sparse machine learning formula that we used is as follows: where and are parameters controlling L1 and L2 normalization, n is the sample number, X is the extracted metabolic signal, and Y is the diagnostic label ('0' for HCs and '1' for CHD/MI/non-MI patients).
The OPLS-DA algorithm was derived from partial least square discriminant analysis (PLS-DA) algorithms to reduce model complexity through the removal of non-predictive variation in X (orthogonal to Y), [6] and thereby improve model interpretation. The OPLS-DA machine learning formula that we used is as follows: where ppis the Y-predictive loading matrix for X, tpis the Y-predictive score matrix, pois the Y-orthogonal loading matrix for X, tois the Y-orthogonal score matrix, qpis the Y-predictive loading matrix for Y, and E and F are the residual matrices for X and Y, respectively.
For the cross-validation, we repeated the process for 20 times with shuffled X and Y each time for the consideration of overfitting effects and the choice of best cross-validation models, where X represents the extracted serum metabolic profiles (SMPs), and Y is the diagnostic label ('0' for HCs and '1' for CHD patients).For the permutation test, we randomly permuted diagnostic labelfor 1,000 times and calculated the distribution of AUC using the uninformative data obtained by random permutation.For valid comparison of the two algorithms, both were tested with the same methodology using Python (version 3.7).

Deep learning for enhanced MS and diagnosis
The overall architecture of the deep stabilizer, which included both a generator and discriminator, was designed based on a generative adversarial network (GAN) and trained by Wasserstein GAN (WGAN) with a gradient penalty strategy. [7] The generator was designed with convolutional neural networks and consisted of two main branches. One branch was for stabilization-oriented spectrum reconstruction, and the other was for attention-guided peak refinement. [8] Final stabilized spectra were reconstructed as the element-wise sum of coarse-grained spectra and refined peaks.
The branch used for stabilization-oriented spectra reconstruction comprised a decoder and an encoder. A skip connection was introduced to convey information from the encoder to the decoder for fast gradient update and information reuse. The encoder was designed for feature extraction, and the decoder was designed for reconstruction, as follows: where , , , , , and represented the input spectra, reconstructed coarse-grained spectra and convolution kernels in the encoder (enc) and decoder (dec), respectively, and the asterisk indicated the convolution operation. The leaky rectified linear unit (LeakyReLU) was chosen as our activation function, and was defined as follows: with =0.01 to repair the "dying ReLU" problem.
To strengthen feature propagation and alleviate the vanishing gradient problem, we chose the following dense block as the basic block: [9] where represented the concatenation of previous layers' feature maps.
The branch for attention-guided peak refinement contained only one dense block with l and k set as 4 and 32, respectively: where , , , and represented the refined peaks, input spectra, and their corresponding convolutional kernels, respectively.
For final stabilized spectrum reconstruction based on refined peaks and reconstructed coarse-grained spectra, the stabilized spectra were built by their element-wise product and element-wise sum: The discriminator contained 7 convolutional layers followed by 2 fully connected layers. Each convolutional layer was followed by a LeakyReLU activation layer with a negative slope of 0.2. Every convolutional layer had a stride size of 2. The loss function of the deep stabilizer consisted of two components: adversarial loss and reconstruction loss.
For adversarial loss, a variant of GAN, WGAN with gradient penalty, was adapted as the GAN framework, which solved the slow convergence and mode collapse problems. The MS reconstruction's objective function of WGAN was defined as follows: where , and represented the reconstructed spectra, high-quality spectra and interpolated spectra, respectively, and represented the gradient penalty.
For reconstruction loss, the network was trained with a sliding window to make the deep stabilizer more robust. To decrease sensitivity to outliers, the Huber loss was selected over the mean square error (MSE) loss as the loss function. The Huber loss was defined as follows: where was given as follows: with and representing the observed value and the predicted value, respectively. Implementation was carried out in Python 3.7 with PyTorch (version 1.3.1) and torchvision (version 0.4.2).
For optimization, deep learning performance was correlated to its architecture settings. The filter number, filter size, dense block size and number of blocks were modified to study the relationships between performance and parameters. During training, the sliding-window strategy was applied to the training dataset. The Adam optimizer was used with an initial learning rate of 0.0001, [10] of 0.9 and of 0.999.
Fora small subset of the training setas minibatch, 128 random sequences were used as the input. The training process was carried out on a Nvidia GeForce GTX 1080Ti GPU (Nvidia Corporation, California, USA) for 300 epochs. When testing, a whole spectrum obtained with few laser shots was used as the input to reconstruct a spectrum obtained with many laser shots.
The capability of stabilization-oriented spectrum reconstruction and attention-guided peak refinement was critical for reconstructing high-quality MS data.
For attention-guided peak refinement, an ablation study was adopted, which remove single part of the model to obtain the influence on performance systematically.
Specifically, the study was performed for an attention operation,while diagnostic performance was compared with and without an attention mechanism.

Statistical Analysis
Pre-processing of data Resampling, smoothing, peak extraction, and peak alignmentwere performed in the pre-processing of MS.

Sample size (n)
The minimal sample size was decided by power analysis.Power analysis was performed by uploading 16 samples as the pilot metabolicdata into MetaboAnalyst and the predicted power for estimating the effect sample size was set as 0.8.

Statistical methods
Sensitivity, specificity, and accuracy were defined as follows: The confidence interval was defined as follows: where represented the average value of the observed data, represented 1-the confidence coefficient, n represented the sample number, and S represented the standard variation in the observed data.
CV was defined as follows: where represented the standard deviation of intensity or resolution and represented the mean intensity or resolution.
The S/N was calculated as follows: where y(n) and x(n) represented the spectra before and after baseline correction, respectively, and N represented the signal length.
PSNR was defined as follows: where y represented the predicted sequence and represented the ground truth sequence.

Software used for statistical analysis
The chi-square test and t-test were implemented with scipy (version 1.3.3, www.scipy.org). [11] AUPRCs and AUCs were measured with sklearn (version 0.21.3, www.scikit-learn.org). [12]  Three independent experiments were performed to determine both size and zeta potential distribution.                         f PSNR referred to the peak signal-to-noise ratio for the stabilized spectra and reference spectra. g Diagnostic performance (deep-stabilized) referred to the AUC for the diagnosis of CHD.