Virtual Calibration Quantitative Mass Spectrometry Imaging for

Jan 14, 2019 - It is highly challenging to quantitatively map multiple analytes in biotissues without specific chemical labeling. Quantitative mass sp...
1 downloads 0 Views 4MB Size
Subscriber access provided by United Arab Emirates University | Libraries Deanship

Article

Virtual Calibration Quantitative Mass Spectrometry Imaging for Accurately Mapping Analytes Across Heterogenous Bio-tissue Xiaowei Song, Jiuming He, Xuechao Pang, Jin Zhang, Chenglong Sun, Luojiao Huang, Chao Li, Qingce Zang, Xin Li, Zhigang Luo, Ruiping Zhang, Ping Xie, Xiaoyu Liu, Yan Li, Xiaoguang Chen, and Zeper Abliz Anal. Chem., Just Accepted Manuscript • DOI: 10.1021/acs.analchem.8b04762 • Publication Date (Web): 14 Jan 2019 Downloaded from http://pubs.acs.org on January 14, 2019

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 17

Virtual Calibration Quantitative Mass Spectrometry Imaging for Accurately Mapping Analytes Across Heterogenous Bio-tissue Xiaowei Song,†,§ Jiuming He,†,§ Xuechao Pang,† Jin Zhang,† Chenglong Sun,† Luojiao Huang,† Chao Li,† Qingce Zang,† Xin Li,† Zhigang Luo,† Ruiping Zhang,† Ping Xie,† Xiaoyu Liu,† Yan Li,† Xiaoguang Chen,† Zeper Abliz *,†,‡ † State Key Laboratory of Bioactive Substance and Function of Natural Medicines, Institute of Materia Medica, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100050, China. ‡ Centre for Imaging and Systems Biology, School of Pharmacy, Minzu University of China, Beijing 100081, China. ABSTRACT: It is highly challenging to quantitatively map multiple analytes in bio-tissues without specific chemical labelling. Quantitative mass spectrometry imaging (QMSI) has that potential but still poses technical issues for its variant ionization efficiency across complicated, heterogenous biomatrix. Herein, a self-developed air-flow assisted desorption electrospray ionization (AFADESI) is introduced to present a proof-of-concept method, virtual calibration (VC) QMSI. This method screens and utilizes analyte response-related endogenous metabolite ions from each mass spectrum as native internal standards (IS). Through machine learning-based regression and clustering, tissue-specific ionization variation can be automatically recognized, predicted and normalized region-by-region or pixel-by-pixel. Therefore, the quantity of analytes can be accurately mapped across highly structural bio-samples like whole-body, kidney, brain or tumor, etc. VC-QMSI has the advantages in simple sample preparation without laborious isotopic IS synthesis, extrapolation for those unknown tissues or regions without previous investigation, and automatic spatial recognition without histological guidance. This strategy is suitable for mass spectrometry imaging using varieties of in situ ionization techniques. It is believed that VC-QMSI has wide applicability no matter for drug candidate’s discovery, molecular mechanism elucidation, biomarker validation or clinical diagnosis.

histological region pixel-by-pixel or for same regions came from different individuals.

Mass spectrometry imaging (MSI), as an evolving technique, has shown great power in biomedical research 1-5 and drug discovery 6-9 due to high chemical specificity without loss in spatial information.10-12 This advantage makes MSI a powerful tool to provide direct evidence of the native distribution of analytes (endogenous metabolites, drugs and their metabolic products) in vivo and better understand molecular events in disease progression 13 or the potential pharmacological and toxicological action of the drug.14-16 For new chemical entities or drug candidates, it is especially critical to quickly acquire their accurate concentration in target or non-target regions before a final decision is made for further development or suspension at the early stage of discovery.

It has been well known that the biochemical micro-environment will also cause ion suppression to the detected drug ion signals [14, 20-21, 25]. Based on this fact, we proposed the assumption that this complex relationship between the analyte’s relative response and regional biochemical components can be mathematically illustrated by regression model using machine learning technique. Herein, we proposed a proof-of-concept strategy, virtual calibration (VC), to effectively normalize the analyte response variation among different organs, tissues, sub-organs and even at each pixel. Thus, the quantity of analyte in any region could be accurately measured using QMSI.

17-19

Currently, introduction of a stable isotope labelling internal standard (SILIS) 20-25 is the most recommended strategy to correct the ion suppression variability in quantitative MSI (QMSI) because the SILIS theoretically has an identical ionization efficiency with the tested drug compound. Unfortunately, extra labor/time cost is required either to synthesize that SILIS for each new chemical entity or to coat it uniformly onto a large whole-body area with special equipment. Some reports introduced SILIS-free method using the normalization factor termed the “Tissue/Signal Extinction Coefficient (TEC/SEC)” 26, 27 or “Tissue-Specific ionization efficiency Factor (TSF)” 28 to account for the regional matrix effect. This method may be well suited for whole-body imaging experiments only if the normalization factor for each type of region has been thoroughly evaluated. However, this empirical coefficient-based calibration method mainly summarizes the drug’s matrix effect in organs which are treated as the homogenous entities and reflect the average trend of drug responses. Therefore, it is still hard to give accurate compensation for tiny, unknown

In the VC strategy, the endogenous metabolite ions acquired in mass spectrum of each organ serve as potential natural IS candidates. Among these, analyte response-related metabolite ions will be screened and utilized as the input features for the regression model of inter-region analyte response variation. This regression model will fit the response relationship between the analyte ion and those related metabolite ions. Then, the analyte ions will be corrected with virtual calibration factors predicted by this regression model under the aid of machine learning technique. In this research, we employed a self-developed air-flow assisted desorption electrospray ionization (AFADESI) method 29, 30 to undertake the imaging of target analyte as well as those related metabolites, making a case study for this IS-free QMSI strategy. The general framework of the VC-QMSI strategy is illustrated in Figure 1 using an example on drug quantification in a whole-body tissue section. Two sets of analyte-spiked reference tissues were constructed. One group contains dilution series of the analyte standards in parallel simulative sections (Figure 1A) for fitting the 1

ACS Paragon Plus Environment

Page 3 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

standard curve for quantitation. The other group containing the same analyte content in different types of organ surrogates (Figure

1B) was used to screen analyte response metabolite ions and to train the regression model for calibrating tissue-specific ionization

Figure 1. Schematic illustration of the virtual calibration quantitative mass spectrometry imaging to accurately map drug in whole-body animal tissues. (A) The reference tissue spiked with different analyte contents. (B) The different organ reference tissues spiked with the same quantity of analyte. Ht, Li, Sp, Lu, Ki and Br denote heart, liver, spleen, lung, kidney and brain, respectively. (C) Scheme of machine learning-based regression modelling. The analyte response-related metabolite ions were screened as the input features, and the analyte ion intensities were set as the training target. (D) The optical image of the whole-body section of mouse dosed with drug. (E) The drug ion variation across different organs via isotopic internal standard calibration (Iso. Calib.), virtual calibration (Virt. Calib.) or without calibration (No Calib.). (F) The image of the relative calibration factor across the whole-body section predicted by the regression model and analyte response-related metabolite ions. (G) The original ion image of drug without any calibration. (H) The standard curve constructed with the drug quantities versus calibrated drug ion intensities. (I) The quantitative visualization result of the drug in whole body. (J) The statistical result of drug distribution in each organ. (K) The clustering results of all bio-informative pixels in the whole body by t-SNE Kmeans clustering analysis. (L) The image of whole-body sample segmentation by automatic pixel labelling.

variation. A panel of suitable endogenous metabolite ions (Figure 1C) and an ideal regression model were validated to best predict the tissue-specific suppression for the drug ionization. In this strategy, the AFADESI-obtained endogenous metabolite ions, considered as “natural ISs”, were used to calibrate the analyte response for QMSI.

Pharmaceutical Factory. Chemical structures of four model compounds were shown in Figure S-1. Balb/C mouse (purchased from Beijing Vitalstar Biotechnology Co., Ltd.) were selected as the tested animals. MDA-MB-231 breast cancer cell suspension mixed with matrigel were injected into mouse to build the xenograft tumor model.



Whole-body Section Preparation. The Balb/C mice weighing 1417.0 g were randomly selected to be dosed with 10 mg/kg LXY6006, 20 mg/kg PTX, PTX-CH, or MTX, respectively. After drug administration for 30 min, the mice were euthanized with CO2 gas and fixed into a freezing block with 2.5 % (w/v, g/100 mL) aqueous carboxymethyl cellulose embedded outside. The 25 μm sagittal whole-body animal cryo-sections were made using Leica CM 3600 XP instrument (Leica Microsystem Ltd., Germany) at 20 °C and thaw mounted on the epoxy resin adhesive-covered superfrost plus slide (Thermo Technologies). All sections were stored at -80 °C until use and dried in a vacuum desiccator (6 h) before AFADESI-MSI analysis. The optical images were acquired with a Microtek Scan Maker i360 at 300 dpi resolution.

EXPERIMENTAL SECTION

Case-study reagent and materials. Acetonitrile and isopropanol were purchased from Merck (Muskegon, MI). Purified water was obtained from Wahaha (Hangzhou, China). The new drug candidates coded LXY6006 (LXY) and modified paclitaxel (PTXCH) were provided by Professor Ping Xie’s laboratory and BioDuro Shanghai Co. Ltd., respectively. The drug standard methotrexate (MTX) was purchased from Zhenzhun Biotechnologies Co. Ltd., Shanghai. The formula of LXY and PTX-CH were prepared into the form of 15 % (w/v) hydroxypropyl-β-cyclodextrin complexes. The drug formula of paclitaxel (PTX) were purchased from the Beijing Union 2

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 17

tolerance of ± 0.005 Da. Each pixel was also registered with an unique pair of coordinate indexes (row and column No.). Each row in the spectrum was presented as the relative intensity normalized against the base peak m/z 149.0233 to reduce the influence of system fluctuation.

Reference Tissue Preparation. Considering the homogenized tissue surrogate and its spiked analyte inside were in a relatively uniform distribution, this model section may be more suitable as the reference tissue to provide the external quantitation curve and the average level of the relative matrix effect for different organs. The drug-spiked tissue homogenates were prepared as reported in the literature.31,32,33 To make the reference tissue histologically similar to the native one, the cell suspension was also mixed with these tissue homogenates before molding and cryo-sectioning. Among these mimetic tissues, a dilution series of drug standards was spiked into the homogenized liver tissue to construct the external standard curves. The contents were 0.36, 0.71, 1.78, 3.56, and 7.12 pmol/mm2 for LXY; 0.25, 0.51, 1.27, 2.54, and 5.09 pmol/mm2 for PTXCH; 1.47, 2.93, 7.33, 14.6, and 29.3 pmol/mm2 for PTX; and 7.27, 18.16, 36.33, 72.66, 145.3, and 290.6 pmol/mm2 for MTX. In addition, different types of homogenized tissue surrogates (heart, liver, spleen, lung, kidney and brain) spiked with the same content of drug standards (1.78 pmol/mm2 for LXY, 1.27 pmol/mm2 for PTXCH, 7.33 pmol/mm2 for PTX, and 18.16 pmol/mm2 for MTX) were also employed to build the regression model for predicting the relative response factor. More details about the refence tissues preparation protocol was specified in supporting information.

For spatial segmentation, the whole data matrix was logarithmically transformed to improve the power of these features, which had low abundance but high contribution to clustering. For regression modelling, each column of peak intensity “Ii” was normalized with the formula “Ii’=Ii - min (I) / range (I)” to eradicate the influence of intensity level and make the intensity value range within 0 - 100 %. All of the above steps were schematically summarized (Figure S-3) and performed using self-written script in MATLAB (Supporting Information Zip file). Metabolites Screening. The MS features including region-specific metabolites and analyte response-related metabolites are of great importance for spatial segmentation and virtual calibration. They can directly determine the performance of a model, no matter for clustering or regression. To achieve the best region recognition result, we first screened 110 endogenous metabolite ions (Table S1) from the representative mass spectrum from each type of region as the input variables to implement Kmeans clustering for the pixels in the whole-body section. After standardization using “zscore” function, linear and non-linear dimension reduction methods, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), were employed to extract the representative features from these 110 endogenous metabolite ions. The functions “pca” and “kmeans” in MATLAB and the “tsne” function in the open source toolbox (http://lvdmaaten.github.io/tsne/) were used to carry out the above tasks. For regression modelling, we selected the top 10 endogenous metabolite ions (Figure S-4) that had relatively high Pearson correlation with the drug ion intensity change among different types of reference tissues (heart, liver, spleen, lung, kidney, brain, intestine, tumor, Figure S-5).

AFADESI-MSI Acquisition. The MSI analyses were performed on an AFADESI as previously reported. 29, 30, 34 The AFADESI interface was built in house with a stainless steel transport tube (I.D. 3 mm, O.D 4 mm, Length 500 mm) and vacuum pump (MZ2C NT, Vacuubrand, Germany). The extracting gas flow was 45 L/min under control by a gas flow meter (0-65 L/min, Tianjin Flow Meter Co., China). The sprayer and transport tube voltages were set at 7.5 kV and 3.5 kV, respectively. Acetonitrile-isopropanol-water (4:4:2, v/v/v) was used as the spraying solvent with the aid of 0.7 MPa nitrogen as the spraying gas. The solvent flow was set at 0.010 mL/min. A Q-Orbitrap mass spectrometer (Q Exactive, Thermo Scientific, San Jose, CA) was coupled with AFADESI in positive ion mode. The capillary temperature was set at 350 °C and S lens voltage was 55 V. The drug and endogenous metabolites were monitored using an alternative scan mode combining Full MS with t-SIM (Figure S-2). The drug ions [LXY+Na]+ (m/z 725.3296), [PTXCH+H]+ (m/z 983.4172), [PTX+Na]+ (m/z 876.3202) and [MTX+H]+ (m/z 455.1786) were monitored in target-SIM mode. The maximum injection time was set at 200 ms and the AGC target value was set at 3E6 for the t-SIM scan. Because the AGC target value was large enough than the sampled ion amounts that the tissues could actually yield within 200 ms, the scan time could keep consistent. The MSI data was acquired in a continuous line scan mode. The whole-body, organs, tumor tissue and reference tissue sections mounted on glass slides were scanned by a sprayer driven by a 3D electrical moving stage (MTS225, Beijing Optical Instrument Factory, Beijing, China). The raster speed of 3D electric moving stage was 0.3 mm/sec and the interval between two lines was 0.3 mm.

Regression Modelling. After the features for regression were screened as the “Natural Internal Standards”, the next key step was to build a suitable fitting model. The MSI data matrix generated from the different types of drug-spiked reference tissues was taken as the training dataset, including all of the pixels scanned mass spectra features as the input variable and the corresponding drug ion relative intensity within the same scanned pixel as the output target. Four regression methods, including principle component regression (PCR), lasso regression (LR), support vector regression (SVR) and back propagation neural network (BPNN), were chosen as candidate regression models to interpret the complex relation between the features and drug ions. The functions “pca”, “fitlm”, “lasso”, and “fitrsvm” and the Neural Network Toolbox were used to develop the above-mentioned regression models. Analyte Identification. The endogenous metabolic ions for regression modelling were identified mainly through literature confirmation and our previous experiment summary. 34, 36, 37 The delta m/z value between different adducted ions ([M+H]+, [M+Na]+ or [M+K]+), the isotope abundance from HR-MS in combination with the data searching from HMDB (http://hmdb.ca/) or LIPID MAPS (http://www.lipidmaps.org/) help to give the elemental composition and possible list of endogenous metabolites. The analyte identification results were listed in the Table S-2.

MSI Data Processing. The raw data were first converted into cdf format by Xcalibur (Thermo Fisher) for following image construction in MassImager 35 (Chemmind Technologies) and then imported into MATLAB 2017a (MathWorks) for further spatial segmentation and relative response factor prediction. A 2D data matrix was constructed as the modelling subject. In the constructed data matrix, each row represents one pixel acquired from a single AFADESI scan and each column represents one ion peak. These acquired mass spectra were aligned along the m/z axis with a mass 3

ACS Paragon Plus Environment

Page 5 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Figure 2. The comparison of different regressive modeling for calibration factor prediction. (A) Schematic illustration about the inner relationship between the input features (analyte response-related metabolite ions, Xi, i=1,2,3…) and analyte’s relative response factor (fr). (B) The predictive results of the relative calibration factor in comparison with the analyte isotopic ion as internal standard in different types of reference tissues. (C) The relative response factor image of the whole-body section using the predictive models.



RESULTS AND DISCUSSION We successfully profiled the relative calibration factor at each region by inputting the relative intensities (Xi, i=1, 2, 3…) of the metabolite ions into the developed regression model “fr = F(Xi)”. Thus, the matrix effect of the analyte ions across the whole-body tissue section can be virtually profiled. Then, the analyte ion signal “Ipix” at each pixel was corrected with the predicted relative calibration factor “fr” using the formula “ I’pix=Ipix / fr”. Therefore, the calibrated signal “I’pix” of the analyte at any point had a linear relationship only related with its absolute content. The analyte content can be finally calculated by putting the calibrated ion intensity into the standard curve, which was constructed from the spiked analyte contents (Cj, pmol/mm2, j=1, 2, 3, 4, 5) versus corresponding mean calibrated intensities (Iarea).

Calibration factor modelling. We investigated four types of regression models, including principle component regression (PCR), lasso regression (LR) 38, support vector machine (SVM) 39 and artificial neural network (ANN) 40, to establish a predictable relation between analyte ion intensity and the screened metabolite ions intensities. We compared the fitting results of the relative response of the drug in the main model organs. It was shown that LR and ANN provided the best fitting results (Figure 2). The predicted values of the average response in each type of model tissue were in close agreement with the actual intensities of the drug. The mean RE% was below approximately 10 %, and the correlation coefficient was over 0.99 (Table S-3). After a pixel-bypixel correction of the signal of the drug with the predicted calibration factor, the inter-region RSD% of the mean intensity of the drug ion was found to decrease to within 15% (Table S-4). Furthermore, a clear regional difference was observed in the possible response level of the drug from the LR and ANN predictive maps of whole-body sections. The predicted relative calibration factor “fr” can be considered as the intensity of the “virtual internal standard”. In this case study, the ANN model was employed because it achieved a better training accuracy (Figure S6 and Table S-4).

The greatest strength of this VC-QMSI method was its ability to calibrate the differences in MS response induced by tissue regionspecific matrix effects pixel by pixel. The accurate region of interested organ can also be automatically assigned instead of manual selection under the guidance of histological image. In the context of VC strategy, we neither employ equipment for uniformly coating the IS onto the large-sized whole-body sample 4

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 17

nor synthesize the isotopic IS, both of which are time-consuming and cost-prohibitive, there is also no need to construct sets of virtually calibrated ion images of serial LXY standard solution spotted on renal cortex and medulla. (I) The quantitative standard curves before and after calibration with the predicted relative response factor. GM: gray matter; WM: white matter.

curves for each type of tissue due to the computational normalization of the relative ion response among multi-regions. Virtual calibration at sub-organ level. Moreover, this VC-QMSI strategy is further demonstrated to be applicable to analyte measurement in highly heterogenous tissues, such as kidney, brain and tumor tissues. The predicted calibration factor could effectively map the sub-regions, indicating that the region-specific composition would truly cause a non-negligible influence on suppression of the drug ion. We initially designed an experiment to confirm if the predictive relative calibration factor could compensate the sub-organ-specific suppression variability (Figure 3A-3C). Duplicates of the analyte dilution series were micropipetted onto the renal cortex and medulla regions, respectively (Figure 3D-3F). The slope of the curve constructed in the renal medulla was clearly higher than that in the renal cortex region. However, the slopes of the two standard curves become more consistent with each other after the drug ion signal was corrected with the relative calibration factors of the sub-structure region. The standard curves also become much more linear, further proving the feasibility of this VC strategy to correct the viability in suppression in inter-compartments in QMSI without SILIS (Figure 3F).

We have demonstrated that VC-QMSI is feasible to predict its relative matrix effect for analyte response using the metabolomic features as natural IS. The predicted calibration factors based on this strategy can correct the analyte ion intensities in each organ or sub-organ region as well as isotopic calibration strategy. Due to benefits from the computational modelling approach, the calibration factor at any pixel could be calculated even for those unidentified regions or the micro-compartment regions in highly structured organs (kidney and brain) or heterogenous tissues (tumor). It made VC-QMSI superior to bulk measurement or isotope-labelled calibration. Unlike traditional MSI which characterize the analyte’s distribution with ion intensity, with the VC instead of externally incorporated isotopic IS, all types of target analytes’ in situ quantities could be conveniently normalized and compared within same quantity scale (pmol/mm2), eliminating the influence of tissue-specific signal variation, chemical structuredependent analyte ion efficiency or imaging technique-dependent spatial resolution. Validation of VC-QMSI. In the imaging experiment taking methotrexate (MTX) as the study case, we compared the developed VC method with the current isotopic calibration (IC) method. In one group, we spiked the same amount of MTX and its stable isotope labelled internal standard (MTX-D3) into different simulative organ samples. In another group, varied contents of MTX and the same amount of MTX-D3 were spiked into different model samples. As results, the linearity of the standard curve calibrated by VC is as good as that by isotopic IS (Figure 4). The drug ion signal intensities were only dependent on its native content without interference by matrix changes (Figure S-7 and Table S-5). The inter-organ and inter-pixel variations of the MTX ion intensities were decreased from 21.09 % and 51.60 % to 5.52 % and 28.97%, respectively (Table S-6). Finally, the images of quantitative visualization results of MTX and its main metabolite, 7-hydroxyl MTX, in real whole-body sections could be calculated and estimated, respectively (Figure S-8 and Figure S-9).These demonstrated that VC performed quite ideal as IC in compensating for the ion intensity variation caused by organ-specific ion suppression. In another result of regression modeling for PTXCH as study case, the feasibility of VC-QMSI was also demonstrated. Dilution series of PTXCH standards-spiked tissues (heart, liver, spleen, lung, kidney, brain) performed good linearity in ionization response after virtually calibration with predicted relative calibration factors, even the drug with varied concentrations were spiked into different types of reconstituted organs (Figure S10A and S10B). Furthermore, we tested two other drug-spiked tissues (muscle and tumor) which were not included in the regression modeling training sets. As our expected, the distorted PTXCH ion intensities in muscle and tumor were successfully calibrated by VC (Figure S10C and S10D). It means that the established regression model not only has a good fitting ability, but also performs good in extrapolation for these unknown tissue regions.

Figure 3. The prediction of relative calibration factor at substructure level in those highly heterogenous organs and tissue. The hyperspectral images of relative calibration factors of (A) kidney, (B) brain, (C) xenograft breast tumor predicted by ANN regression modelling. The relative calibration factors of their correspondent sub-organ regions were displayed in (D), (E) and (F). (G) The overlay of optical images with ion images obtained with serial of LXY standard solution spotted onto renal cortex and medulla. (H) The overlay of optical images with

Since the principle of VC was built on general endogenous features within biological tissue for regression modelling and 5

ACS Paragon Plus Environment

Page 7 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

clustering, it was worth noted that the developed VC-QMSI is not only applicable for those sprayed-based MSI (such as DESI 41 or nanoDESI 42) but also for those laser or plasma-based MSI (like MALDI, 43 SIMS, 44 LAESI, 45 LA-ICP, 46 etc.), no matter for exogenous drugs or endogenous biological molecules. Therefore, this VC-QMSI method is quite important for both drug’s pharmacokinetic and the quantitative metabolomics study. It is especially suitable for research into new drug candidates at the preclinical stage of drug discovery, when the accurate quantity of the drug candidate in the target or non-target region should be quickly determined to indicate the potential pharmacological or toxic effects. Moreover, if a dilution series of the isotopic analyte is used to fit the quantitation standard curve, VC-QMSI can also be applied to quantitatively image endogenous molecules in bio-tissue for biomarker discovery and validation or clinical diagnosis.

histological image followed by manual selection of certain organ or sub-organ region. But this would inevitably influence the quantitative accuracy and result’s repeatability. Therefore, an unsupervised, metabolomic feature-based region clustering model was built instead of manual selection to localize the drug’s physiological position. Results of Kmeans clustering were compared using different sets of features including 110 metabolite ions, three PCA components and three t-SNE components. It can be seen that t-SNE features-based clustering performed best. The whole-body section pixels in this group were well resolved into 14 clusters with the least amount of overlap (Figure 5A) in comparison with PCA feature-based clustering (Figure 5B) and original metabolite ions based Kmeans clustering (Figure 5C). We assigned these pixels based on their designated clusters, which were coded as No. 1-14 (representing the heart, liver, spleen, lung, brain, kidney, large and small intestine, tumor, oral cavity, gastric contents, intestinal excretion, skin, muscle and testis). When the label numbers of the pixels were registered back and visualized according their position index, most of the pixels within same histological region gave the labelling result. It was furtherly proved that t-SNE combined with Kmeans clustering gave the most accurate segmentation result over other two methods.

Figure 5. Comparison in pixel clustering and spatial segmentation using different metabolomic features. Scatter diagrams of whole-body section pixels were plotted and clustered in the feature space constructed with (A) t-SNE features, (B) PCA features or (C) 110 metabolite ions. The segmentation maps were reconstructed based on the pixel index and assigned cluster labelled with different colors according to Kmeans clustering using (D) t-SNE, (E) PCA, or (F) original metabolite ions. Br: brain; M: muscle; Lu: lung; G.C: gastric contents; Sp: spleen; K: kidney; Tu: tumor; O.C: oral cavity; Ht: heart; Sk: skin; Li: liver; In: intestine; I.E: intestinal excretion; Ts: testis.

Figure 4. The standard curves constructed with different model organs spiked with varied quantity of methotrexate standards. (A) The standard curve without calibration. (B) The standard curve with the drug ion calibrated via isotopic calibration by deuterated methotrexate. (C) The standard curve via virtual calibration by endogenous metabolite ions as the natural internal standard. The drug concentration in each tissue were as follows: S1=7.3; S2=18.0; S3=36.0; S4=73.0; S5=145.0; S6=291.0 pmol/mm2.

These unsupervised clustering results proved the feasibility of automatic assignment of the drug ion signal’s position. This is especially useful for improving the accuracy when quantifying the content of the drug in certain organ or sub-organ compartment and avoiding manual selection under the guidance of optical or histological images.

Automatic spatial recognition. In this study, VC-QMSI was used to quantify the drug at whole-body level to show its accurate distribution in organs or sub-structures. Traditionally, the drug signals are discerned and quantified under aid of optical or 6

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Pharmacokinetic properties evaluation for drugs by VCQMSI. Based on the developed VC-QMSI method, the general quantitation results of LXY distribution in whole-body section were displayed in Figure 6. Since LXY performed less response in the model heart and lung region (Figure 6B), it indicated that the matrix effect of heart and lung were higher than other organ regions (Figure 6D). Therefore, more compensation in ion intensity should be given to those two regions in order to normalize the ionization efficiency into the same level. As it can be seen in Figure 6H, more accurate drug quantity could be measured from the heart and lung regions after virtual calibration in comparison with the results achieved directly by putting the raw drug ion signal into the quantitation curve. This quantitation result was furtherly crossvalidated by the reported method using “Tissue Extinction Coefficient (TEC)” as the normalization factor. [26] The relative

Page 8 of 17

To illustrate the applicability of developed VC-QMSI method in preclinical drug research, we selected two anti-tumor drug candidates, termed LXY and PTXCH as well as one clinically used anti-tumor drug, paclitaxel (PTX, positive contrast) as the study cases. As the water-insoluble drug candidate, improving the bioavailability (BA) of LXY is of first important factor that should be taken into consideration. By comparison with oral administration, we could quantitatively estimate the accurate increase in BA after intravenous administration (Figure S-12 and Table S-8) or changing the formula from suspension to HP-β-CD complex solution (Figure S-13 and Table S-9). Furtherly, the quantitative distribution maps of LXY, PTXCH and PTX in whole-body were visualized (Figure S-14) for direct comparison in the tumor-targeting efficiencies. It could be seen that the exposure level (characterized with AUC0-4h) of PTXCH and LXY in tumor were 2.25 and 11.34 folds higher than that treated with PTX, respectively (Table S-10). It was supposed that two drug candidates may perform more pharmacological actions in cancer treating. Furthermore, the tumor-targeting efficiencies (characterized by AUC0-4h ratio) were also significantly enhanced, especially for PTXCH which was structurally modified (Figure S15 and Figure S-16). Therefore, it was predicted that the PTXCH may have a more focusing therapeutic effect in target-region (tumor) with less adverse reactions than PTX in those non-target regions.

error was within an acceptable range (≤±35%). The quantitation results given by VC-QMSI was in a good correlation with that determined by TEC-QMSI (r = 0.9855) (Figure S-11). Moreover, the VC-QMSI method not only quantify the drug amounts in those known organ regions in which the relative matrix effect had been fully investigated, but also provided the accurate calibration and quantitation results in those regions which empirical TEC values were not easy to get or previously summarized (Table S-7). Compared with VC-QMSI and TEC-QMSI, there were obvious systematic errors in drug contents acquired from direct QMSI method which only employed a non-calibrated standard curve to convert the drug ion intensity into its local quantity.

These study cases fully demonstrated that VC-QMSI could give an accurate drug’s spatial content, making analytes contents more comparable among different regions or structurally different drug candidates in the same region. This VC-QMSI method greatly supports the preclinical pharmaceutical research, no matter for drug’s formulation study to improve BA, or structural design or modification to discover new drug candidates. 

CONCLUSION

In general, we demonstrated the feasibility of label-free calibration of tissue specific matrix effect as well as histology-free selection of the region in VC-QMSI. The predicted values of virtual calibration factors could simulate the drug ion’s relative intensities in each type of organ region as ideal as its isotopic ion. This computational model has the potential to give a reasonable prediction value of the relative matrix effect for the microcompartment regions in those multi-organ sections, highly structural organs or heterogenous tissues, entailing the homogenate surrogate-based QMSI strategy beyond the bulk measurement. We neither have to synthesize the SILIS which are time-consuming and cost-prohibitive nor employ the equipment for uniformly spraying the SILIS onto the large size of whole-body sample. There is also no need to construct sets of curves for each type of tissue due to the computational normalization of relative ion response among multiregions, making drug contents in two different regions or two structurally different drug candidates in the same region even more comparable. It’s especially very suitable for those new drug candidates research at the preclinical stage of drug discovery when the accurate quantity information of the drug candidate in the target or non-target region should be quickly acquired to indicate the potential pharmacological effect or toxicity.

Figure 6. The quantitation result of LXY in the whole-body tissue section. (A) The ion image of [LXY+Na]+ (m/z 725.32) in the region of mimetic liver sections spiked with the dilution series of the LXY standard. (B) The ion image of [LXY+Na]+ (m/z 725.32) in the region of the reference sections. Ht: heart; Li: liver; Sp: spleen; Lu: lung; Ki: kidney; Br: brain. (C) The ion image of LXY without calibration. (D) The hyperspectral image of the relative response factors for the drug predicted by the regression model. (E) The quantitative visualization of LXY in the whole-body section. (F) The regional segmentation in the whole-body section based on t-SNE-Kmeans clustering. (G) The external standard curve constructed from the drug content versus its intensity calibrated with the predicted relative calibration factor. (H) The final quantitative result by VC-QMSI which is cross-validated with previously reported TEC-QMSI method.



ASSOCIATED CONTENT

Supporting Information 7

ACS Paragon Plus Environment

Page 9 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry (6) Eberlin, L. S.; Mulcahy, J. V.; Tzabazis, A.; Zhang, J.; Liu, H.; Logan, M. M.; Roberts, H. J.; Lee, G. K.; Yeomans, D. C.; Du Bois, J.; Zare, R. N.; J. Am. Chem. Soc. 2014, 136, 6401-6405. (7) Margulis, K.; Neofytou, E. A.; Beygui, R. E.; Zare, R. N.; ACS Nano. 2015, 9, 9416-9426. (8) Chen, S.; Xiong, C.; Liu, H.; Wan, Q.; Hou, J.; He, Q.; BaduTawiah, A.; Nie, Z.; Nat. Nanotechnol. 2015, 10, 176-182. (9) Feist, P. E.; Sidoli, S.; Liu, X.; Schroll, M. M.; Rahmy, S.; Fujiwara, R.; Garcia, B. A.; Hummon, A. B.; Anal. Chem. 2017, 89, 2773-2781. (10) Cornett, D. S.; Reyzer, M. L.; Chaurand, P.; Caprioli, R. M.; Nat. Met. 2007, 4, 828-833. (11) Phan, N. T. N.; Li, X.; Ewing, A. G.; Nat. Rev. Chem. 2017, 1, 0048-0065. (12) Ellis, S. R.; Paine, M. R. L.; Eijkel, G. B.; Pauling, J. K.; Husen, P.; Jervelund, M. W.; Hermansson, M.; Ejsing, C. S.; Heeren, R. M. A.; Nat. Met. 2018. (13) He, J.; Luo, Z.; Huang, L.; He, J.; Chen, Y.; Rong, X.; Jia, S.; Tang, F.; Wang, X.; Zhang, R.; Zhang, J.; Shi, J.; Abliz, Z.; Anal. Chem. 2015, 87, 5372-5379. (14) Sun, N.; Walch, A.; Histochem. Cell Biol. 2013, 140, 93-104. (15) Cobice, D. F.; Goodwin, R. J.; Andren, P. E.; Nilsson, A.; Mackay, C. L.; Andrew, R.; Br. J. Pharmacol. 2015, 172, 3266-3283. (16) Prideaux, B.; Stoeckli, M.; J. Proteomics 2012, 75, 4999-5013. (17) Nilsson, A.; Goodwin, R. J.; Shariatgorji, M.; Vallianatou, T.; Webborn, P. J.; Andren, P. E.; Anal. Chem. 2015, 87, 1437-1455. (18) Lietz, C. B.; Gemperline, E.; Li, L.; Adv. Drug Deliv. Rev. 2013, 65, 1074-1085. (19) Buchberger, A. R.; DeLaney, K.; Johnson, J.; Li, L.; Anal. Chem. 2018, 90, 240-265. (20) Pirman, D. A.; Reich, R. F.; Kiss, A.; Heeren, R. M.; Yost, R. A.; Anal. Chem. 2013, 85, 1081-1089. (21) Vismeh, R.; Waldon, D. J.; Teffera, Y.; Zhao, Z.; Anal. Chem. 2012, 84, 5439-5445. (22) Bokhart, M. T.; Rosen, E.; Thompson, C.; Sykes, C.; Kashuba, A. D.; Muddiman, D. C.; Anal. Bioanal. Chem. 2015, 407, 2073-2084. (23) Chad W. Chumbley; Michelle L. Reyzer; Jamie L. Allen; Gwendolyn A. Marriner; Laura E. Via; Clifton E. Barry; Caprioli, R. M.; Anal. Chem. 2016, 88, 2386-2398. (24) Papachristou, E. K.; Kishore, K.; Holding, A. N.; Harvey, K.; Roumeliotis, T. I.; Chilamakuri, C. S. R.; Omarjee, S.; Chia, K. M.; Swarbrick, A.; Lim, E.; Markowetz, F.; Eldridge, M.; Siersbaek, R.; D’Santos, C. S.; Carroll, J. S.; Nat. Commun. 2018, 9, 2311. (25) Pirman, D. A.; Yost, R. A.; Anal. Chem. 2011, 83, 8575-8581. (26) Hamm, G.; Bonnel, D.; Legouffe, R.; Pamelard, F.; Delbos, J. M.; Bouzom, F.; Stauber, J.; J. Proteomics 2012, 75, 4952-4961. (27) Luo, Z.; He, J. J.; He, J.M.; Huang, L.; Song, X.; Li, X.; Abliz, Z.; Talanta. 2018, 179, 230-237. (28) Stoeckli, M.; Schweitzer. A.; Int. J. of Mass Spectrom. 2007, 260, 195-202. (29) He, J.; Tang, F.; Luo, Z.; Chen, Y.; Xu, J.; Zhang, R; Wang, X.; Abliz, Z.; Rap. Commun. Mass Spectrom. 2011, 25, 843-851. (30) Luo, Z.; He, J.; Chen, Y.; He, J.; Gong, T.; Tang, F.; Wang, X.; Zhang, R.; Huang, L.; Zhang, L.; Lv, H.; Ma, S.; Fu, Z.; Chen, X.; Yu, S.; Abliz, Z.; Anal. Chem. 2013, 85, 2977-2982. (31) Takai, N.; Tanaka, Y.; Saji, H.; Mass Spectrom. 2014, 3, A0025A0031. (32) Groseclose, M. R.; Castellino, S.; Anal. Chem. 2013, 85, 1009910106. (33) Song, X.; Luo, Z.; Li, X.; Li, T.; Wang, Z.; Sun, C.; Huang, L.; Xie, P.; Liu, X.; He, J.; Abliz, Z.; Anal. Chem. 2017, 89, 6318-6323. (34) He, J.; Sun, C.; Li, T.; Luo, Z.; Huang, L.; Song, X.; Li, X.; Abliz, Z.; Adv. Sci. 2018, 1800250. (35) He, J.; Huang, L.; Tian, R.; Li, T.; Sun, C.; Song, X.; Lv, Y.; Luo, Z.; Li, X.; Abliz, Z.; Anal. Chim. Acta. 2018, 1015, 50-57. (36) Li, T.; He, J.; Mao, X.; Bi, Y.; Luo, Z.; Guo, C.; Tang, F.; Xu, X.; Wang, X.; Wang, M.; Chen, J.; Abliz, Z.; Sci. Rep. 2015, 5, 1408914101.

Chemical structure of selected four model drugs. Schematic illustration of alternative AFADESI acquisition method and the general workflow of QMSI data processing. The selected top 10 metabolomic features which had relatively high agreement with the drug response. The selected metabolomic features which could be used to discern the organ-specific region. Comparison results between virtual calibration and isotopic calibration. The spatial segmentation and quantitation results of methotrexate and its metabolic product in whole-body section. The spatial-temporal change of LXY content via different drug administration. The difference in spatial content distribution of LXY via different drug formula. The spatial-temporal change of three anti-tumor drugs in whole-body animals. All the raw MS data are stored at the format of *.xls or *.mat within the zip file in the Supporting Information. The additional data that support the findings of this study are available from the corresponding author upon request.

AUTHOR INFORMATION Corresponding Author *E-mail: [email protected], [email protected] Tel: (+86)-010-63165218 Author Contributions §These authors contributed equally to this work. Z.A. guided all of the research work; Z.A., J.M.H. and X.W.S. designed the research, wrote and revised the manuscript; X.W.S. planned and carried out the experiments and performed the data analysis; J.Z., X.C.P., C.L.S., and L.J.H. helped bio-fabricate the mimetic tissues and prepare the whole-body animal samples; C.L., Q.C.Z., Y.L., and X.G.C. provided the tumor cell suspension, xenograft tumor model and dosing animal; X.L., Z.G.L., and R.P.Z. supported the AFADESI-MSI platform; P.X. and X.Y.L. provided the new drug candidate coded LXY6006.

Notes The authors declare no competing financial interests. 

ACKNOWLEDGMENT

This work was supported by the National Natural Science Foundation of China (Grant No. 21335007, 81773678), the National Instrumentation Program (Grant No. 2016YFF0100304). We would like to thank Dr. Runtao Tian from Chemmind Technologies for his helpful technical support with data processing in this study. 

REFERENCES

(1) Kuznetsov, I.; Filevich, J.; Dong, F.; Woolston, M.; Chao, W.; Anderson, E. H.; Bernstein, E. R.; Crick, D. C.; Rocca, J. J.; Menoni, C. S.; Nat. Commun. 2015, 6, 6944-6949. (2) Van de Plas, R.; Yang, J.; Spraggins, J.; Caprioli, R. M.; Nat. Methods 2015, 12, 366-372. (3) Jarmusch, A. K.; Pirro, V.; Baird, Z.; Hattab, E.M.; Cohen-Gadol, A. A.; Cooks, R. G.; PNAS. 2016, 113, 1486-1491. (4) Sans, M.; Gharpure, K.; Tibshirani, R.; Zhang, J.; Liang, L.; Liu, J.; Young, J. H.; Dood, R. L.; Sood, A. K.; Eberlin, L. S.; Cancer Res. 2017, 77, 2903-2913. (5) Eberlin, L. S.; Norton, I.; Dill, A. L.; Golby, A. J.; Ligon, K. L.; Santagata, S.; Cooks, R. G.; Agar, N. Y.; Cancer Res. 2012, 72, 645654.

8

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(37) Mao, X.; He, J.; Li, T.; Lu, Z.; Sun, J.; Meng, Y.; Abliz, Z.; Chen, J.; Sci. Rep. 2016, 6, 21043-21055. (38) Eberlin, L. S.; Tibshirani, R. J.; Zhang, J. L.; Longacred, T. A.; Berry, G. J.; Bingham D. B.; Nortone, J. A.; Zare, R. N.; Poultside, G. A.; PNAS. 2013, 111, 2436-2441. (39) Du, Y. M.; Hu, Y.; Xia, Y.; Ouyang, Z.; Anal. Chem. 2016, 88, 3156-3163. (40) Xiong, X.; Xu, W.; Eberlin, L. S.; Wiseman, J. M.; Fang, X.; Jiang, Y.; Huang, Z.; Zhang, Y.; Cooks, R. G.; Ouyang, Z.; J. Am. Soc. Mass Spectrom. 2012, 23, 1147-1156. (41) Takats, Z.; Wiseman, J. M.; Cooks, R. G.; J. Mass Spectrom. JMS 2005, 40, 1261-1275. (42) Lanekoff, I.; Thomas, M.; Laskin, J.; Anal. Chem. 2014, 86, 18721880. (43) Berry, K. A. Z.; Hankin, J. A.; Barkley, R. M.; Spraggins, J. M.; Caprioli, R. M.; Murphy, R. C.; Chem. Rev. 2011,, 111, 6491-6512. (44) Zheng, L.; McQuaw, C. M.; Ewing, A. G.; Winograd, N.; J. Am. Chem. Soc. 2007, 129, 12730-1273.

9

ACS Paragon Plus Environment

Page 10 of 17

Page 11 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Insert Table of Contents artwork here

(TOC Graphic Only)

10

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Paragon Plus Environment

Page 12 of 17

Page 13 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Figure 1. Schematic illustration of the virtual calibration quantitative mass spectrometry imaging to accurately map drug in whole-body animal tissues. (A) The reference tissue spiked with different analyte contents. (B) The different organ reference tissues spiked with the same quantity of analyte. Ht, Li, Sp, Lu, Ki and Br denote heart, liver, spleen, lung, kidney and brain, respectively. (C) Scheme of machine learningbased regression modelling. The analyte response-related metabolite ions were screened as the input features, and the analyte ion intensities were set as the training target. (D) The optical image of the wholebody section of mouse dosed with drug. (E) The drug ion variation across different organs via isotopic internal standard calibration (Iso. Calib.), virtual calibration (Virt. Calib.) or without calibration (No Calib.). (F) The image of the relative calibration factor across the whole-body section predicted by the regression model and analyte response-related metabolite ions. (G) The original ion image of drug without any calibration. (H) The standard curve constructed with the drug quantities versus calibrated drug ion intensities. (I) The quantitative visualization result of the drug in whole body. (J) The statistical result of drug distribution in each organ. (K) The clustering results of all bio-informative pixels in the whole body by t-SNE Kmeans clustering analysis. (L) The image of whole-body sample segmentation by automatic pixel labelling.

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure 2. The comparison of different regressive modeling for calibration factor prediction. (A) Schematic illustration about the inner relationship between the input features (analyte response-related metabolite ions, Xi, i=1,2,3…) and analyte’s relative response factor (fr). (B) The predictive results of the relative calibration factor in comparison with the analyte isotopic ion as internal standard in different types of reference tissues. (C) The relative response factor image of the whole-body section using the predictive models.

ACS Paragon Plus Environment

Page 14 of 17

Page 15 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Figure 3. The prediction of relative calibration factor at substructure level in those highly heterogenous organs and tissue. The hyperspectral images of relative calibration factors of (A) kidney, (B) brain, (C) xenograft breast tumor predicted by ANN regression modelling. The relative calibration factors of their correspondent sub-organ regions were displayed in (D), (E) and (F). (G) The overlay of optical images with ion images obtained with serial of LXY standard solution spotted onto renal cortex and medulla. (H) The overlay of optical images with virtually calibrated ion images of serial LXY standard solution spotted on renal cortex and medulla. (I) The quantitative standard curves before and after calibration with the predicted relative response factor. GM: gray matter; WM: white matter.

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure 4. The standard curves constructed with different model organs spiked with varied quantity of methotrexate standards. (A) The standard curve without calibration. (B) The standard curve with the drug ion calibrated via isotopic calibration by deuterated methotrexate. (C) The standard curve via virtual calibration by endogenous metabolite ions as the natural internal standard. The drug concentration in each tissue were as follows: S1=7.3; S2=18.0; S3=36.0; S4=73.0; S5=145.0; S6=291.0 pmol/mm2.

ACS Paragon Plus Environment

Page 16 of 17

Page 17 of 17 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Figure 5. Comparison in pixel clustering and spatial segmentation using different metabolomic features. Scatter diagrams of whole-body section pixels were plotted and clustered in the feature space constructed with (A) t-SNE features, (B) PCA features or (C) 110 metabolite ions. The segmentation maps were reconstructed based on the pixel index and assigned cluster labelled with different colors according to Kmeans clustering using (D) t-SNE, (E) PCA, or (F) original metabolite ions. Br: brain; M: muscle; Lu: lung; G.C: gastric contents; Sp: spleen; K: kidney; Tu: tumor; O.C: oral cavity; Ht: heart; Sk: skin; Li: liver; In: intestine; I.E: intestinal excretion; Ts: testis.

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure 6. The quantitation result of LXY in the whole-body tissue section. (A) The ion image of [LXY+Na]+ (m/z 725.32) in the region of mimetic liver sections spiked with the dilution series of the LXY standard. (B) The ion image of [LXY+Na]+ (m/z 725.32) in the region of the reference sections. Ht: heart; Li: liver; Sp: spleen; Lu: lung; Ki: kidney; Br: brain. (C) The ion image of LXY without calibration. (D) The hyperspectral image of the relative response factors for the drug predicted by the regression model. (E) The quantitative visualization of LXY in the whole-body section. (F) The regional segmentation in the whole-body section based on t-SNE-Kmeans clustering. (G) The external standard curve constructed from the drug content versus its intensity calibrated with the predicted relative calibration factor. (H) The final quantitative result by VC-QMSI which is cross-validated with previously reported TEC-QMSI method.

ACS Paragon Plus Environment

Page 18 of 17