VIBRANT: spectral profiling for single-cell drug responses

Abstract

The measurement of single-cell drug responses is important to drug discovery, as it facilitates understanding drug mechanism of action (MoA), evaluating drug efficacy, overcoming drug resistance and optimizing drug therapy. Despite its importance, a suitable method satisfying the requirements of high throughput, high content and low instrument cost is still lacking. Here, we present a new high-content spectral profiling method named Vibrational Painting (VIBRANT), which integrates vibrational imaging, multiplexed vibrational probes and optimized data analysis pipeline for measuring single-cell drug responses. Three infrared-active vibrational probes were designed to measure general but distinct metabolic activities in human cancer cells. More than 20,000 single-cell drug response data were collected, corresponding to 23 different drug treatments. The resulting spectral profile is highly sensitive to drug-perturbed cell phenotypes. Utilizing this property, we built a machine learning classifier to predict drug MoAs at single-cell level with high accuracy and minimal batch effects, which is difficult for other methods. We further designed a novelty detection algorithm to discover drug candidates with novel MoAs and evaluate drug combinations. Overall, VIBRANT has demonstrated great potential across multiple key areas of phenotypic screening, establishing it as a promising approach for advancing drug discovery.

1. Introduction

Cellular drug responses refer to the changes that occur in cells in response to drugs, including alterations in gene expressions, protein abundance, metabolism, etc¹. The study of cellular drug responses is vital to drug discovery, as it facilitates understanding of drug MoAs, evaluating drug efficacy and safety, overcoming drug resistance, and optimizing drug therapy^2,3. However, drug response in many diseases varies dramatically due to the complex nature of cellular phenotypes and disease context⁴. Even among the same kind of cells residing in the same human tissue, there exists significant heterogeneity in molecular phenotypes and gene expression⁵. Consequently, the averaged result from traditional bulk/ensemble measurement might deviate significantly from actual cellular drug responses and mask cell-to-cell heterogeneity. Therefore, it has been suggested that only the drug response from a single cell can accurately reflect drug efficacy⁶.

Despite the perceived importance of single-cell drug response in drug discovery, its measurements present significant challenges. An ideal technique should satisfy the following requirements⁵: (1) sufficient signal-to-noise ratio (SNR) at single-cell level; (2) high throughput to measure a large number of cells; (3) high content to detect multiple cell features to construct a profile; (4) non-invasive to measure drug responses from intact or live cells; (5) low device cost and ease-of-operation to adapt to large-scale drug research; (6) robust performance with minimal batch or plate layout effects. Current techniques for single-cell drug response can be mainly divided into three categories: optical methods, mass spectrometry methods and single-cell sequencing methods. Unfortunately, none of them can fully satisfy the above-mentioned criteria. For optical methods, label-free fluorescence metabolic imaging measures the signal from autofluorescent metabolic coenzymes (reduced NADH and FAD)⁷, but is constrained by its low content. Image-based profiling methods^8,9 such as Cell Painting^10–14 provide enriched morphological content from organelle staining. However, due to the subtle cell morphological changes upon drug response, they are known to have batch effects and plate layout effects for identifying drug MoAs^3,15. Mass spectrometry methods can offer multiplexed metabolic or antigenic cell features at single-cell level, but they are intrinsically destructive and require expensive instruments^16–20. Single-cell RNA sequencing methods provide valuable insights into the drug-induced molecular changes^21–23. Nevertheless, they still suffer from high costs and complicated operations for large-scale drug development.

Vibrational microspectroscopy techniques are promising for assessing single-cell drug responses by providing a biochemical fingerprint of the structure and function of cells in the vibrational spectrum²⁴. These techniques are in principle non-destructive and contain multiplexed cellular features regarding biochemical compostitions^25–27. In general, Raman-based methods often have lower speed (spontaneous Raman)^28–32 or limited spectral coverage and require expensive instruments (e.g. stimulated Raman)^33–36. On the other hand, infrared (IR)-based methods usually possess higher sensitivity, higher throughput and wide spectrum coverage³⁷. Although label-free IR spectroscopy has been reported in distinguishing drug mechanisms in ensemble studies^38–42, relevant research on single-cell drug responses is largely lacking^43,44. A pioneering study measured a few single-cell IR spectra of cells exposed to 5 different drugs in a label-free manner⁴⁴. However, it failed to discriminate drug-treated cell spectra from control at single-cell level, likely due to insufficient sensitivity and/or specificity of the label-free approach.

Mid-infrared (MIR) metabolic imaging was recently developed by our group⁴⁵. By coupling MIR microscopy with IR-active vibrational probes, the metabolic sensitivity and specificity of this technique have been drastically improved compared with the traditional label-free measure. It can achieve large-area metabolic imaging with cellular-level spatial resolution, rich metabolic information, and high throughput. However, our previous work mainly focused on tissue imaging towards tissue metabolic profiling⁴⁶. Thus, the utilities of MIR metabolic imaging for measuring single-cell drug responses have not been fully unleashed. It remains unknown if this technique is able to discriminate drug-perturbed cell phenotypes at single-cell level. Whether the profile is specific enough to infer MoAs of drug candidates is unexplored. In addition, the specific probe design, workflow and data science methods of this technique have not been optimized.

Here, we present a new spectral profiling method named Vibrational Painting (VIBRANT), which integrates vibrational imaging (MIR imaging demonstrated in this study while general to all vibrational spectroscopy), multiplexed vibrational probes and optimized data analysis pipeline for large-scale single-cell drug response measurement. To the best of our knowledge, this is the first work that has systematically studied the use of MIR metabolic imaging in singlecell drug responses and how it can impact drug discovery. Several developments were made to tailor for measuring single-cell drug responses. Probe-wise, we introduced a new IR-active vibrational probe, deuterated oleic acid (d₃₄-OA), to monitor unsaturated fatty acid metabolism. Combining this new probe with two other probes, three-color metabolic imaging was achieved on human cancer cells. Scale-wise, we collected more than 20,000 single-cell metabolic response data corresponding to 23 different drug treatments. Analytical-wise, we introduced machine learning and novelty detection for single-cell drug response, especially for inferring MoAs. Overall, VIBRANT has satisfied most aforementioned criteria for single-cell drug response measurements including high sensitivity, high throughput, high content, non-invasive, relatively low instrument cost and minimal batch effects. The demonstrated potential in identifying drug MoAs, discovering novel drugs and accessing combination therapy here paves the way for its translation toward phenotypic drug discovery.

2. Results

2.1. Development of VIBRANT for measuring large-scale single-cell drug responses

Several important aspects were considered for the development of VIBRANT. For cell line selection, human cancer cell lines represent the cancer of origin and are widely used for anticancer drug screening^47,48. Here, metastatic breast cancer cell line MDA-MB-231 was chosen as the model cell line to mimic the drug screening process.

Probes are crucial to improve metabolic sensitivity and specificity, considering that label-free approach failed to identify cell responses of drugs at single-cell level^42,47. Several layers of requirements are expected. These probes should reflect general but different metabolic activities and be able to work together. Moreover, their signals should be captured simultaneously and distinguishable by MIR imaging. In our previous work, we demonstrated the separate use of two IR-active vibrational probes, ¹³C amino acids (¹³C-AA) and azido-palmitic acid (azido-PA), to reflect protein synthesis and saturated fatty acid (FA) metabolism, respectively. Here we introduced a new IR-active vibrational probe, deuterated oleic acid (d₃₄-OA), capable of reporting unsaturated FA metabolism. To the best of our knowledge, this is the first time that d₃₄-OA has been utilized as a probe in MIR imaging. Combining this new probe with the reported probes, 3-color metabolic imaging can be achieved to report three distinctive metabolic activities. It is important that the captured metabolic activities, including protein synthesis, saturated and unsaturated FA metabolism, are essential to support the survival and functions of cells. These essential metabolisms are regulated by a wide range of metabolic pathways and signaling networks^49–51, thereby interrogating a large biological space. As long as the drug MoA has a projection on these metabolic and signaling pathways, the corresponding drug candidate is likely to be identified.

The average single-cell IR spectrum of cells co-cultured with these three IR-active vibrational probes for 48hrs is shown in Fig. 1a. It can be observed that the signals of these three probes are captured simultaneously and spectrally separable. The red-shifted amide I band at 1616 cm⁻¹ from the original 1650 cm⁻¹ band corresponds to the newly synthesized proteins from ¹³C-AA labeling. The 2096 cm⁻¹ peak mainly stems from the signal of azide bonds in azido-PA. The new probe, d₃₄-OA has two separate peaks at 2092 cm⁻¹ and 2196 cm⁻¹ in the cell-silent region of IR spectrum, corresponding to asymmetric and symmetric CD₂ vibrations, respectively. The average IR spectrum of cells cultured by individual vibrational probes and without any labeling is shown in Supplementary Fig. 1. Noteworthy, although the azido-PA and d₃₄-OA have overlapped absorbance at 2096 cm⁻¹, the unique peak of d₃₄-OA at 2196 cm⁻¹ can be used to reflect the concentrations of d₃₄-OA in cells. To obtain pure signals from azido-PA, linear unmixing was used to readily separate the signal of azido-PA from d₃₄-OA (Methods). The FTIR images of the three vibrational probes are shown in Supplementary Fig. 2. An advantage of FTIR spectroscopic imaging in measuring single-cell drug responses is its minimal background, as shown in Supplementary Fig. 1–2. The main reason is that FTIR imaging measures MIR absorbance of cells, without concerning the autofluorescence background. Moreover, its millimolar level sensitivity dictates that the measured MIR signal is mainly from the biochemical compositions of cells, without interference of added drug whose concentrations are at micromolar or nanomolar level.

Fig. 1. VIBRANT workflow.

(a) Average single-cell IR spectrum of MDA-MB-231 cells with the labeling of three IR-active vibrational probes in 48hrs. (b) Workflow of VIBRANT. The first step is to add drugs under its IC50 concentration together with three IR-active vibrational probes to the cell culture media for 48hrs, followed by cell fixation, air-drying and FTIR imaging. Next, single-cell segmentation was performed to extract single-cell IR spectrum from imaging data, forming the single-cell spectral profiles. Downstream analysis is added to further map cell phenotypes after drug perturbations. In (3) FTIR imaging, BS represents beam splitter, MM represents moving mirror, FM represents fixed mirror.

Another aspect to consider is how to ensure the cells are in a similar state under different categories of drug treatments. To achieve this, we measured the half maximal inhibitory concentration (IC50) of different drugs separately by cell viability assays in 48hrs (Supplementary Fig. 3, Supplementary Table 1), in agreement with the probe culturing time. This method is commonly used in pharmaceutical studies to compare the effect of various products^42,52. Besides the experimental operations, data analysis is vital to map cell phenotypes from the obtained large-scale data. We thus built a complete analytical pipeline, including data preprocessing, single-cell segmentation, single-cell spectral profile extraction and downstream analysis.

The complete workflow of VIBRANT is presented in Fig.1b. The first step is to add the three vibrational probes and the drug under study at its IC50 concentration to the cell culture media and co-cultured for 48hrs. After that, cells are fixed and air-dried. Then Fourier transformed infrared (FTIR) imaging, a popular modality of MIR imaging, was performed to collect the full spectral data of cells. The collection speed is around 2 cells/second for human cancer cells with sufficient co-scans (~64 scans) with each pixel of 3.3 μm. In the following data analysis, quality test, baseline correction and data normalization were included for data preprocessing (Methods). Next, single-cell segmentation was performed with open-source software CellProfiler⁵³ to generate masks (Supplementary Fig. 4). These masks can be further used to extract single-cell IR spectrum from the MIR imaging data, forming the single-cell spectral profile. We then introduced a series of analytical methods for the downstream analysis, including machine learning, novelty detection and statistical methods to map cell phenotypes and identify drug MoAs. Overall, the workflow of VIBRANT is straightforward in experimental operations and optimized in single-cell data analysis, which laid the foundation to reveal valuable insights toward phenotypic profiling and potential drug discovery.

2.2. Systematic sensitivity and specificity evaluation toward a broad range of drugs

It remains unknown whether our method has sufficient sensitivity to discriminate cell responses from different drugs. To evaluate this, we tested 13 drugs belonging to 9 different drug MoAs. Note that we follow the standard definition of MoA used in the literature⁵⁴, but the exact mechanism can be complicated to identify for certain cases. These 9 drugs can be separated into two main panels: one consists of drugs that inhibit general metabolism; another one contains drugs targeting specific proteins or interfering with particular metabolic pathways. The single-cell metabolic responses toward different drug treatments were visualized in the 3D scatter plots, where x-axis represents saturated FA metabolism (azido-PA/CH₂), y-axis represents unsaturated FA metabolism (d₃₄OA/CH₂), and z-axis represents protein synthesis (¹³C amide I/(¹³C amide I + ¹²C amide I) ). The use of ratios is meant for more accurate quantification of metabolic activities. For each drug treatment, around 500–1000 single-cell data were analyzed and presented in Fig. 2. Each dot represents the metabolic responses of an individual cell in the plots. The corresponding 2D plots of the paired three metabolic activities under drug treatments are also provided for better quantitative visualization (Supplementary Fig. 5).

Fig. 2. 3D scatter plots on three metabolic activities of cells treated by various drugs.

Each dot indicates the metabolic responses from an individual cell. X-axis represents saturated FA metabolism (azido-PA/CH₂), y-axis represents unsaturated FA metabolism (d₃₄OA/CH₂), and z-axis represents protein synthesis (¹³C amide I/(¹³C amide I + ¹²C amide I) ). Shaded areas are error ellipses with 70% confidence. (a) Metabolic inhibitor panel from 5 different drugs inhibiting metabolism with control. ACS represents acyl CoA synthetase. FASN represents fatty acid synthase. Topo II represents topoisomerase II. (b) Specific protein/metabolic pathway inhibitors panel includes 4 different drugs targeting specific protein/pathway with control. (c) Cell responses to drugs with the same MoAs, including protein synthesis inhibitors, DNA intercalation, protein degradation inhibitors and PI3K/mTOR inhibitors. Experiments were repeated 3 times with similar results.

In the metabolic panel (Fig. 2a), five drugs that inhibit known metabolism were selected to validate our method’s ability to report protein synthesis and lipid metabolism. They were also used to assess the sensitivity of our method toward drugs targeting general metabolism. Remarkably, cells treated under different drug MoAs form identifiable clusters, demonstrating our method's sensitivity toward diverse metabolic inhibitors. For instance, cells exposed to triacsin-C, which inhibits lipid metabolism, showed significant decreases in both saturated and unsaturated FA metabolisms, whereas their protein synthesis remained at a similar level compared with the control group. This result indicates the ability of our method to specifically probe lipid metabolism. Additionally, cells treated by cycloheximide, which inhibits protein synthesis, displayed much lower protein synthesis ratios, supporting the capability of our method to reflect protein synthesis. Interestingly, for cells treated with doxorubicin, both FA metabolisms and protein synthesis were largely inhibited. This is consistent with the mechanism of doxorubicin interfering with DNA replication and inhibiting multiple macromolecule biosynthesis^55,56. For TVB-3166 and bortezomib, all three metabolic activities are reduced to different levels. For TVB-3166 (FA synthase inhibitor), FA metabolisms were significantly reduced while protein synthesis was slightly decreased. In contrast, for bortezomib (protein degradation inhibitor), protein synthesis was decreased more evidently. Together, these varied responses under different drug treatments demonstrated the sensitivity of our method in identifying potential drug candidates inhibiting general metabolism.

We further tested drugs that target specific proteins or interfere with particular metabolic pathways, including dactolisib (dual PI3K/mTOR inhibitor), olaparib (PARP inhibitor), gefitinib (EGFR inhibitor) and lapatinib (dual EGFR/HER2 inhibitor). These drugs were selected due to their reported therapeutic effects on breast cancer cell lines⁵⁷. Remarkably, cells treated with these drugs also exhibited distinctive metabolic responses (Fig. 2b), indicating the sensitivity of our approach toward drugs targeting specific proteins or metabolic pathways. For olaparib, all three metabolic activities were significantly reduced. This is consistent with its mechanism of PARP inhibition, which impedes the repair of single-stranded DNA breaks and further affects macromolecule metabolism⁵⁸. Considering this drug has been under phase-III clinical trial and recently approved for the adjuvant treatment of breast cancer⁵⁹, the evident decrease of metabolic activities observed here echoes its promising therapeutic effect. The other three drug groups also lead to different metabolic responses. It is interesting to observe that our approach can differentiate gefitinib (EGFR inhibitor) and lapatinib (EGFR/HER2 inhibitor), which further supports the sensitivity of our method to infer refined drug MoAs.

To evaluate the specificity of our method, which is defined as the ability to recognize drugs with similar MoAs, we selected drugs belonging to the same MoA and analyzed their metabolic activities (Fig. 2c). Interestingly, drugs with the same MoA display very similar metabolic responses. This finding is valid for both drugs inhibiting general metabolism (protein synthesis inhibitors, DNA intercalations/topoisomerase II inhibition, protein degradation inhibitors) and drugs inhibiting specific proteins or metabolic pathways (PI3K/mTOR inhibitors). Moreover, we tested the influence of batch effects on the signals from metabolic vibrational probes (Supplementary Fig. 6). Cells treated by the same drug from different batches are highly overlapped in the 3D scatter plots. In addition, we evaluated the dose-response of our approach with varied drug concentrations, using protein synthesis inhibitors and DNA interaction/topo II inhibitors as demonstrations (Supplementary Fig. 7). For cycloheximide that inhibits protein synthesis, only ¹³C-AA signal showed a clear dose-response curve, which further reflects the ability of our probe to specifically target protein synthesis. For daunorubicin that interferes with DNA intercalation and affect macromolecule metabolisms, all three vibrational probes exhibit similar dose responses, which shows the sensitivity and dynamic range of our approach in evaluating drug responses. Together, these results demonstrate the sensitivity and specificity of our method in discriminating drug MoA with minimal batch effects, which lays the foundation for VIBRANT to identify drug MoAs.

2.3. Advantages of multiplexed vibrational probing approach over label-free approach

Despite the high sensitivity and specificity demonstrated by VIBRANT, it remains unknown whether the use of multiplexed vibrational probes is indeed advantageous in discriminating drug responses compared with the conventional label-free approach. We thus conducted a systematic comparison of these two approaches.

We first extracted the average single-cell spectrum of cells treated by different drugs. Drugs belonging to two different MoAs, protein synthesis inhibitors (anisomycin, cycloheximide) and DNA intercalation (doxorubicin, epirubicin), were used for demonstration (Fig. 3a, 3b). For the probing approach, it can be observed that the cellular spectra of drugs with different MoAs are obviously distinct from each other and separable from the control group. Moreover, the cellular spectra of drugs within the same MoAs (e.g. doxorubicin and epirubicin) are very similar. However, for the label-free approach, the cellular spectra under all different conditions (including the control) are highly overlapped, exhibiting much subtle deviation between each other. These results directly showed the improved sensitivity for discriminating different drug MoAs with the use of multiplexed vibrational probes. It should be noted that due to the split of the amide I peak from ¹³C-AA labeling, the other vibrational peaks became elevated after min-max normalization, which contributes to the difference in cellular spectrum in the labeling approach compared with the label-free approach.

Fig. 3. Comparison between multiplexed probing approach and label-free approach.

(a-b) Averaged singlecell spectrum of drug treatments for (a) multiplexed vibrational probing approach and (b) label-free approach. (c-d) HCA dendrogram of cellular spectrum of multiple drug treatments for (c) multiplexed vibrational probing approach and (d) label-free approach. Dox represents doxorubicin, Epi represents epirubicin, Ani represents anisomycin, Cyc represents cycloheximide, Bor represents bortezomib, MG represents MG-132, Eve represents everolimus, Dac represents dactolisib, Tri represents triacsin-C, ctrl represents control group. (e) UMAP plot of cellular spectrum of multiple drug treatments using multiplexed vibrational probing approach. (f-g) UMAP plots of cellular spectrum from different batches for (f) multiplexed vibrational probing approach and (g) label-free approach. Each dot indicates cellular spectrum from an individual cell.

Next, we performed hierarchical clustering analysis (HCA) on cellular spectra of multiple drug treatments (Fig. 3c, 3d). The ability to group drugs with the same MoAs together is an important aspect in phenotypic drug discovery, known as the guilt-by-association approach to determine drug MoAs⁵⁸. Drugs with similar MoAs shall generate similar phenotypic signatures for identification. For the probing approach (Fig. 3c), the dendrogram clearly grouped drugs within the same MoAs together, which are coded by similar colors (e.g., anisomycin and cycloheximide). This result is consistent with the previous finding that drug within the same MoAs share similar metabolic responses (Fig. 2c). Moreover, the related UMAP⁶⁰ plot also demonstrates the separation of drugs with different MoAs. For the label-free approach (Fig. 3d), instead, the dendrogram failed to group drugs with similar MoAs. For example, anisomycin and cycloheximide are located far away from each other; bortezomib and MG-132, both protein degradation inhibitors, are also not grouped together.

We further evaluated batch effects of these two approaches, which is important in measuring single-cell drug responses to provide reliable readouts. The corresponding UMAP plots of drug-perturbed cellular spectra from different batches are shown in Fig. 3f–3g, where each dot indicates the cellular spectra from an individual cell. Data from one batch includes both the control group and cells under different drug treatments. For the probing approach, data from different batches are well mixed together, indicating minimal batch effects. This result is also consistent with the overlapped metabolic responses for data from different batches (Supplementary Fig. 6). In contrast, for the label-free approach, data from two batches obviously form two separate branches in its UMAP plot, suggesting severe batch effects in this approach. Overall, our studies demonstrated key advantages of using multiplexed vibrational probes compared with the label-free approach, including improved sensitivity to discriminate different drugs, improved specificity to identify drugs with similar MoAs and reduced batch effects.

2.4. Accurate prediction of drug MoAs at single-cell level using machine learning

Phenotypic drug screening has been tremendously powerful for identifying novel small molecules and potential therapeutics⁴⁷. Compared with traditional target-based approaches, the advantage of phenotypic screening is to identify drug leads and clinical candidates that are more likely to possess therapeutically relevant drug MoAs without concerning their target-binding affinity⁶¹. Our previous results showed clear separations of drug MoAs in HCA clustering and UMAP plots (Fig. 3c, 3e), with drugs of the same MoAs grouped together using multiplexed vibrational probes. Thus, we are exploring whether our method can be further used to identify and predict drug MoAs.

To perform classification and predictions of drug MoAs, we employed the guilt-by-association approach⁶¹, based on the premise that drugs with similar MoAs should generate similar spectral profiles. We used the cellular spectra of 13 drug treatments plus control group as the dataset. This dataset contains around 6,000 single-cell data with 288 features (vibrational peaks in 1000–1800 cm⁻¹, 2000–2300 cm⁻¹). Cells exposed to drugs with the same MoA belong to the same class, which corresponds to 10 different classes (9 MoAs plus the control group). The dataset was split into 70% for training and 30% for testing, where each drug MoA was split with the same ratio in training and testing data. Feature selection is an important step in machine learning to improve performance and avoid overfitting. We tested multiple popular machine learning models and utilized their corresponding feature importance for feature selection. The best prediction performances of these models are shown in Table 1. It can be observed that many classifiers, such as linear discriminant analysis (LDA), multi-layer perceptron (MLP), random forest (RF) and XGBoost, have extremely high accuracies (over 99%) to predict drug MoAs at single-cell level. Moving forward, we further tested the influence of batch effects on these classifiers, which are important in practical drug screening. The ideal situation is that the chosen model provides stable prediction results over different batches. Our strategy was to use classifiers trained on the same batch data to directly predict data collected from different batches, which would ensure models are generalizable to batch effects. Interestingly, only LDA has high prediction accuracies on the dataset from different batches (Table 1). The decreased accuracy in other classifiers might be caused by overfitting due to model complexity⁶².

Table 1.

Average accuracy of machine learning classifiers for drug MoA prediction at single-cell level on testing data with feature selection based on feature importance.

	DT	LDA	KNN	MLP	NB	QDA	RF	SVM	XGBoost
Same Batch	94.02%	99.95%	90.93%	99.71%	86.73%	91.61%	99.08%	99.71%	99.42%
Different Batches	65.15%	98.28%	56.57%	53.60%	61.85%	62.97%	40.66%	67.79%	46.07%

Note: DT represents decision tree; LDA represents linear discriminant analysis; KNN represents k-nearest neighbors; MLP represents multi-layer perceptron; NB represents naïve bayes; QDA represents quadratic discriminant analysis; RF represents random forest; SVM represents support vector machine; XGBoost represents extreme gradient boosting.

We thus selected LDA as the proper machine learning classifier to predict drug MoAs. The 3D LDA dimension reduction plot is shown in Fig. 4a. It can be observed that the single-cell data under different drug MoAs formed distinct clusters. This result agrees well with the high prediction accuracy achieved by LDA. The feature selection process is demonstrated in Fig. 4b–c. First, a LDA model was trained for feature selection, using all 288 features as input data. The ranking of permutation feature importance in this model was plotted in the average IR spectrum (Fig. 4b). Color gradient coding was used for better visualization, where redder colors represent more important features. It can be noticed that features arising from vibrational probes (1600–1616 cm⁻¹, 2000–2300 cm⁻¹) are observed to display deep red colors. The intrinsic feature importance in RF classifier was also included in Supplementary Fig. 8, showing a similar pattern. These results strongly indicated that features arising from metabolic labeling play important roles in predicting drug MoAs. Several features in the label-free region (coded by red or orange colors), such as phosphorate stretching vibrations from nucleic acids and carbonyl group from lipids, also assist the prediction. Next, we plotted the average prediction accuracy versus the number of features in Fig. 4d, where features with higher importance were selected first. It can be observed that when the feature number is 110, it reaches the maximum accuracy in both testing data from the same batch and from different batches. We thus selected these top 110 features to construct the final LDA model. The corresponding confusion matrix is shown in Fig. 4e. It can be seen that the prediction accuracy of individual drug MoA is also very high (nearly 100%). Additionally, the ROC curve is presented both for the predictions on testing data from the same batch and different batches (Fig. 4f), to further showcase the accurate prediction results. It turns out that the dataset size plays an important role. Although high accuracy prediction can be achieved within 1000 single-cell drug responses for data from the same batch, a dataset size above 2000 is needed to achieve accuracy over 90% on data from different batches (Fig. 4d). This suggests the necessity of collecting large-scale single-cell data for accurate prediction.

Fig. 4. Accurate prediction of drug MoAs at single-cell level.

(a) 3D dimension reduction LDA plot based on LDA classifier, where PSI represents protein synthesis inhibitors, PDI represents protein degradation inhibitors, DNAI/Topo II represents DNA intercalation/topoisomerase-II inhibition, LFACS represents long-chain fatty acyl CoA synthetase inhibitors and FASN represents fatty acid synthase inhibitors. (b) Feature importance ranking projected on IR spectrum with color gradient coding, where redder color indicates more important features. (c) Plot of average prediction accuracies versus the number of features, based on training data (grey), testing data from the same batch data (blue) and testing data from different batches (red). (d) Plot of prediction accuracies versus training dataset size based on training data (grey), testing data from the same batch data (blue) and testing data from different batches (red). (e) Confusion matrix of prediction results of 110 selected features, based on testing data from the same batch. (e) ROC curves of the prediction results of data from the same batch data (blue) and different batches (red).

In control experiments, we further tested whether single-cell spectrum measured in the label-free approach is sufficient to predict drug MoAs (Supplementary Fig. 9). Although the average accuracy is high for label-free data from the same batch, it drops dramatically to only 21.12% after including data from different batches. This further confirmed the advantage of the probing approach in drug MoA prediction. Additionally, we evaluated whether simply using the three vibrational probe ratios (Supplementary Fig. 10) and features mainly from the vibrational labeling is sufficient to have accurate prediction results (Supplementary Table 4). Although features mainly from the vibrational labeling offer better results compared with the label-free approach, it is still worse compared with the existing result. The conclusion is the combination of both probed and label-free spectral features provides the best prediction performance.

Overall, we systematically studied the possibility of VIBRANT to predict drug MoAs. A reliable classifier (LDA) was selected with stable and superb prediction performances across different batches. The trained model can be used as an initial reference set for further screening applications. This study demonstrates the tremendous potential of our method in phenotypic drug screening.

2.5. Discover drugs with novel MoAs using novelty detection

The ability to identify the MoA of lead compounds as either known or novel is crucial in phenotypic drug screening, as it can help assess the potential novelty of test compounds over existing drugs and guide further decision-making in drug discovery⁶³. By evaluating the similarity of the test compounds to annotated drugs in the reference set, it can either elucidate the MoA of test compounds as a known MoA or discover compounds with completely novel MoAs. This is a strong advantage of phenotypic profiling methods in drug discovery⁶⁴. If the MoA is known, it allows for comparisons to existing drugs and may provide insights into chemical structures, potential efficacy or resistance issues. If the MoA is deemed novel, it suggests the compound may represent a previously unexplored therapeutic approach or target, opening up possibilities for innovative drug discovery.

We specifically designed an algorithm that integrates LDA and novelty detection to identify drugs with novel MoAs. In this algorithm, the first step is to calculate the Mahalanobis distance between single-cell data from the test compound and each reference class in the trained LDA model. The class with the minimum distance represents the known drug closest to the test compound. Next, isolation forest⁶⁵, a popular and robust method in novelty detection, is introduced to determine whether the test compound is novel to the selected known drug class. The prediction output in this step is 1 as inliers (i.e., known) and −1 as outliers (i.e., novel) for each single-cell data from the test compound. Integrating the prediction results from all the single-cell data, a mean prediction score can be calculated for the test compound with a dynamic range of [1, −1], where a score of 1 represents all the single-cell data of the test compound are predicted as inliers (known) and −1 represents the opposite. We thus chose a mean score of 0.1 as the threshold value, which suggests 60% percent of single-cell data are predicted as inliers (known). Test compound with a mean prediction score below the threshold value is thus detected as novel. Test compound with a mean prediction score above the threshold value is detected as known, with the same MoA assigned from the drug that has been previously identified to be closest to the test compound.

The analytical results of 6 tested compounds are shown in Table 2, using the dataset in Fig. 4 as an annotated reference set. For compounds belonging to known mechanisms in the reference, such as daunorubicin (DNA intercalation) and emetine (protein synthesis inhibitor), their mean prediction scores are close to 1, hence identified as known. Their predicted MoAs are assigned from the annotated drugs in the reference that are identified to be the closest, which are found correct. In contrast, for the other four compounds, their mean prediction scores are quite negative, suggesting the high probability of these compounds being novel. To validate the prediction results, we plotted their projections onto the LDA-reduced dimensions calculated from the reference (Fig. 5). Data from test compounds are highlighted in black. For compounds that were predicted as known, their projected data is indeed closely distributed to the corresponding drug MoAs (Fig. 5a, 5b). For compounds that were predicted as novel, they generally form a new cluster away from any annotated drugs (Fig. 5c–f). The CDF of Mahalanobis distance of each test compound to the annotated reference groups is shown in Supplementary Fig. 11 to further indicate the quantitative distance measure in LDA. Thus, these results corroborate well with the novelty prediction result from our algorithm.

Table 2.

Prediction of drug MoAs as known or novel using novelty detection

Drugs	Mean prediction score	Outlier percentage	Prediction	Mechanism of Action (ground truth)
Daunorubicin	0.8497	7.515%	DNA intercalation/Topo II inhibitor	DNA intercalation/Tope II inhibitor
Emetine	0.6208	18.96%	Protein synthesis inhibition	Protein synthesis inhibition
Iniparib	−0.9923	99.62%	Novel	Unknown
Apicidin	−0.5289	76.45%	Novel	HDAC inhibitor
Taxol	−0.5052	75.26%	Novel	Microtubule stabilization
Vincristine	−0.9466	97.33%	Novel	Microtubule stabilization

Fig. 5. Identify test compounds with novel mechanism of action.

(a-f) Projections of test compounds in the calculated LDA dimension reduction space from the reference: (a) daunorubicin; (b) emetine, single-cell data are highly overlapped with the PSI cluster so that partial single-cell data of emetine (black dots) are blocked by data from the PSI cluster; (c) iniparib; (d) apicidin; (e) taxol; (f) vincristine. Ctrl represents the control group; DNAI/Topo II represents DNA intercalation/topoisomerase-II inhibition; PSI represents protein synthesis inhibitors.

An interesting finding is about iniparib, which was initially suggested as PARP inhibitor and went-through FDA clinical trials treating breast cancer⁶⁶. This compound eventually failed at Phase III with disappointing results. Later experiments proved that iniparib is not a PARP inhibitor at all, and its MoA is still unknown today⁶⁷. We found that iniparib has a quite negative mean prediction score (i.e., novel) and indeed distributes far away from the annotated PARP inhibitor cluster in the LDA plot (Fig. 5c). Hence, our finding of iniparib not being a PARP inhibitor agrees well with the latest experiments⁶⁷, further demonstrating the ability of our method to reveal insights for real-world lead compounds. Another interesting finding is that for taxol and vincristine, which share the same MoA, their distributions are quite similar to each other in the projection plots (Fig. 5e, 5f). This further supports the ability of our method to identify drug with similar MoAs. Overall, the designed novelty prediction algorithm is found reliable in determining compounds with known MoA or novel MoA.

2.6. Discriminate drug combinations

Combination therapy, a treatment modality that combines two or more therapeutic agents, has become a cornerstone of cancer therapy⁶⁸. It provides a complementary strategy to enhance the efficacy of chemotherapeutics. It can offer synergistic or additive anti-cancer effects and potentially reduce drug resistance⁶⁸. To explore whether our method can further assess drug combinations, we used two different drug combinations under FDA clinical trials for breast cancer as technical demonstrations⁶⁹: everolimus and lapatinib (Eve-Lap) and everolimus and doxorubicin (Eve-Dox). The synergy of drug combinations was determined using cell viability assay (Supplementary Fig. 12) with an open-source package SynergyFinder⁷⁰. For each drug combination, we selected two drug concentration groups with high synergy scores, including Eve-Lap (1–1), Eve-Lap (1–2), Eve-Dox (2–1) and Eve-Dox (1–2). The groups were named based on their concentration with respect to half of IC50 values from single drug treatments.

If the drug combinations possess synergistic effects on cells, the resulting cell phenotypes should be different from single drug treatment and can be identified through phenotype profiling. To further test whether our method can discriminate drug combinations from single drug treatment, we first used UMAP to better visualize their data distributions (Fig. 6a). Remarkably, we found that the drug combination groups form separable clusters in the plot. Moreover, they are distributed exactly between single drug treatments, which agrees well with their MoAs. For instance, both Eve-Lap (1–1) and the Eve-Lap (1–2) groups lie between the everolimus and lapatinib (single drug treatment), with Eve-Lap (1–1) closer to the everolimus cluster and Eve-Lap (1–2) closer to the lapatinib cluster, consistent with their relative concentrations. The Eve-Dox groups are also distributed between single drug treatments but closer to the doxorubicin group. The UMAP plot qualitatively demonstrated the ability of our method in discriminating drug combinations.

Fig. 6. Discriminating drug combinations.

(a) UMAP plot of drug combinations. Drug combination groups are highlighted in dashed circles. For Eve-Lap (1–1), the added concentration was half of everolimus IC50 (14.5μM) and half lapatinib IC50 (1.12 μM); for Eve-Lap (1–2), the added concentration was half of everolimus IC50 (14.5μM) and the IC50 value of lapatinib (2.14 μM). The Eve-Dox groups follow the same logic. Eve represents everolimus, Lap represents lapatinib, Dox represents doxorubicin. (b) Projections of drug combination data in the calculated LDA dimension reduction space from the reference. Drug combination groups are highlighted in dashed circles. DNAI/Topo II represents DNA intercalation/topoisomerase-II inhibition. Ctrl represents control group.

To provide quantitative results, we applied the novelty detection algorithm to drug combination data (Table 3), again using the training dataset as an annotated reference set. According to the mean prediction scores, all four drug combination groups were detected as novel, consistent with expectation. These results were confirmed in the LDA projection plot (Fig. 6b), where the drug combination groups form distinct clusters and separate from the annotated drugs in the reference set. We noted an interesting dose response: when the concentration of a certain single drug treatment increases in drug combinations, its prediction score is also increased and inclined to this single drug treatment. For instance, for Eve-Lap (1–2), whose concentration of lapatinib doubled compared with Eve-Lap (1–1), its mean prediction score is improved to slightly above zero with the closest known drug as lapatinib. The LDA plot also showed its closer distribution to lapatinib (EGFR/HER2 inhibitor). Overall, our method can discriminate drug combinations in both qualitative and quantitative manners.

Table 3.

Prediction of drug combinations as known or novel using novelty detection

Drugs	Mean prediction score	Outlier percentage	Prediction
Eve-Lap (1–1)	−0.5880	79.40%	Novel
Eve-Lap (1–2)	0.05565	47.22%	Novel
Eve-Dox (2–1)	−0.8683	93.41%	Novel
Eve-Dox (1–2)	−0.1482	57.41%	Novel

3. Discussions and Conclusions

We developed VIBRANT, a new spectral profiling method integrating vibrational imaging, multiplexed vibrational probes and optimized data analysis pipeline, for measuring single-cell drug responses. Compared to label-free approach, our method showed much higher sensitivity and robustness in discriminating cell phenotypes perturbed by a broad range of drugs. This likely benefits from vibrational-probe-interrogated essential metabolisms, which are regulated by broad metabolic and signaling networks and thus sensitive to drug perturbations. Utilizing this property, we achieved predicting drug MoAs with high accuracy and minimal batch effects at single-cell level, identifying drugs with novel MoAs and evaluating drug combination therapy.

VIBRANT offers a novel and alternative way for cell phenotyping after drug perturbation. Different from image-based profiling (e.g. Cell Painting) measuring fluorescence signals from organelle staining, our method measures metabolism-driven biochemical compositions of cells. Given the consensus that metabolism provides the readout closest to cell phenotypes^71,72, the biochemical features shall be more biologically relevant and interpretable, compared to pure morphology features. In addition, VIBRANT is intrinsically high-content with hundreds of spectral features. Extra feature extraction and analysis are not necessary as in Cell Painting which typically needs thousands of features^71–74. This overcomes the burden of redundant features and greatly reduces the technical barrier for data science⁷³. Another advantage of our method is its minimal batch effects, which mitigates the essential concern of phenotypic assays in capturing technical noise over biological information¹⁵. In fact, it is non-trivial and still in the research-phase to handle batch effects in Cell Painting, requiring both experimental and algorithmic efforts⁷⁴. Besides batch effects, we further calculated the Z-factor of the VIBRANT features to evaluate batch-to-batch variability (Supplementary Table 5), which is an important metric in high throughput screening (HTS). The existence of features with Z-factor greater than 0.5 indicates a strong separation between the control group and the drug treatment group in our assay, showing great promise for HTS. To better compare VIBRANT with other methods, we concluded the pros and cons of different methods in Supplementary Table 2.

VIBRANT is still in its early stage of development when compared to other commercialized methods. Several future lines of improvement can be anticipated. Feature-wise, further expansion of the vibrational probe palette is beneficial to increase the content and specificity of our method. The probe can be designed to target general metabolism to detect a broad range of drug-treated phenotypes or specific metabolic pathways to probe interested small molecules. Throughput-wise, the limit has not been reached yet. The current throughput can be further improved by using DFIR imaging, which has been reported to be 10–100 times faster than FTIR imaging^75–78. As we have tested that around 50 features can achieve accurate drug MOA prediction (Fig. 4f), this marvelous speed is expected to be achieved by DFIR imaging with selective frequency imaging. Resolution-wise, due to the diffraction limit, the current spatial resolution of FTIR imaging is around 5 μm, which impedes the possible use of fine intracellular features. Employing other techniques, such as advanced Raman microscopy and photothermal MIR imaging could further improve the resolution of VIBRANT. Analytical-wise, deep learning models can be introduced when larger datasets containing various drug-perturbed phenotypes are collected.

VIBRANT can also be applied in multiple other areas. For drug discovery, it can be used for lead generation after drug screening to select candidates that have been narrowed down³. The measured single-cell drug response should provide valuable insights into drug MoAs and drug resistance of the hits. In addition, further data collection combined with machine learning models can be used to explore whether our method can predict assay activity⁷⁹ and drug-drug interactions⁸⁰. In terms of precision medicine, patient-derived organoids (PDO) emerged as robust preclinical models for evaluating drug responses and improving drug treatment⁸¹. Although FTIR imaging does not have 3D sectioning ability, advanced Raman microscopy and photothermal MIR imaging^82–85 can potentially perform volumetric imaging for PDO model. Additionally, as a profiling method, VIBRANT can be applied to detect cell phenotypes perturbed by other genetic strategies, such as RNA interference⁸⁶ and CRISPR screening⁸⁷, to provide insights into genetic interactions. Overall, VIBRANT can be further tailored for applications in diverse areas.

Methods

Cell line and materials.

MDA-MB-231 (ATCC HTB-26) was purchased from ATCC. For reagents, azido-palmitic acid (1346) was purchased from Click chemistry tools; algal amino acid mixture (U-13C, 97–99%, CLM-1548) was purchased from Cambridge; deuterated oleic acid (683582) was purchased from Sigma-Aldrich. For drugs, anisomycin (A9789), bortezomib (179324–69-7), cycloheximide (01810), emetine (SMB01061), everolimus (94687), epirubicin hydrochloride (E9406), gefitinib (184475–35-2), lapatinib (231277–92-2), TVB-3166 (SML1694), taxol (PHL89806), vincristine sulfate (V8388) and apicidin (A8851) were purchased from Sigma-Aldrich. Daunorubicin hydrochloride (AAJ60224MA), iniparib (AC469161000), triacsin C (24–721-00U), doxorubicin hydrochloride (BP25165), MG 132 (AAJ63250LB0), dactolisib (NC0298104) were purchased from Fisher Scientific. For cell culture agents, DMEM medium (11965), FBS (10082), penicillin/streptomycin (1514), were purchased from ThermoFisher Scientific. CaF₂ substrates (CAFP13–1) were purchased from Crystran.

Probe preparation and media recipe for cell labeling.

Azido palmitic acid-bovine serum albumin (BSA) solution.

For the solution, couple azido-palmitic acid with BSA to prepare a 2-mM stock solution. Prepare 20 mM sodium palmitic acid solution by dissolving palmitic acid in NaOH solution with the following recipe: azido-PA (5.5 mg) + 1.0 ml dd-H2O + 35 μl 1 M NaOH. Mix and incubate the solution in 70 °C water baths until no oil droplets are visible. Then slowly add the sodium palmitic acid solution into 2.7 ml 20% BSA under room temperature water baths. Quickly add 6.3 ml DMEM culture medium and filter the solution with a 0.22-μm sterile filter.

¹³C-amino acids DMEM.

4 mg ml⁻¹ algae ¹³C-amino acids mix was dissolved in dd-H2O with 10% FBS and 1% penicillin, which matched the concentrations of regular amino acids in DMEM.

deuterated oleic acid-bovine serum albumin (BSA) solution.

For the solution, couple d₃₄-oleic acid with BSA to prepare a 2-mM stock solution. Prepare 20 mM oleic acid solution by dissolving oleic acid in NaOH solution with the following recipe: d₃₄ oleic acid (6.3 mg) + 1.0 ml dd-H2O + 24 μl 1 M NaOH. Mix and incubate the solution in 70 °C water baths until no oil droplets are visible. Then slowly add the d₃₄ oleic acid solution into 2.7 ml 20% BSA under room temperature water baths. Quickly add 6.3 ml DMEM culture medium and filter the solution with a 0.22-μm sterile filter.

Cell culture.

MDA-MB-231 cells were cultured in DMEM media supplemented with 10% FBS and 1% penicillin. Cells were grown in a humidified atmosphere containing 5% CO₂ at 37 °C in the incubator. At ~80% confluence, cells were dissociated with trypsin and passaged.

Cell viability assay and drug IC50 calculus.

The IC50 values of the drugs were determined by Alamar blue assay. Cells were seeded at 10,000 per well in 96-well plates. After 24 h, the cells were washed twice with phosphate buffered saline (PBS) and treated with drugs at different concentrations in cell culture media for 48hrs. Each drug concentration has 6–8 replicates. After the drug treatments, cells were washed with PBS twice and the cell viability was determined by Alamar blue assay following the manufacturer’s protocol (Invitrogen) using plate reader. The IC50 values were determined by fitting the data using a dose response model with variable Hill slope built in Prism.

Drug combination and synergy score calculus.

Cell viability assay was performed on cells treated by drug combinations to evaluate synergy scores. Everolimus-Lapatinib and Everolimus-Doxorubicin were chosen as model systems to study drug combinations. The concentration combinations of these two groups are shown in Supplementary Fig. 12. Each drug concentration combination has 5 replicates. To calculate synergy scores, an open-source package SynergyFinder⁷⁰ was applied. The synergy scores were calculated based on the ZIP model available in the package.

Sample preparation for drug-treated cells with labeling.

MDA-MB-231 cells were seeded on clean CaF₂ substrates with 5*10⁴ cells per well in cell culture media (DMEM, 10% FBS, 1% penicillin) overnight for control and other drug treatment conditions. Then the culture media was replaced by ¹³C-amino acids DMEM with 50 μM azido-palmitic acid, 50 μM d₃₄ oleic acid, either single drug at its IC50 concentration or drug combinations at chosen concentrations. Drugs were prepared in 100% DMSO and diluted to 0.1% DMSO in labeling media. For the control group, only cell labeling media with 0.1% DMSO was added (without any drugs). Cells were treated for 48hrs. The culturing time was selected based on the trade-off between signal and experimental time. After that, cells were fixed by 4% PFA at room temperature for 15 min and washed three times with PBS buffer and five times with dd-H2O. The samples were then air-dried before FTIR imaging.

FTIR imaging.

Agilent Cary 620 Imaging FTIR equipped with an Agilent 670-IR spectrometer and 128 × 128-pixels FPA mercury cadmium telluride (MCT) detector was used in the transmission mode. A background spectrum was collected on a clean CaF₂ substrate using 128 scans at 8 cm⁻¹ spectral resolution, suggesting that the IR absorbance were measured every 4 cm⁻¹. Cell spectra were recorded using 64–128 scans at 8 cm⁻¹ spectral resolution. A ×25 IR objective (pixel size, 3.3 μm, 0.81 numerical aperture (NA)) was used for cell imaging.

Data preprocessing.

Data preprocessing was performed using both the commercial software Cytospec and home-built MATLAB scripts with the following steps: 1) PCA noise reduction to denoise and reconstruct the spectra, the top 30 principal components were kept for imaging data collected from each condition, including control and drug treatment; 2) quality test to remove pixels with low SNR in both the fingerprint region and the cell silent region; 3) rubber-band baseline correction for spectral correction; 4) single-cell segmentation on cell images using Cell Profiler; 5) single-cell spectrum extraction applying generated single-cell masks on the processed FTIR imaging data; 6) min-max spectral data normalization for supervised machine learning.

Single cell image segmentation:

The single-peak Amide I FTIR images (1650 cm⁻¹) of cells after rubber-band baseline correction were used to perform the image segmentation using CellProfiler. The baseline correction was used to eliminate the Mie scattering effect in cells. Large-area FTIR imaging data was split into individual tile or field of view (FOV) before image segmentation, to ensure that each FOV has good segmentation results. In addition, cells that occurred at the edges of each FOV were eliminated from image segmentation, as they might not possess the whole cell contour. Meanwhile, the appropriate seeding density of 5*10⁴ on CaF₂ with 35mm diameter ensures that the seeded cells were not directly conjoint to a large extent, reducing possible segmentation errors. We also shrank one pixel between the neighboring cells in segmentation to further reduce errors. The image segmentation results are shown in Supplementary Fig. 4.

Linear unmixing.

To obtain the pure signal of the three vibrational probes, linear unmixing was performed. For ¹³C-amide I and ¹²C-amide I, the unmixing coefficients were measured from spectra of bacteria growing in the media with ¹²C₆-glucose or ¹³C₆-glucose as the only carbon source, which obtained pure ¹²C or ¹³C labeled amide peak for unmixing and has been proved sufficient for unmixing 13C-AA labeled amide peak in MDA-MB-231 cells⁴⁵. The final linear combination coefficients are:

$${}^{\mathrm{12}}\mathrm{C}-\mathrm{a}\mathrm{m}\mathrm{i}\mathrm{d}\mathrm{e}=1.2813\mathrm{*}{\mathrm{A}}_{1651}-0.5484\mathrm{*}{\mathrm{A}}_{1616}$$

$${}^{\mathrm{13}}\mathrm{C}-\mathrm{a}\mathrm{m}\mathrm{i}\mathrm{d}\mathrm{e}=1.2813\mathrm{*}{\mathrm{A}}_{1616}-0.6573\mathrm{*}{\mathrm{A}}_{1651}$$

A represents the absorbance of each peak.

For azido-PA (2096 cm⁻¹) and d₃₄OA (2092 cm⁻¹ and 2196 cm⁻¹), the unmixing coefficients were measured from spectra of cells growing in media added with either azido-PA or d₃₄OA (Supplementary Fig. 1). The final linear combination coefficients are:

$${\mathrm{d}}_{34}\mathrm{O}\mathrm{A}=1.0379\mathrm{*}{\mathrm{A}}_{2196}-0.0520\mathrm{*}{\mathrm{A}}_{2096}$$

$$\text{azido-PA}=1.0379\mathrm{*}{\mathrm{A}}_{2096}-0.7559\mathrm{*}{\mathrm{A}}_{2196}$$

3D scatter plots.

To construct the 3D scatter graph, three ratio values (azido-PA/CH₂, d₃₄OA/CH₂,¹³C amide I/(¹³C amide I + ¹²C amide I)) were calculated for each cell based on thee extracted single-cell spectrum data after baseline correction but not min-max normalization, considering that the lipid vibrational peaks will be elevated in min-max normalization. These paired ratio values of each cell were then plotted on the 3D graph using MATLAB with shaded areas indicating error ellipse with 70% confidence. Color coding was used to separate the data points from different drug treatments. Cells treated by drugs with the same MoA have similar color coding.

Machine learning.

The tests of all the machine learning classifiers were performed using Python (Scikit-learn). Specifically, single-cell data was split into 70% training and 30% testing with labels corresponding to drug MoAs. The hyper-parameters of each classifier were carefully tuned using RandomizedSearchCv in Scikit-learn. The performances of these classifiers were evaluated in the testing dataset using accuracy as the main metric. To test the performance of classifiers on batch effects, single-cell data collected from different batches were loaded as another testing dataset. For feature importances of LDA classifier, permutation feature importance (Scikit-learn) was used. The intrinsic feature importance values from random forest classifier were also extracted in Supplementary Fig. 8.

Novelty detection algorithm.

In the first step, the Mahalanobis distances between the test compound and each reference class in the annotated dataset were calculated by home-built codes in python. Drug class with minimum distance was selected. Next, isolation forest, which is available in Python, was imported to predict whether single-cell data from the test compound is known (inliers) or novel (outliers) to the selected drug class. The mean of the prediction results from single-cell data is reported as mean prediction score.

UMAP and HCA analysis.

UMAP graphs were plotted based on single-cell spectrum data using available packages in Python. Both full-spectrum data using probing approach and label-free approach were tested. Data were z-score normalized as input for UMAP. HCA dendrograms were constructed on averaged single-cell spectrum data using packages in Scipy (scipy.cluster.hierarchy). For the probing approach, all the 288 spectral features in fingerprint and probed regions were used. For the label-free approach, all the 209 spectral features in the fingerprint region were used.

Statistical analysis.

In Fig. 2, for the control group, the cell number is 1000; for anisomycin, the cell number is 589; for cycloheximide, the cell number is 700; for bortezomib, the cell number is 583; for MG-132, the cell number is 579; for doxorubicin, the cell number is 500; for epirubicin, the cell number is 950; for triacsin-C, the cell number is 562; for TVB3166, the cell number is 474; for dactolisib, the cell number is 443; for everolimus, the cell number is 495; for gefitinib, the cell number is 821; for lapatinib, the cell number is 951; for olaparib, the cell number is 417. The total cell number is 9,064 for all 14 groups (with control)

In Fig. 3a, 3c, 3e, the statistics of data for specific drug treatment from the probing approach is the same as in Fig. 2. In Fig 3b, 3d, the statistics of the label-free approach are: control (727 cells), anisomycin (557 cells), cycloheximide (698 cells), bortezomib (585 cells), MG-132 (611 cells), dactolisib (637 cells), everolimus (346 cells), doxorubicin (706 cells), epribucin (412 cells), tiracsinc-C (470 cells). The total number is 5,749 cells. In Fig. 3f (probing approach), batch 1 data contains 6,910 cells; batch 2 data contains 4,511 cells. In Fig. 3g (label-free approach), batch 1 data contains 5,749 cells; batch 2 data contains 2,982 cells.

In Fig. 4, to ensure class balances, the single-cell data number from each drug MoA is confined to be around 500–700, making up a dataset with cell number of 6,298 in total. In Fig. 5, the statistics of test compounds are: daunorubicin (512 cells), emetine (370 cells), iniparib (397 cells), apicidin (978 cells), taxol (720 cells), vincristine (487 cells). The total cell number is 3,464. In Fig. 6, the statistics for drug combinations are: control (609 cells), everolimus (740 cells), lapatinib (873 cells), doxorubicin (485 cells), Eve-Lap (1–1) group (988 cells), Eve-Lap (1–2) group (976 cells), Eve-Dox (2–1) group (471 cells), Eve-Dox (1–2) group (459 cells). The total cell number is 5,601.

For experiment replicates, experiments in Fig. 2 have been repeated in 3 different batches with similar results. The statistics of supplementary figures are included in supporting information.

Data availability:

The raw FTIR imaging data generated in this work are available from the corresponding author on reasonable request. The processed single-cell FTIR spectrum data are available at https://github.com/MinLabColumbia/VIBRANT.