High-throughput combinatorial screening of antiplatelet drugs for personalized medicine

Abstract

Cardiovascular disease (CVD) remains the leading cause of death worldwide. Platelet activation plays a critical role in arterial thrombotic events such as myocardial infarction. Although antiplatelet drugs are standard therapies, they are associated with risks including bleeding, gastrointestinal adverse effects, and drug resistance. Furthermore, substantial inter-individual variability in patient responses underscores the need for personalized antiplatelet regimens. These factors emphasize the importance of screening for optimal antiplatelet drugs and drug combinations tailored to individual patients. However, traditional platelet detection assays are reagent-hungry and low-throughput, making them unsuitable for high-throughput screening of antiplatelet agents. Here, we present the C-chip, a high-throughput platform for on-chip parallel screening of antiplatelet drug combinations. The C-chip miniaturizes individual screening reactions into picoliter-volume, color-coded droplets, enabling the generation of thousands of screening data points in a single experiment. We demonstrate that the C-chip can effectively identify the optimal combinations of three clinically relevant antiplatelet drugs: Aspirin, Tirofiban, and Ticagrelor. We further applied this platform to identify optimal drug combinations for five healthy volunteers, revealing marked inter-individual variability in antiplatelet drug responses.

Introduction

Cardiovascular diseases (CVD) remain the leading cause of mortality worldwide, with a high prevalence of arterial thrombotic disorders¹. In patients with CVD, pathological platelet activation plays a pivotal role in thrombus formation^2,3. Given the complexity and heterogeneity of platelet activation pathways^3,4, various antiplatelet drugs, such as aspirin, clopidogrel, cilostazol, and tirofiban, have been developed to target distinct molecular mechanisms^5,6. However, interindividual variability in drug response, along with adverse effects including bleeding, gastrointestinal complications, granulocytopenia, and multidrug resistance, poses significant challenges in identifying optimal therapeutic strategies^5,7–11. These limitations underscore the urgent need for improved approaches to antiplatelet therapy selection that support personalized treatment¹². Currently, patients with high thrombotic risk are often prescribed dual antiplatelet therapy based on population-level evidence to ensure efficacy¹³; however, both the therapeutic benefit and the risk of adverse effects remain unpredictable at the individual level.

Currently, three main types of platelet function assays exist: aggregation-based assays (e.g., light transmission aggregometry [LTA], VerifyNow), biomarker-based assays (e.g., flow cytometry to detect P-selectin), and biomechanical-based assays (e.g., thrombelastogram [TEG], PFA-200 system)^14–20. However, these methods share limitations, including low throughput, slow detection, and high reagent costs. While LTA (the gold standard) requires large sample volumes and complex procedures, which limits its clinical use²¹. Biomarker-based assays (e.g., flow cytometry, ELISA) assess platelet function at the population level but are expensive, difficult to operate, and difficult to implement clinically^22–24. Proof-of-concept platforms like VerifyNow and PFA-200 offer faster results but are narrow in scope, costly, and unable to replace traditional aggregation assays^23,25,26. Despite these options, clinical studies confirm current platelet function tests offer limited value for guiding patient medication^27,28, highlighting the need to develop new high-throughput assays for combinatorial antiplatelet therapy screening to advance personalized treatment and drug discovery.

High-throughput screening (HTS) is widely used in drug discovery^29–32. It has been successfully applied to single-drug screening using 96- and 384-well plate formats³³, but these formats are not well-suited for evaluating drug combinations. To address this gap, microfluidic technologies have been introduced to enable high-throughput combinatorial drug screening, utilizing platforms such as programmable microvalves^34–37, microwells³⁸, micropillars^39,40, and multilayered channels³⁷. Among microfluidic strategies, droplet-based systems have recently been developed to assess drug effects on single cells or multicellular aggregates^41,42. These systems allow individual cells to be encapsulated in discrete droplets, enabling thousands of parallel, miniaturized experiments. Notably, Xie et al. further advanced droplet-based combinatorial screening with CP-seq, which used microwell-based droplet random pairing and single-cell RNA sequencing for deep transcriptomic profiling of drug-perturbed cells, though it focused on adherent cell lines rather than platelets and relied on costly sequencing⁴². Despite such progress in droplet-based screening, these assays have been effectively applied to general cell phenotyping but not widely used in platelet research^43,44. Recently, Jiang et al. developed the NebulaPlate chip, a droplet microfluidic platform specifically designed for platelet function analysis⁴⁵, providing a foundation for platelet-focused microfluidic screening but lacking combinatorial drug evaluation capabilities.

Here, we developed a combinatorial screening chip (C-chip) by integrating microdroplet-based high-throughput screening technology with a combinatorial fluorescence encoding strategy. This platform enables the simultaneous assessment of platelet activation across multiple drug combinations in a highly multiplexed format, providing a promising tool for personalized antiplatelet therapy. The C-chip encapsulates platelets and fluorescently encoded drugs into discrete droplets, which are subsequently paired within custom-designed microwells. In each microwell, two drug droplets are randomly paired with one platelet-containing droplet, enabling combinatorial drug perturbation. The effects of different drug combinations on platelet aggregation are then quantitatively assessed. We first demonstrated the effective manipulation of droplets, with the majority of microwells being successfully used for droplet pairing and merging. We then validated the C-chip’s sensitivity by performing dose-response analyses of individual antiplatelet drugs. As a proof-of-concept, we demonstrated that the C-chip could identify the most effective drug combinations using two concentrations of three clinically approved agents. Looking ahead, this combinatorial fluorescence encoding strategy could enable the screening of thousands of drug combinations on a single chip the size of a standard glass slide. We envision the C-chip as a versatile high-throughput platform for antiplatelet drug discovery and a valuable tool in advancing precision antiplatelet therapy.

Results

Design and fabrication of the C-chip

To enhance the scalability of assays, we employed the combinatorial fluorescence encoding strategy developed by our group⁴⁶. This strategy enables ultra-multiplex fluorescence encoding, significantly increasing the number of experimental conditions that can be tested in a single assay (Fig. 1). For our fluorescent markers, we selected Alexa 488 (Ex/Em: 495/519) and Alexa 594 (Ex/Em: 590/618) due to their distinct spectral profiles and robust signaling at specific wavelengths, which are critical for analyzing complex biological samples. We modulated the dye concentrations to exploit variations in fluorescence intensity, facilitating the differentiation between multiple encoding levels. This methodology involved selecting three distinct intensities for each fluorophore, enabling six unique barcoding capacities suitable for platelet drug combination assays. Each antiplatelet drug solution was mixed with a specific fluorophore type and intensity, encoding both the drug molecule and its concentration (Table S2). The mixture was then dispensed into droplets, as shown in Fig. 1a. The continuous phase consisted of fluorinated oil supplemented with a surfactant to prevent premature droplet coalescence. The antiplatelet drug droplets were pooled together to form a drug library in droplet form. Meanwhile, platelets were encapsulated in larger droplets than the drug droplets.

Fig. 1. Overview of the combinatorial screening chip (C-chip).

a Fluorescent barcode drug droplets and platelet droplets with varying diameters were generated using microfluidics. The drugs were pre-mixed with different fluorescent markers. b Platelet droplets were loaded into the microwell array device and occupy the larger microwells, followed by the loading of drug droplets that occupy the smaller microwells. Each small microwell randomly captured one drug droplet. Through corona treatment, the droplets in each microwell unit were merged, enabling combinatorial drug treatment for platelets and fluorescent labeling. c Using a fluorescence microscope, the aggregation and dispersion of platelets were continuously observed and recorded, with real-time images captured for each area of the chip. Subsequently, decoding software was used to re-identify the platelets and calculate the anti-aggregation efficiency of platelets for each individual drug or drug combination

The C-chip was a microfluidic device with an upper layer of downward-facing microwell arrays and a lower layer of flow chambers (Fig. 1b). Each microwell unit included a circular chamber (80 μm lateral dimension, 70 μm depth) for capturing larger “main droplets” (platelet-containing) and two identical semicircular chambers (40 μm diameter, 40 μm depth) for smaller “encoding droplets” (drug-containing; Fig. S1). A C-chip, approximately the size of a standard glass slide, contained 2106 microwell units (39 × 54) (Fig. S1). Platelet droplets, approximately 80 μm in diameter, were first loaded into the flow chamber at 1000 μL/h (Fig. S2). Due to the lower density of water compared to the continuous oil phase (≈ 1.6 g/mL), platelet droplets were loaded into the circular chamber of individual microwells, leaving the semicircular chambers empty. Smaller drug droplets, approximately 40 μm in diameter, were then introduced at 1000 μL/h, randomly occupying the semicircular chambers (Fig. S2).

After loading, the C-chip was imaged via fluorescence microscope (Fig. 1c), followed by corona treatment to induce droplet fusion. Fluorescence microscopy confirmed complete droplet fusion, and platelet aggregation was imaged at the experiment’s conclusion. By distinguishing fluorescence intensities, we generated codes, and an AI-based aggregation evaluation model was then used to automatically decode drug combinations and assess the degree of platelet aggregation in each droplet.

Platelet microdroplet optimization and activity preservation

Maintaining platelet activity is critical for evaluating antiplatelet drug efficacy. We therefore characterized and optimized key parameters for droplet manipulation. Uniform droplets were generated using flow-focusing microfluidic devices operated under negative outlet pressure (Fig. 2a). The resulting platelet droplets and drug droplets had diameters of 77.01 ± 0.92 μm (n = 500) and 40.54 ± 1.11 μm (n = 499), respectively, with a coefficient of variation (CV) < 5% (Fig. 2b and Fig. S3). Each platelet droplet contained 66.40 ± 5.65 platelets, consistent with Poisson statistics (Fig. S4), and no red blood cells (RBCs) or white blood cells (WBCs) were detected in the platelet suspension (Fig. S5).

Fig. 2. Optimization of platelet microdroplet generation.

a Micrographs of the generated platelet droplets and drug droplets. The scale bar for all images is 50 μm. b Diameter distribution of platelet droplets (n = 500) and drug droplets (n = 499). c Platelet aggregation of washed control (without corona treatment, n = 6) and corona-treated platelets (n = 6) measured by aggregometry. d Platelet aggregation of washed control (n = 6) and fluorophore-treated platelets (n = 6) measured by aggregometry. e Platelet droplets exposed to various thrombin concentrations (0, 0.01, 0.05, 0.1, 0.2, 0.5, 1, and 5 U/mL). f Variation in the number of platelets within droplets after exposure to a constant thrombin concentration. The data are expressed as mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 vs. 0 U/mL group

While droplet-based cell culture is well-studied, the impact of corona treatment on platelet function was previously unelucidated. To verify that platelets remain functionally responsive after corona treatment, we performed platelet aggregation assays via light transmission aggregometry (LTA) (Fig. 2c). The mean aggregation efficiency for the corona-treated group and the untreated control group was 70.83 ± 8.38% and 71.67 ± 5.99%, respectively, with no statistically significant difference (Fig. 2c, inset). Drugs and fluorescent dyes were premixed as a single aqueous solution immediately before droplet formation; the resulting encoded drug droplets were then fused with platelet droplets via corona treatment. To evaluate whether fluorescent dyes themselves affect platelet activity, we mixed 1 µM Alexa 488 with platelets and performed LTA. As shown in Fig. 2d, the mean aggregation efficiency of the Alexa 488-treated group and the control group was 69.00 ± 9.96% and 71.83 ± 6.46%, respectively, with no statistically significant difference (Fig. 2d, inset). These results confirm that the fluorescent dye does not interfere with platelet function.

To further validate platelet functionality in our system, we simulated physiological conditions by titrating thrombin to determine the optimal concentration for platelet activation (Fig. 2e). In the absence of thrombin, platelets remained dispersed, and aggregation increased progressively with thrombin concentration. As shown in Fig. 2f, at 0.01 U/mL thrombin, the anti-aggregation efficiency was approximately 55.50 ± 4.97%; at 0.2 U/mL, it decreased to 15.50 ± 1.51%; and at 5 U/mL, it dropped further to 2.30 ± 1.49%. Based on these results, aimed at maintaining platelet responsiveness without excessively suppressing drug effects, 0.2 U/mL thrombin was selected as optimal²². Additionally, we optimized corona treatment duration to maximize fusion efficiency while preserving platelet activity. As shown in Fig. S6, a 5-s corona treatment enabled fusion in 88.2 ± 2.57% of droplet pairs; this sub-optimal fusion rate is mainly attributable to minor droplet-size variations that restrict physical contact between pairs, while 10 s yielded a similar rate (87.2 ± 3.45%) with no added benefit. Therefore, 5 s was selected as the optimal corona treatment duration.

Droplet manipulation and AI-based aggregation analysis on the C-chip

Robust droplet manipulation is essential for C-chip operation. As shown in Fig. 3a, after platelet droplet loading, >99% (2087/2106) of the microwells contained a platelet droplet. After drug droplet loading, >92% (1939/2106) of the microwells successfully captured two drug droplets. We then applied a brief corona treatment to the microwell array device, which destabilized interfaces and induced subsequent droplet fusion. By slightly tilting the device and ensuring all three droplets (1 platelet + 2 drug) were in contact, nearly 90% (1893/2106) of microwells achieved successful droplet fusion post-treatment.

Fig. 3. The performance of the C-chip and optimization of the AI-based aggregation evaluation model.

a Micrographs showing the process of blank droplet loading, drug droplet loading and droplet merging. Scale bar for all images is 100 μm. b AI-based platelet aggregation evaluation model trained via transfer learning from a general image classification network. The model predicted the probability distribution of platelet states within microwells after decoding the fluorescent barcodes. It classified four distinct states: (1) droplet with aggregated platelets (clustered); (2) droplet with dispersed platelets (dispersed); (3) droplet present but no platelets (drop wo_cell); and (4) microwell without any droplet (wo_drop). c Training and validation accuracy of the AI-based platelet aggregation model. d Training and validation loss of the AI-based platelet aggregation model

Microscopic imaging was performed pre-droplet fusion and post-aggregation. To enhance decoding accuracy and recognition efficiency, we employed an AI-based aggregation evaluation model with a pre-trained MobileNetV2 model for image classification (Fig. 3b). The model identified four distinct platelet states (Fi. S7): (1) Microwells with droplets but no platelets; (2) Empty microwells; (3) Microwells with droplets and aggregated platelets; and (4) Microwells with droplets and dispersed platelets. As shown in Fig. 3c, d, the model demonstrated significant improvement during the training process. In the feature extraction phase, the validation accuracy reached 88.73%, with a validation loss of 0.4346. After fine-tuning, the validation accuracy further increased to 98.59%, and the validation loss decreased to 0.0368. Notably, the model achieved 100% test accuracy, demonstrating excellent generalization. Learning curves showed steady increases in accuracy and decreases in loss across training epochs, confirming that the model effectively learned and adapted to the dataset. These results confirm that the AI model efficiently identified the state of platelet aggregation, reduced misclassifications, and minimized the occurrence of false positives.

Generation of antiplatelet titration curves on the C-chip

We next validated the performance of the C-chip in determining IC50 values of Aspirin, Tirofiban, and Ticagrelor. These drugs are commonly prescribed antiplatelets with distinct mechanisms for blocking platelet aggregation. Aspirin inhibits thromboxane A2 (TXA2) synthesis in platelets by suppressing cyclooxygenase (COX), particularly COX-1, thereby blocking platelet aggregation⁴⁷. Tirofiban, a GPIIb/IIIa receptor antagonist that prevents platelet aggregation by blocking binding of GPIIb/IIIa receptors on the platelet surface to fibrinogen and other platelet-activating molecules⁴⁸. Ticagrelor, a P2Y12 receptor antagonist that reduces platelet activation and aggregation by inhibiting ADP (adenosine diphosphate) binding to platelet P2Y12 receptor⁴⁹.

Using single-color, three-intensity barcoding, we encoded low/medium/high drug concentrations into three fluorescence levels. Random pairing of two droplets per well generated 10 dose combinations: blank, 3 single-dose, 3 mixed-dose and 3 identical-dose states, of which 9 drug-containing points formed the dose-response curve (Fig. 4a). Concentrations ranged from 0 to 40 µM (aspirin) and 0–20 µM (tirofiban/ticagrelor). Antiplatelet efficacy was assessed via platelet anti-aggregation efficiency. As shown in Fig. 4a, the initial concentrations of Aspirin tested were 0.2, 2, and 20 µM, showing platelet anti-aggregation efficiencies of 7.33 ± 11.50%, 50.65 ± 7.15%, and 73.72 ± 5.76%, respectively. The highest concentration of 40 µM achieved an anti-aggregation efficiency of 99.05 ± 0.88%. Similarly, for Tirofiban, the initial drug concentrations used were 0.1, 1, and 10 µM, showing platelet anti-aggregation efficiencies of 20.81 ± 2.44%, 71.37 ± 4.61%, and 80.75 ± 2.69%, respectively. The highest concentration of 20 µM reached 97.01 ± 3.28% (Fig. 4c). For Ticagrelor, the anti-aggregation efficiencies were 10.90 ± 4.15% at 0.1 µM, 49.59 ± 4.38% at 1 µM, 76.82 ± 3.78% at 10 µM, and 94.75 ± 5.71% at 20 µM (Fig. 4e).

Fig. 4. IC50 evaluation of different antiplatelet drugs on the C-chip.

a The concentrations of Aspirin tested were 0, 0.2, 0.4, 2, 2.2, 4, 20, 20.2, 22, and 40 μM. b Nonlinear fitting curve for determining the IC50 value (2.149 μM) of Aspirin. c The concentrations of Tirofiban tested were 0, 0.1, 0.2, 1, 1.1, 2, 10, 10.1, 11, and 20 μM. d Nonlinear fitting curve for determining the IC50 value (0.0997 μM) of Tirofiban. e The concentrations of Ticagrelor tested were 0, 0.1, 0.2, 1, 1.1, 2, 10, 10.1, 11, and 20 μM. f Nonlinear fitting curve for determining the IC50 value (0.9242 μM) of Ticagrelor. The data were expressed as mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 vs. 0 μM group

Next, we fitted these data to a sigmoid function to calculate IC50 values. In pharmacology, IC50 (half-maximal inhibitory concentration) quantifies a drug’s inhibitory capacity on biological processes. Here, IC50 refers to the concentration that halves maximal thrombin-induced aggregation amplitude, a standard in platelet pharmacology that is commonly used to assess the inhibitory capacity of a drug or inhibitor on a specific biological process^50,51. We calculated IC50 values for each drug separately. The IC50 value for Aspirin was 2.1490 µM, with R² = 0.9516 (Fig. 4b), the IC50 value for Tirofiban was 0.0997 µM, with R² = 0.9423 (Fig. 4d), and the IC50 value for Ticagrelor was 0.9242 µM, with R² = 0.9678 (Fig. 4f). These values align with LTA literature (within typical one- to three-fold inter-laboratory variation), confirming the C-chip validly quantifies antiplatelet properties of the three drugs (Table S1)^50–52.

Personalized antiplatelet drug screening using C-chip

Clinically, antiplatelet drugs are typically prescribed at low concentrations to CVD patients, whereas higher concentrations are often used for acute myocardial infarction. Combinatorial therapy, such as Aspirin and Ticagrelor, is frequently prescribed, but optimal combinations are usually determined empirically. To validate whether the C-chip can enable rapid determination of optimal antiplatelet pairings for individuals, we selected two concentrations for each of Aspirin, Ticagrelor, and Tirofiban, aiming to identify combinations that deliver optimal anti-aggregation efficacy.

To achieve this, we used a dual-encoding strategy (fluorophore type and intensity) to label the three drugs. Fluorescence coding was designed with fluorophore type corresponding to drug concentrations and intensity encoding drug identity (Table S2). Specifically, 20 μM and 2 μM Aspirin were labeled with 5 μM Alexa 488 and 5 μM Alexa 594, respectively. Similarly, Tirofiban was labeled with 3 μM Alexa 488/594, while Ticagrelor was labeled with 1 μM Alexa 488/594 (Fig. S8). This approach yielded 27 combinations: 6 single-drug treatments, 6 identical two-droplet (same drug) pairs, and 15 unique two-drug pairs.

Next, we determined optimal antiplatelet combinations for five healthy volunteers via the C-chip (Table S3). We decoded fluorescence barcodes and quantified platelet anti-aggregation values for all combinatorial pairs using the AI-based aggregation analysis model, then plotted results for each volunteer (Fig. 5 and Fig. S9).

Fig. 5. Combinatorial screening of antiplatelet drugs on the C-chip.

a Heatmap of drug combination tests for volunteer 1. b Anti-aggregation efficiency of single-drug combinations for platelets in the combinatorial screening tests for Volunteer 1. For Aspirin (Asp), the concentrations were 2 μM, 4 μM (2 μM + 2 μM), 20 μM, 22 μM (2 μM + 20 μM), and 40 μM (20 μM + 20 μM). For Tirofiban (Tif), the concentrations were 0.2 μM, 0.4 μM (0.2 μM + 0.2 μM), 2 μM, 2.2 μM (0.2 μM + 2 μM), and 4 μM (2 μM + 2 μM). For Ticagrelor, the concentrations were 1 μM, 2 μM (1 μM + 1 μM), 10 μM, 11 μM (10 μM + 1 μM), and 20 μM (10 μM + 10 μM). c Anti-aggregation efficiency of two-drug combinations for platelets in the combinatorial screening tests for Volunteer 1. Based on the concentration of drug combinations, the two-drug combinations were divided into three groups: low concentration group (low group) including 2 μM Asp + 0.2 μM Tif, 2 μM Asp + 1 μM Tig, 0.2 μM Tif + 1 μM Tig; high concentration group (high group) including 20 μM Asp + 2 μM Tif, 20 μM Asp + 10 μM Tig, 2 μM Tif + 10 μM Tig; combinations of low and high concentrations (medium group) including 2 μM Asp + 10 μM Tig; 2 μM Tif + 1 μM Tig; 2 μM Asp + 2 μM Tif; 20 μM Asp + 1 μM Tig; 20 μM Asp + 0.2 μM Tif; 0.2 μM Tif +10 μM Tig. d Total heatmap of the anti-aggregation efficiency of two-drug combinations for the five volunteers. e The therapeutic effects of different drug combinations across various samples were compared, with samples having an anti-aggregation degree greater than 50% scored as 1 (indicated in orange) and those less than 50% scored as 0 (indicated in gray). f The top three drug combinations with the highest anti-aggregation efficiency in the medium group for the five volunteers

We generated a heatmap of all drug combinations for Volunteer 1 (Fig. 5a). Based on droplet drug concentrations, we separated single-drug treatments from pairwise combinations, then divided the latter into three groups: low group (combinations involving low concentrations of all three drugs), high group (combinations involving high concentrations of all three drugs), and medium group (all other combinations). As shown in Fig. 5b, 4 μM Tif (Tirofiban) exhibited the strongest platelet aggregation inhibition. In the pairwise combinations (Fig. 5c), the medium group showed that all combinations had >50% anti-aggregation, with 0.2 μM Tif + 10 μM Tig (Ticagrelor) reaching 86.93 ± 12.47%. In the high group, all multi-drug combinations had over 70% anti-aggregation, with 20 μM Asp (Aspirin) + 2 μM Tif achieving 89.12 ± 8.05%. Overall, Volunteer 1 was highly sensitive to Tirofiban, with strong anti-aggregation in Tirofiban-containing combinations. Data for the remaining four volunteers were presented in Fig. S10. For example, for Volunteer 2, two-drug combinations (e.g., 10 μM Tig + 2 μM Tif) outperformed single-drug treatments (e.g., 20 μM Tig). Figure S10c revealed that, except for the 2 μM Asp + 0.2 μM Tif combination, all pairwise combinations exhibited anti-aggregation greater than 50%. Notably, in the medium group, the anti-aggregation of 2 μM Asp + 10 μM Tig, 0.2 μM Tif + 10 μM Tig, and 2 μM Tif + 1 μM Tig reached 77.94 ± 4.01%, 79.45 ± 3.01%, and 87.40 ± 4.98%, respectively. In the high group, the anti-aggregation of 2 μM Tif + 10 μM Tig was 97.12 ± 3.45%, confirming that combination therapy was more effective than high-dose single-drug therapy for this volunteer. As shown in Fig. S10f, compared to the first two volunteers, Volunteer 3 exhibited relatively low platelet anti-aggregation in both low- and medium-concentration groups. The highest anti-aggregation (59.44 ± 2.98%) was observed for 0.2 μM Tif + 10 μM Tig; in the high-group two-drug combinations, the highest was 70.89 ± 2.89% (2 μM Tif + 10 μM Tig), and single-drug 20 μM ticagrelor achieved 71.82 ± 4.67% (Fig. S10e, f). This indicated Volunteer 3 was less sensitive to both high-concentration single-drug and two-drug treatments.

Across all 27 combinations, mean anti-aggregation efficiencies ranged from 20.05 ± 9.15% to 87.54 ± 10.44%; 23 combinations had 95% confidence interval (CI) lower bounds >50% and Cohen’s d = 2.51–17.04, confirming robust antiplatelet efficacy (Table S4). However, when screening for optimal antiplatelet drug combinations tailored to individual needs, inter-individual variability in drug responses became evident (Fig. 5d). As a positive control, high concentrations of Aspirin, Tirofiban and Ticagrelor completely blocked platelet aggregation, though these doses carried clinical bleeding risks. In contrast, the three drugs failed to inhibit aggregation in all volunteers except Volunteer 1 (Fig. 5d). To identify effective yet safe combinations, we used a 50% anti-aggregation rate as the threshold. Combinations marked in orange indicated >50% platelet aggregation inhibition, while gray combinations showed <50% anti-aggregation (Fig. 5e). Among the 27 combinations, 9 (40 μM Asp, 4 μM Tif, 20 μM Tig, 0.2 μM Tif + 1 μM Tig, 0.2 μM Tif + 10 μM Tig, 2 μM Tif + 1 μM Tig, 20 μM Asp + 2 μM Tif, 20 μM Asp + 10 μM Tig, 2 μM Tif + 10 μM Tig) achieved >50% anti-aggregation in all volunteers (Fig. 5e). Person-to-person heterogeneity was still evident when we examined the top three drug combinations in medium drug combination group (Fig. 5f). Notably, the combination of 0.2 µM Tif + 10 µM Tig achieved >75% anti-aggregation in all five volunteers, emerging as a universal medium-dose candidate for future clinical validation.

Discussion

In this study, we developed the C-chip, a high-throughput microfluidic platform for personalized antiplatelet drug screening. We benchmarked C-chip against standard clinical assays on a per-combination basis (Table S5). A single C-chip run yielded over 1900 evaluable microwells, and the total experiment time (droplet generation, pairing, fusion and imaging) was about 2 h, resulting in a throughput of 1000 wells/h. Using a small blood sample (2–3 mL), we generated about 1 × 10⁷ droplets, theoretically sufficient for about 4700 C-chip runs (1 × 10⁷ droplets ÷ 2160 microwells per chip). Each C-chip consumed 20 µL platelet suspension, versus 300 µL for LTA^22,53 or 100 µL for flow cytometry⁵⁴, reagent costs (dyes + surfactant) were about 0.10 per combination, over 100-fold lower than LTA (> $10) or flow cytometry ( > $100)⁵⁵.

In high-throughput platelet-function testing, microplate-based aggregation assays have become the default⁵⁶, yet they infer aggregation from turbidity, a readout easily biased by colored drugs, and lack direct platelet count quantification⁵⁷. The C-chip consumed only 20 µL of platelet suspension (≥ 40 µL per well in microplates). Each drug combination was fluorescence-encoded and randomly distributed across about 2000 microwells, enabling a full aggregation distribution via one automated scan, whereas achieving equivalent statistical power with conventional plates would require 2000 manual pipetting steps and 80 mL of blood.

In clinical practice, the three antiplatelet drugs have well-established therapeutic doses and plasma concentrations. For Aspirin, 75–100 mg orally once a day, total plasma concentration is 70–140 ng/dL (3.9–7.8 µM)^58,59; For Tirofiban, intravenous administration of 25 µg/kg over 5 min results in a total plasma concentration of 20 nM–50.5 µM)^60,61; For Ticagrelor, oral administration of 60–90 mg twice daily results in a total plasma concentration of 5–5000 ng/mL (9.6 nM–9.6 µM)^62,63. The medium-dose combination identified in our study aligns well with clinical therapeutic ranges. Notably, some in vitro droplet concentrations (e.g., 20 μM Aspirin) exceed clinical total plasma concentrations. This discrepancy stems from the protein-free nature of droplets (platelets in Tyrode’s Buffer), where 100% of the drug remains free and active. In contrast, about 80% of Aspirin binds to plasma proteins clinically, leaving only 20% as free active drug⁶⁴. Moderately higher in vitro concentrations were thus necessary to match the clinically relevant free drug levels, ensuring the C-chip’s antiplatelet effects are biologically translatable to clinical settings.

Building on the current work, the C-chip has room for further improvement and expansion of the C-chip. We previously established DropAI for cell-free gene-expression optimization, where 9 intensity levels × 4 fluorescent dyes deliver 6561 codes per run in 250 pL droplets at 300 Hz, and the lysate tolerates dyes up to 20 µM without time constraints⁴⁶. However, C-chip, designed for living-platelet phenotyping, currently uses the matrix to 27 fluorescence-coded combinations to ensure biological and statistical reliability. With 2160 microwells per chip, assigning wells to 27 conditions yields ~80 physical replicates per combination, providing a robust aggregation value distribution. Scaling to 9-intensity × 4-dye encoding (≥ 6561 codes) would reduce replicates to <0.5 replicates, destroying statistical power. Moreover, unlike our previous cell-free system, human platelets lose agonist responsiveness within 120 min of leaving 37 °C, completing droplet generation, pairing, fusion, and imaging for thousands of extra codes would exceed this viability window, producing artifactual loss-of-function rather than true drug effects. To address this limitation, we will integrate far-red dyes, a six-color line-scan confocal reader and >10,000-microwell C-chip formats, enabling nine-level encoding while keeping replicate counts and total assay time within the 120-min biological limit.

Additionally, during combinatorial drug screening, the antiplatelet effects of drug combinations varied substantially among healthy volunteers. While our data supported the feasibility of personalized antiplatelet combination selection, further validation in larger cohorts of healthy volunteers and CVD patients is needed. Given its low platelet consumption per reaction, the C-chip could be used for preclinical sensitivity testing in patients, guiding personalized clinical medication decisions. Finally, our experiments used only thrombin as a platelet agonist; future studies and clinical applications should test additional agonists (e.g., ADP, collagen) to mimic diverse clinical thrombotic triggers.

Conclusion

In summary, the C-chip platform offered a high-throughput and efficient route to personalized antiplatelet drug screening. It is a valuable tool for personalized antiplatelet therapy and serves as a universal combinatorial screening tool applicable to a wide range of related applications, including drug discovery and diagnostics.

Methods

Design and fabrication of microfluidic devices

Three microfluidic devices were designed and employed in this study: two droplet generators for platelet encapsulation and drug droplet generation, and a microwell reaction array for droplet fusion and data collection. These devices, shown in Figs. S1 and S2, were designed using AutoCAD software. The microfluidic droplet-generating component was fabricated through soft lithography using polydimethylsiloxane (PDMS). First, the negative photoresist SU-8 3025 (MicroChem) was spin-coated onto a 7.62 cm silicon wafer (University Wafer) at 4000 rpm, forming a uniform SU-8 layer approximately 20 μm thick. After a pre-bake at 95 °C for 10 min, the wafer was exposed to UV light (M365L2, Thorlabs) at 120 mW for 2 min through a plastic photomask. A post-bake at 95 °C for 10 min followed, after which the mold was developed using SU-8 developer solution (MicroChem) for about 15 min. The resulting master mold was cleaned with isopropanol and ethanol, then blow-dried with nitrogen. The PDMS precursor (SYLGARD 184, Dow Corning) was mixed with a curing agent in a 10:1 weight ratio, poured onto the master mold, and cured at 60 °C for 3 h. After curing, the PDMS slab was detached from the mold, and the inlet and outlet ports were created using a 0.7 mm hole punch. The PDMS slab was then bonded to a clean glass slide using oxygen plasma treatment. Prior to use, the fabricated devices were treated with Aquapel (PPG Industries), a commercial water repellent, to render the channel surfaces hydrophobic.

Platelet extraction and processing. Blood was collected from healthy volunteers via venipuncture and mixed with White’s Buffer in a 9:1 (v/v) ratio. The mixture was then thoroughly mixed with an equal volume of preheated 0.9% NaCl (37 °C). Apyrase (A6535, Sigma) and Prostaglandin E1 (P5515, Sigma) were added to achieve final concentrations of 1 U/mL and 0.1 μg/mL, respectively. The mixture was subsequently centrifuged at 290 rcf and 26 °C for 10 min. After centrifugation, the platelet-rich plasma (PRP) was collected from the top layer. Apyrase and ethylenediaminetetraacetic acid (EDTA, ST066, Beyotime) were added to achieve final concentrations of 1 U/mL and 5 mM, respectively. The mixture was then centrifuged again at 850 rcf and 26 °C for 10 min, and the supernatant was discarded. The platelets were resuspended in Tyrode’s Buffer, preheated to 37 °C. Platelet counts were determined using a hematology analyzer (XN-1000VB1, Sysmex). Resuspended samples were stained with 2 µg/mL DAPI and examined under bright-field and fluorescence microscopy (IX83, 40× objective). All blood collection procedures were approved by the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (SH9H-2019-T160-6), and informed consent was obtained from all participants.

Fluorescence encoding

In the single-drug testing mode, three concentrations of each drug were selected, and different intensities of a fluorescent dye (Alexa 488; Thermo Fisher Scientific) were used to label the drug concentrations. Non-linear fitting was performed using the different drug concentrations to calculate the IC50 values.

For multi-drug combination testing, antiplatelet drugs were doubly encoded with different fluorescent dyes and intensities. Specifically, three antiplatelet drugs—Aspirin (HY-14654, MedChemExpress), Tirofiban (HY-17369B, MedChemExpress), and Ticagrelor (HY-10064, MedChemExpress)—were used, with each drug tested at two concentrations (high and low). Two fluorescent dyes, Alexa 488 and Alexa 594 (Thermo Fisher Scientific), were used to label the high and low concentrations of each drug, respectively, with varying intensities distinguishing the different drug types. Thrombin (T4648, Sigma) was used to stimulate platelet aggregation in the experiments. Detailed drug concentrations and fluorescent labeling conditions are provided in Table S2.

Droplet generation and manipulation

The aqueous phase of the droplets consisted of the previously mentioned buffer solution and platelet solution, while the oil phase was composed of 2% PEG-PFPE. Each solution was drawn into syringes and connected to the device inlets via polytetrafluoroethylene (PTFE) tubing. The flow rates of the oil phase (800 μL/h) and the aqueous phase (400 μL/h) were adjusted to generate drug droplets with a diameter of 40 μm and platelet droplets with a diameter of 80 μm, respectively, by applying positive pressure using a syringe pump (NE-510, New Era). The droplets were then collected through the device outlet, which was also connected via PTFE tubing.

The collected platelet droplets were first transferred into a 10 μL pipette tip, which was then connected to the device inlet. Negative pressure at a flow rate of 1000 μL/h was applied at the device outlet using a syringe pump to load the platelet droplets onto the chip. Once the platelet droplets were loaded, drug droplets were introduced into the pipette tip, and the flow rate was maintained until all droplets were loaded. The negative pressure was then released. The C-chip, containing the captured droplets, was placed under a fluorescence microscope (IX83, OLYMPUS) for imaging. After imaging, an electric field was applied within the chip using a corona processor (BD-20ACV, Electro-Technic Products) for approximately 5 s to induce droplet fusion⁴². Once fusion was complete, the corona processor was turned off, and the chip was left to incubate, allowing the platelets and drugs within the fused droplets to interact⁴². The experiment was performed in triplicate to ensure reproducibility and accuracy of the results.

Platelet aggregation by light transmission aggregometry (LTA)

The platelet aggregometer (CHRONO-LOG) was preheated to 37 °C. The washed platelets were diluted to a concentration of 3 × 10⁸ platelets/mL using Tyrode’s Buffer. Next, 300 μL of the washed platelets was added to a glass aggregating cuvette (312, CHRONO-LOG), which was magnetically stirred and inserted into the detection chamber of the platelet aggregometer. The baseline was calibrated and stabilized before the agonist was added to the cuvette. The platelet aggregation curve was recorded over a 5-min period. The experiment was conducted in triplicate to ensure the reproducibility and accuracy of the results.

Recognizing merged droplets through transfer learning and fine-tuning

In this study, we used transfer learning with a pre-trained MobileNetV2 model68 to classify droplet images⁶⁵. The training dataset included 1200 manually verified images ( ≈ 300 per class) under controlled conditions, categorized into four classes: (1) microwells with droplets but no platelets; (2) empty microwells; (3) microwells with droplets and aggregated platelets; (4) microwells with droplets and dispersed platelets. Images were randomly split into training (70%), validation (15%), and hold-out test (15%) sets.

To improve generalization, data augmentation (random flips, rotations) was applied, and pixel values were normalized to [−1, 1] (consistent with MobileNetV2 input requirements). The pre-trained MobileNetV2 model was loaded with include_top = False to remove its original classification layers; its base weights were frozen to serve as a feature extractor. We added a global average pooling layer, a dropout layer (for regularization), and a dense layer with softmax activation for final classification.

Training proceeded in two phases. Initial training: Only newly added layers were trained using the Adam optimizer (learning rate = 0.0001) for 10 epochs. Fine-tuning: The top layers of the base model (starting from layer 100) were unfrozen, and the model was fine-tuned with the RMSprop optimizer (reduced learning rate = 0.00001) for 10 additional epochs. Model performance was monitored on the validation set during training, with final accuracy evaluated on the test set.

Data acquisition and analysis

First, droplets were identified in the fluorescence images using image processing software, and the fluorescence intensities of the encoded droplets were extracted and exported. The exported data were then decoded to determine the specific drug combinations present in each microwell. Finally, another custom script was applied to analyze the bright-field images, assessing the degree of platelet dispersion and aggregation within the fused droplets. This analysis enabled the quantification of both aggregation and anti-aggregation efficiencies.

Statistical analysis

The results were presented using GraphPad Prism (Version 9.5.0) as mean ± standard deviation (SD) from at least three independent experiments, unless specified otherwise. Statistical significance was defined as p < 0.05, and p-values with 95% confidence intervals (95% CI) were calculated via GraphPad Prism’s built-in analysis functions.

Author contributions

ChenGuang Wang: formal analysis, investigation, data curation, writing—original draft, and Visualization. Wenjie Zhu: formal analysis, data curation. JiaWei Zhu: validation, formal analysis. Tian Gao: software. ZheYi Jiang: formal analysis. TianTian Zhang: investigation. Long Chen: formal analysis. JunFen Zhang: resources, funding acquisition. Yifan Liu: conceptualization, funding acquisition, writing—review & editing. Alex Chia Yu Chang: methodology, conceptualization, project administration, supervision, writing—review & editing and funding acquisition.

Data availability

All data supporting the findings of this study are available within the article and its Supplementary Information, or available from the corresponding author upon request.

Code availability

All the codes regarding the AI model are available at the following GitHub repository: https://github.com/c31io/tect/tree/biochip.

Conflict of interest

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41378-025-01126-8.