Enhancing single-cell biology through advanced AI-powered microfluidics

Abstract

Microfluidic technology has largely benefited both fundamental biological research and translational clinical diagnosis with its advantages in high-throughput, single-cell resolution, high integrity, and wide-accessibility. Despite the merits we obtained from microfluidics in the last two decades, the current requirement of intelligence in biomedicine urges the microfluidic technology to process biological big data more efficiently and intelligently. Thus, the current readout technology based on the direct detection of the signals in either optics or electrics was not able to meet the requirement. The implementation of artificial intelligence (AI) in microfluidic technology matches up with the large-scale data usually obtained in the high-throughput assays of microfluidics. At the same time, AI is able to process the multimodal datasets obtained from versatile microfluidic devices, including images, videos, electric signals, and sequences. Moreover, AI provides the microfluidic technology with the capability to understand and decipher the obtained datasets rather than simply obtaining, which eventually facilitates fundamental and translational research in many areas, including cell type discovery, cell signaling, single-cell genetics, and diagnosis. In this Perspective, we will highlight the recent advances in employing AI for single-cell biology and present an outlook on the future direction with more advanced AI algorithms.

INTRODUCTION

Microfluidic systems have found application in diverse areas, such as identifying individual cancer cells, conducting liquid biopsies, evaluating drugs, simulating angiogenesis, and detecting metastasis.^1,2 The advantages of microfluidic systems, such as high-throughput capabilities, single-cell resolution, data integrity, and broad accessibility, collectively facilitate the generation of vast and comprehensive datasets.³ The massive and complex data generated by microfluidic devices in single-cell biology, which enable multi-omics investigation at the level of a single cell, is the perfect playground for data-hungry artificial intelligence (AI) models, providing a solid foundation for developing machine learning (ML) and, in particular, deep learning (DL) algorithms.⁴ The diversity of single-cell biology tasks drives the data acquisition and training process of artificial intelligence models to adapt to data patterns and environments.^2,5,6 As compared to traditional methods of data analysis, AI models require less manual intervention in classification and prediction.⁷ By combining the merits of both the microfluidic systems and AI models, the automatic and intelligent processing of big data can be realized. Thus, the two approaches are mutually beneficial and combined to create new ideas and research areas.

In this Perspective, our focus is on elucidating the recent advancements in AI model development tailored for single-cell biology. We will delve into their applications across a range of areas and present an outlook on the future direction, highlighting the possibilities of employing even more sophisticated AI algorithms in microfluidics.

AI MODELS IN MICROFLUIDIC SINGLE-CELL BIOLOGY

AI models in single-cell biology aim to extract cellular features from different modalities, such as fluorescence microscopy images,⁸ holographic flow cytometry images,⁹ electrical signals,¹⁰ RNA-seq,¹¹ or a combination of these.¹² According to whether the data are labeled, the model's training method is divided into supervised and unsupervised.

AI MODELS: FROM ML TO DL

Machine learning (ML) algorithms such as support vector machine (SVM), logistic regression (LR), or random forest are widely used in single-cell biology research.¹³ These models focus on classification and regression tasks based on selected features, and the feature manufacturing processes are of vital importance to the performance of the models.

Manak et al.¹⁴ developed a live-cell phenotypic biomarker assay for the risk stratification of cancer patients at single-cell resolution. An extracellular matrix (ECM) coated microfluidic device was employed in live-cell imaging assays. Phenotypic biomarkers were analyzed by a random forest classifier with optimal ranking. The risk stratification prediction score for the area under the curve value in the receiver operating characteristic curve exceeded 80%, validating the analysis and its potential clinical applicability.

Ferguson et al.¹⁵ designed a wideband electrical sensor combined with a data analysis model for the detection of size-changed nuclei in single-cell microfluidic devices. S-parameters and cell images of single cells were collected to infer important frequencies related to the nucleus size. SVM and LR models were used to improve classification models in binary and multiclass scenarios. With the aid of an ML model in cell classification, the results achieved 94% accuracy in a binary classification task in distinguishing ciprofloxacin treated and untreated cells and 96% in multiclass prediction. The results could lead to more flexible approaches to cancer diagnosis.

With the development of high-throughput microfluidic devices and complex biological data beyond the capabilities of ML models, the end-to-end workflow of deep learning (DL) models is more intuitive for data processing and discovering hidden connections among data. DL models are starting to show advantages in single-cell biology in microfluidic devices (Fig. 1).

FIG. 1.

AI models utilized in microfluidics of single-cell biology.

Lamanna et al.¹⁶ introduced the Digital Microfluidic Isolation of Single Cells for -Omics (DISCO) platform, which enables users to select specific cells of interest from a limited initial sample size and links their single-cell sequencing data to their immunofluorescence-based characteristics. DISCO combines digital microfluidics, laser cell lysis, and a Convolutional Neural Network (CNN)-based DenseNet for single-cell segmentation. DISCO's distinctive levels of selectivity, context, and accountability suggest its potential for the in-depth analysis of rare cell populations with contextual dependencies.

Other models, such as variational autoencoder¹⁷ and vision transformer,¹⁸ are also used as a single-cell segmentation model with microscopy image data collected from microfluidic devices. Nevertheless, there are no definite rules for the selection of deep learning models. Most studies lack justification for model choosing.¹⁹ A comparative study of deep learning algorithms in single-cell biology is still needed.

SUPERVISED LEARNING VS UNSUPERVISED LEARNING

Supervised learning aims at maximizing the performance of annotated datasets by implementing models for classification or regression tasks with labeled data prepared in advance.²⁰ Sesen and Whyte²¹ described a single-cell sorting system using supervised ML algorithms with microfluidic devices. Single red blood cells in droplets were differentiated using quadratic support vector machines. A supervised learning algorithm trains a classifier with the unique size and circular features of expert-labeled data. The system provides an efficient method for the analysis of small cell populations.

Supervised learning performance is prone to labeled bias.²² On the other hand, unsupervised learning can significantly mitigate the impact of biases introduced in manual annotations and save effort in obtaining data annotations. Unsupervised or semi-supervised learning results in pre-trained models that can be transferred to other application scenarios to simplify the training procedure.²³ Fleming et al.²⁴ designed a DL model for background removal in scRNA-seq with semi-supervised training. DESC,¹¹ an unsupervised DL algorithm for scRNA-seq clustering, utilizes iterative self-learning to optimize the clustering objective function, which effectively mitigates complex batch effects, preserves biological variation, and uncovers both the discrete and pseudo-temporal cellular structures. However, the unsupervised learning process usually requires a pretext task and carefully designed loss function for training, which leads to unintuitive results and even hurts performance.

FACILITATING SINGLE-CELL BIOLOGY WITH AI-EMPOWERED MICROFLUIDICS

The application of artificial intelligence models in cell analysis, including cell counting, sorting, and classification, is the most common implementation and has a long history of adoption in basic research and clinical diagnosis. In the single-cell biology of microfluidic devices, besides data processing and feature extraction, the interpretation of the underlying structure of big data is also a promising direction, since the models can capture connections in latent space.^11,25,26 The artificial intelligence models in these scenarios focus on data augmentation and feature extraction, and use phenotypic data for prediction and pattern recognition, facilitating the development of a large number of applications (Fig. 2). In various domains, we have selected recent research studies that have been assisted by artificial intelligence methods. Aiming to effectively illustrate the contributions of AI technology to single-cell biology applications, these collaborations serve as a catalyst for inspiring further applications of artificial intelligence in this field.

FIG. 2.

Applications of the AI model in different microfluidic systems. (a) Droplet-based microfluidic system. Reproduced with permission from Sarkar et al., Lab Chip 20(13), 2317–2327 (2020). Copyright 2023 Royal Society of Chemistry.²⁷ (b) Label-free detection of drug-induced morphological variations in microscopy images. Reproduced with permission from Kobayashi et al., Sci. Rep. 7(1), 12454 (2017), licensed under a Creative Commons Attribution (CC BY) license.²⁸ (c) Deep learning-enhanced imaging flow cytometry. Reproduced with permission from Huang et al., Lab Chip 22(5), 876–889 (2022). Copyright 2022 Royal Society of Chemistry.²⁹

Droplet-based microfluidic systems

Droplet-based microfluidic systems offer a high multiplexing capacity, enabling the reactions in single-cell resolution. However, processing the resulting high-content information, which involves cell size, roundness, structures, and other morphological features along with biomarker-based cellular phenotypes, requires robust data processing capabilities. The introduction of deep learning methods has played a crucial role in various aspects of droplet-based microfluidic research studies, including enhancing the accuracy of droplet segmentation, extracting and segmenting single-cell images, and annotating data. Nonetheless, the significance of different parameters varies depending on the experimental scenario, demanding thorough exploration to achieve precise, automated, and high-throughput analysis and experimentation.

Sarkar et al.²⁷ conducted a microfluidic droplet-based investigation, utilizing DL algorithms to assess the cytotoxic impact of natural killer (NK) cells on human cancer cell lines. The researchers employed a CNN model to quantitatively compare the interactions of immunotherapeutic NK-92 cells with different types of target cells. The algorithm yielded overall accuracy of >95% and >93% for death and interaction prediction, respectively, which has been proven to be a highly effective tool, enabling the measurement of NK cell dynamics, functional efficacy, and heterogeneity at a single-cell resolution. The study's outcomes showcased that this semi-automated single-cell assay could uncover NK cell variability and functional potency, rendering it valuable for optimizing immunotherapeutic efficacy in preclinical analyses.

Microscopy images from microfluidic devices

Kobayashi et al.²⁸ investigated cellular drug responses through the analysis of high-throughput brightfield images obtained from microfluidic systems. They focused on a support vector machine (SVM) classifier and trained the model to distinguish images between drug-treated and untreated Michigan Cancer Foundation-7 (MCF-7) cells. The SVM model demonstrated an impressive predictive accuracy of 92%. Notably, this study highlights the potential of machine learning in discerning subtle, dose-dependent morphological alterations induced by drugs.

Flow cytometry

Flow cytometry plays a fundamental role in high-throughput single-cell sorting and finds extensive applications in clinical diagnostics and basic biomedical research. By combining machine learning-driven cell sorting in both brightfield and fluorescence flow cytometry, it is possible to integrate microfluidic flow cells with piezoelectric or pneumatic actuators. This integration allows for the real-time hydrodynamic collection of cells or droplets of interest during the sorting process.^21,30,31

Isozaki et al³² introduced intelligent image-activated cell sorting (iIACS), an advanced image-activated cell sorter, that combined lightweight CNN, high-speed electronics, and graphics processing units (GPUs) to achieve real-time image classification and precise timing control during cell sorting. Cells from a sample tube are injected into the microfluidic chip, where they are focused into a single stream by the hydrodynamic cell focuser. The virtual cryo-fluorescence imaging (VIFFI) microscope captures images, which are then analyzed in real-time by an image processor. Based on the processor's decisions, cells are sorted using a dual-membrane push–pull cell sorter. Their findings highlight the potential of this semi-automated single-cell assay in uncovering the variability and functional potency of NK cells, proving its usefulness in optimizing immunotherapeutic efficacy during preclinical analyses.

Imaging flow cytometry (IFC) has emerged as a potent biomedical tool for various applications, enabling high-throughput imaging of individual cells. However, striking a balance between throughput, sensitivity, and spatial resolution remains a challenge. To address this issue, Huang et al.²⁹ proposed a deep learning-enhanced imaging flow cytometry (dIFC) approach. They implemented an image restoration algorithm on a virtual cryo-fluorescence imaging (VIFFI) flow cytometry platform, which achieved enhanced performance without compromising throughput, sensitivity, or spatial resolution. The core element of dIFC involves a high-resolution VGG16-based CNN architecture, pre-trained on the ImageNet database, acting as an image generator using low-resolution images obtained through low-magnification lenses. During the training phase, two generative adversarial networks (GANs) were designed for high-resolution image generation. The experimental results demonstrated that dIFC images exhibited improved accuracy in FISH spot counts and neck width measurements of budding yeast cells when compared to images acquired with high-magnification lenses.

Single-cell profiling

Microfluidic device-based approaches offer the advantage of single-cell resolution, enabling the acquisition of information on single-cell heterogeneity. Nonetheless, current platforms often encounter limited throughput due to the design or the absence of high-speed data analysis modules, resulting in challenges associated with obtaining statistically unbiased biological data. The utilization of AI models can enhance the data processing capabilities of microfluidic systems in single-cell profiling. At present, the use of AI models has shown advantages in single-cell profiling applications (Fig. 3).

FIG. 3.

Utilizations of the AI model in single-cell profiling in the microfluid system. (a) A high-throughput microwell array chip for the evaluation of the migration and proliferation of tumor cells. Reproduced with permission from Huang et al., Cell Rep. Phys. Sci. 4(2), 101276 (2023), licensed under a Creative Commons Attribution (CC BY) license.³³ (b) A high-throughput living single-cell multi-indicator secreted biomarker profiling platform. Reproduced with permission from Wang et al., Adv. Healthcare Mater. 11(13), 2102800 (2022). Copyright 2022 AIP Publishing LLC.³⁴

Huang et al.³³ propose a high-throughput system featuring an addressable dual-nested microwell array chip (DNMA chip) combined with a Mask R-CNN model trained for image analysis. The DNMA chips allow single-cell capture, label-free encoding, and long-term incubation, facilitating non-destructive evaluation of the migration and proliferation of individual tumor cells under normal culture conditions or chemotherapy. With AI-assisted data processing, this integrated system enables quantitative analysis of cell behavior in a high-throughput manner.

Wang et al³⁴ proposed a high-throughput living single-cell multi-indicator secreted biomarker profiling platform that integrates machine learning technology to achieve accurate tumor cell classification. This platform utilizes a single-cell culture microfluidic chip with self-assembled graphene oxide quantum dots (GOQDs) to enable high-activity single-cell culture, ensuring the normal secretion of biomarkers and high-throughput single-cell separation, provided sufficient statistical data for machine learning. By combining the K-means strategy with machine learning, they analyzed thousands of single tumor cell secretion data, achieving a tumor cell classification recognition accuracy of 95.0%. Furthermore, additional analysis of the grouping results revealed the unique secretion characteristics of subgroups. This study provides an intelligent platform for high-throughput living single-cell multi-index secreted biomarker profiling, which holds broad implications for cancer research and biomedical investigation.

OUTLOOKS

AI models have made great progression in feature extraction and data processing, but the main limitations of AI models in terms of data analysis still exist. The lack of transparency in ML and DL models is a major challenge in biological applications.³⁵ These models primarily aim to predict outputs solely based on inputs, overlooking the logic and underlying mechanisms responsible for specific outcomes. In the context of single-cell biology, these methods often fail to establish plausible connections between predictions and the actual underlying mechanisms. Consequently, enhancing the interpretability of AI model outcomes emerges as a promising avenue for realizing artificial intelligence models in single-cell biology, particularly when integrated with microfluidic devices.

The lack of interpretability presents a significant challenge in the majority of recent studies that investigate complex machine learning models. Most of these studies have primarily concentrated on estimating and discovering posterior explanations for training algorithms concerning particular predictions. In contrast, self-explanatory models, where interpretability is an integral part of the learning process, have not received as much attention. A possible solution to augment AI model interpretation is the cross-modal representation learning for multi-omics data interaction. Cross-validation of different modalities increases not only the power but also the interpretability of the model especially with large language models (LLMs).³⁶ LLMs have demonstrated impressive capabilities in reasoning and instruction prompting. An LLM could be helpful in converting abstract model output (especially in latent spaces) into language representation that helps explain the outcome. Conversely, language can also be used as input data to generate corresponding biological data as the clinical reference. Overall, the number of available AI tools is rapidly increasing; with the proper implementation in the microfluidic systems, we can anticipate a new era of intellectual biomedicines in the recent upcoming future.

AUTHOR DECLARATIONS

Conflict of Interest

The authors have no conflicts to disclose.

Author Contributions

Zhaolong Gao: Conceptualization (equal); Formal analysis (equal); Resources (equal); Writing – original draft (equal); Writing – review & editing (equal). Yiwei Li: Conceptualization (equal); Funding acquisition (lead); Supervision (lead); Writing – original draft (equal); Writing – review & editing (equal).

DATA AVAILABILITY

Data sharing is not applicable to this article as no new data were created or analyzed in this study.