高效准确的药物体外筛选是避免具有风险的小分子化合物进入临床试验阶段、降低药物研发成本的关键。高内涵表型筛选能够通过不同荧光标记的细胞影像反映候选药物对细胞表型影响，是量化评估候选药物的有效性和安全性的重要方法。然而目前的高内涵表型筛选过程仍然依赖于荧光染料对特定细胞结构的标记，这个过程不仅对研究人员的操作技术要求较高，并且染色过程对细胞会造成一定的损伤，从而对检测结果造成影响，此外，由于光谱的局限性，可同时使用的荧光标记的数量会受到限制。AI技术的出现能够为高内涵筛选过程引入新的解决方案。本文提出一种基于荧光标记预测的药物评价方法，并命名为DeepHCA，旨在解决当前高内涵筛选流程中的一些关键问题，改善当前研究中的局限性，为实现基于明场细胞影像的高通量药物筛选提供方案。本文主要工作如下：（1）构建了一套包含2400组同一视野下细胞的明场及其细胞膜、细胞核、脂质荧光染色图像的高质量、标准化的细胞明场-荧光图像配对数据集，扩充了现有细胞图像公开数据集的标记范围。通过对比同一视野下明场及荧光图像的纹理学特征发现了细胞明场图像的数据丰富性及基于明场图像预测细胞状态的可行性。（2）提出了一种基于去噪扩散模型的细胞特定结构的荧光标记预测方法。在细胞核荧光标记预测中，本方法相比于与基于生成对抗网络的图像翻译方法，平均PSNR、SSIM、RMSE、R2相关性分别提升2.50 dB、0.14、1.61、0.23；对HepaRG、HepG2、Huh7三种细胞系，平均SSIM分别为0.77、0.70、0.71，并且扩展了细胞膜和脂质的精细结构预测。此外，开发了一种基于Transformer架构的条件去噪扩散模型，并且在细胞膜荧光标记预测中取得较好的效果。（3）设定四种具体的药物评价场景，通过建立的荧光标记预测模型分别预测同一视野下细胞膜、细胞核及脂质的的荧光标记图像，基于预测的荧光标记图像分别进行数量、面积、圆度等参数的计算，用于量化分析药物作用下细胞的凋亡、增殖、形态学特征及细胞内容物等表型的变化。分析结果显示基于荧光标记预测的细胞表型评价方法能够降低95%以上的实验操作时间，提供更精准、全面、无创的细胞表型评价，简化了体外筛选的流程，对药物筛选和研发具有重要意义。
Efficient and accurate in vitro drug screening is crucial for avoiding the clinical trial phase of molecules with potential risks and for reducing the cost of drug development. High-content phenotypic screening, which reflects the impact of candidate drugs on cell phenotypes through cellular images labeled with different fluorescent markers, serves as a significant method for quantifying the effectiveness and safety of candidate drugs. However, the current high-content phenotypic screening process still relies on the staining of specific cellular structures with fluorescent dyes. This process not only demands high technical skills from researchers but also causes certain damage to cells, affecting the results. Additionally, due to spectral limitations, the number of fluorescent markers that can be used simultaneously is restricted. The emergence of AI technology can introduce new solutions to the high-content screening process. This article proposes a drug evaluation method based on the prediction of fluorescent markers, named DeepHCA， aiming to address some of the key issues in the current high-content screening workflow, improve the limitations present in current research, and provide solutions for high-throughput drug screening based on bright-field cell imaging. The main contributions of this work are as follows:(1) A high-quality, standardized dataset of bright-field and fluorescent images of cellular membranes, nuclei, and lipids from 2400 groups of cells under the same field of view was constructed, expanding the labeling scope of existing public cell image datasets. The comparison of texture features between bright-field and fluorescent images under the same field of view revealed the richness of data in bright-field cell images and the feasibility of predicting cell states based on bright-field images.(2) A method based on a denoising diffusion model for the prediction of cell-specific structural fluorescence labeling has been proposed. In the prediction of nuclear fluorescence labeling, compared to image translation methods based on Generative Adversarial Networks, this method achieves an average improvement of 2.50 dB in PSNR, 0.14 in SSIM, 1.61 in RMSE, and 0.23 in R2 correlation. For the cell lines HepaRG, HepG2, and Huh7, the average SSIM values are 0.77, 0.70, and 0.71, respectively, and the method has been extended to predict the fine structure of cell membranes and lipids. Moreover, a conditional denoising diffusion model based on the Transformer architecture has been developed and has shown promising results in the prediction of membrane fluorescence labeling.(3) Four specific drug evaluation scenarios were established, and the fluorescent labeling images of cellular membranes, nuclei, and lipids under the same field of view were predicted using the developed model. Quantitative analysis of drug-induced changes in cell apoptosis, proliferation, morphological characteristics, and intracellular content was performed based on the predicted fluorescent labeling images, utilizing parameters such as quantity, area, and circularity.The analysis results demonstrated that the cell phenotype evaluation method based on fluorescent marker prediction could reduce experimental operation time by more than 95%, providing more precise, comprehensive, and non-invasive cell phenotype evaluations, thus simplifying the in vitro screening process and bearing significant implications for drug screening and development.
