Predicting Gene Expression from RPE Cell Images

This study aimed to develop RPEGENE-Net, a multi-resolution deep learning framework for the non-destructive prediction of selected gene expression markers from live-cell phase-contrast microscopy images of retinal pigment epithelium (RPE) cells. Images of RPE cells exposed to aflibercept, bevacizumab, dexamethasone, aflibercept plus dexamethasone, or no treatment were acquired at 40×, 100×, 200×, and 400× magnifications. The expression levels of six genes associated with epithelial–mesenchymal transition, fibrosis, cell adhesion, and related RPE phenotypic responses, α-SMA, ZEB1, TGF-β, CD90, β-catenin, and Snail, were evaluated as prediction targets. After preprocessing the image data and gene expression values, we trained and evaluated twelve state-of-the-art deep learning architectures, including three variants of DenseNet, five variants of ResNet, EfficientNet_b5, Inception_v3, RegNet_y_400mf, and a vision transformer model (Swin_b). A two-stage learning pipeline was implemented, in which multiple deep learning backbones were first pretrained using an autoencoder-based strategy to extract histology-relevant features in an unsupervised manner, followed by fine-tuning for supervised gene expression regression and treatment classification tasks. Features extracted from the second stage across four magnifications were concatenated to generate the final prediction, leveraging multi-scale morphological information for improved accuracy. DenseNet121 demonstrated superior performance, achieving the highest Pearson correlation coefficients for four genes: α-SMA (0.79), ZEB1(0.84), TGF-β (0.83), and Snail (0.86). ResNet34 outperformed other models for CD90 (0.87) and β-catenin (0.85) predictions. The average mean-absolute-error (MAE) and average root mean square error (RMSE) on test dataset were 0.0244 and 0.1228, respectively. The R² scores ranged from 0.50 (α-SMA) to 0.74 (TGF-β), indicating strong alignment between predicted and actual gene expression values. A multi-level approach, combining data from 40x, 100x, 200x, and 400x magnifications yielded higher R² scores for almost all genes compared to single-magnification models. For the classification task, DenseNet121 achieved F1 score, precision, recall, and accuracy of 0.98, with a specificity of 0.99. These findings demonstrate the feasibility of RPEGENE-Net to predict selected molecular markers and distinguish treatment conditions in RPE cultures. Following validation in larger and biologically independent datasets, this approach may support experimental studies and non-destructive quality assessment of cultured RPE cells.本研究旨在开发RPEGENE-Net，这是一款多分辨率深度学习框架，可从视网膜色素上皮（retinal pigment epithelium, RPE）细胞的活细胞相差显微镜图像中无创预测选定的基因表达标志物。本研究采集了经阿柏西普（aflibercept）、贝伐珠单抗（bevacizumab）、地塞米松（dexamethasone）、阿柏西普联合地塞米松处理，以及未接受任何处理的RPE细胞的图像，采集时采用了40×、100×、200×及400×四种放大倍率。本研究选取与上皮间质转化、纤维化、细胞黏附及相关RPE表型反应相关的6个基因的表达水平作为预测靶点，分别为α-SMA、ZEB1、TGF-β、CD90、β-连环蛋白（β-catenin）及Snail。在对图像数据与基因表达值完成预处理后，我们训练并评估了12种当前主流深度学习架构，包括3种DenseNet变体、5种ResNet变体、EfficientNet_b5、Inception_v3、RegNet_y_400mf，以及一款视觉Transformer模型（Swin_b）。本研究实施了两阶段学习流程：首先采用基于自编码器的策略对多个深度学习骨干网络进行预训练，以无监督方式提取与组织学相关的特征；随后针对有监督的基因表达回归与治疗分类任务开展微调。将第二阶段从4种放大倍率下提取的特征进行拼接，以生成最终预测结果，借此利用多尺度形态学信息提升预测精度。其中DenseNet121表现最优，在4个基因的预测中取得了最高的皮尔逊相关系数：α-SMA（0.79）、ZEB1（0.84）、TGF-β（0.83）及Snail（0.86）。ResNet34在CD90与β-连环蛋白的预测中性能优于其他模型，相关系数分别为0.87和0.85。测试集上各基因预测的平均绝对误差（mean-absolute-error, MAE）与均方根误差（root mean square error, RMSE）的均值分别为0.0244和0.1228。决定系数（R²）范围为0.50（α-SMA）至0.74（TGF-β），表明预测值与实际基因表达值具有良好的一致性。结合40×、100×、200×及400×放大倍率数据的多尺度方法，相较于单放大倍率模型，几乎所有基因的R²得分均有所提升。在分类任务中，DenseNet121的F1分数、精确率、召回率及准确率均为0.98，特异性为0.99。上述研究结果证实了RPEGENE-Net在预测选定分子标志物及区分RPE培养物处理条件方面的可行性。在更大规模且生物学独立的数据集上完成验证后，该方法可用于支持RPE培养细胞的实验研究及无创质量评估。