Deep Learning-based Prediction Method For Label-free Stimulated Raman Scattering(SRS) Microscopy Images

Proteins are important molecules that support life activities and are also components of important structures of organisms.In biological cells,proteins are not only involved in many life activities,but also undertake many important functions.Therefore,analyzing the localization of protein subcells at the cellular level is important for human biology and the treatment of disease.In recent years,with the breakthrough and rapid development of microscopic imaging technology,protein optical microscopic images that can visually reflect the protein distribution in subcells have become the core data carrier in the study of protein subcellular localization.Therefore,predicting protein subcellular localization based on optical microscopic imaging methods has become one of the most important methods at present.Although deep learning methods have demonstrated excellent performance in image tasks,they are still in their infancy in the field of subcellular localization of proteins,and have not yet been well applied.In this paper,a combination of optical microscopy and deep learning was developed to locate protein subcells.Fluorescence microscopy is a technique for studying labeled single cells of interest and protein subcells.In the current field,the use of fluorescent markers to locate protein subcells is one of the most efficient and commonly used methods.Fluorescence microscopy is highly specific,enabling selective detection of fluorescently labeled targets at very low concentrations in complex mixtures.At the same time,due to its high sensitivity and spatial resolution,fluorescence microscopy can accurately locate individual molecules over the length range with diffraction-limited optical resolution.However,the bottlenecks of conventional fluorescence imaging techniques in live-cell detection mainly include:(1)a substance that uses fluorescent dye molecules to locate protein subcells with limited markers;(2)The development of fluorescent dye molecules is time-consuming,labor-intensive and expensive;(3)Some fluorescent molecules are toxic and harmful to the labeled biological cells,and may even cause damage to the biological structure or function;(4)There are also substances that cannot be calibrated.Stimulated Raman scattering microscopy(SRS)is a recently developed and powerful tool for studying the structure,function,and activity of living cells.Compared with conventional fluorescence microscopy imaging technology,it has the advantages of hyperspectral resolution,no fluorescent molecular labeling and fast and reliable.However,How to localize and predict protein molecule subcells from complex femtosecond dynamic SRS cell images is a key question in current research,which can not only provide useful clues for their function and biological processes,but also help to prioritize drug development and select appropriate targets.However,the bottleneck in the prediction and localization of protein suborganelles for SRS cell imaging is the need to model complex relationships hidden under the original cell imaging data due to spectral overlap information from different protein molecules.In this research work,aiming at the bottleneck problem in live-cell detection of high-resolution fluorescence imaging and label-free stimulated Raman imaging technology,we combine the advantages of the two imaging methods to develop a new method for labeling stimulated Raman molecular cell imaging analysis based on deep learning,and use the hybrid architecture of CNN and Transformer deep learning to realize the rapid prediction from label-free multi-target femtosecond stimulated Raman cell microscopic image to a single target fluorescent cell microscopic image.and complete the accurate localization of multiple protein suborganelles.Through experiments,as well as comparative analysis with other models,it is proved that the multimodal hybrid learning model developed in this institute has the best performance,which is not only more accurate in predicting fluorescence microscopy images by femtosecond stimulated Raman microscopic images,but also more accurate in the localization of protein suborganelles.This method may be applied to the diagnosis of diseases and drug development related to protein subcellular localization in the future.