Max Planck Institute enhances imaging of single proteins

15 Mar 2023

Combining interferometric scattering with AI could detect small traces of disease markers.

Imaging of proteins in as much detail as possible is a key step in understanding neurodegenerative diseases, inflammation and immune responses.

A project from the Max Planck Institute for the Science of Light (MPL) and the University of Erlangen–Nuremberg (FAU) has now demonstrated how combining machine learning with an existing protein imaging modality can enhance the sensitivity limit in label-free visualization of single proteins.

Reported in Nature Methods, the findings open the door to optical investigations of small traces of biomarkers and disease markers such as cytokines and chemokines, according to the project team.

The optical technique at the heart of the new approach is interferometric scattering (iSCAT) microscopy, an interferometry method that detects elastic Rayleigh scattering in addition to reflected or transmission signals from objects being studied. iSCAT can show not only the presence of proteins and sub-wavelength structures, but also their size and mass.

In theory iSCAT imaging resolution is limited by shot noise, but in practice a number of noise sources combined with speckle-like background fluctuations are also present, and proteins with molecular weights below approximately 40 kilodaltons (kDa) have not been detected.

MPL's solution is to apply machine learning to the task. Machine learning methods have recently been used in microscopy applications with an emphasis on correcting the background or enhancing the signal, but the distinction between supervised deep neural networks (DNNs) using labeled input data, and unsupervised DNNs using unlabelled datasets, has been key.

Seeing biomarkers and nanostructures

In trials, an iSCAT platform employing 445 nanometer illumination was employed, with its results benchmarked using fluorescence detection. Under normal operation, the sensitivity of iSCAT microscopy when imaging proteins on video images from this platform would be limited to molecular weights of approximately 40 to 50 kDa, according to the team.

Two deep learning approaches were then applied to the video data. The first algorithm, FastDVDnet, is designed specifically to remove noise from data, and was used to identify iSCAT images of proteins from the recorded video sequences. The extracted spatiotemporal features were then analyzed by iForest, an unsupervised algorithm used to spot "anomalies" in data - the anomalies in this case being the protein signals of interest among the normal residual background speckle.

The combination of iScat and two machine learning algorithms was able to push the sensitivity limit for protein imaging down to 9 kDa, potentially bringing a number of significant disease biomarkers and biological nanostructures into view. Further improvements to the CMOS sensor and imaging substrates employed could help to improve the results.

"We are determined to push the detection limit further both by improving the physical measurement methods and by developing more sophisticated machine learning algorithms," said MPL's Vahid Sandoghdar. "There is really no fundamental reason why we should not be able to detect molecules below 1 kDa, coming close to the weight of even a single lipid molecule."