Using AI to understand the nano world

Modern technology enables science to scrutinize the world on the nanometer scale. Thanks to today’s high-resolution techniques, researchers are even able to observe and measure individual molecules, such as proteins or DNA structures, in detail. This information from the nano world is highly coveted in many scientific disciplines. “Technologies like single-molecule spectroscopy have revolutionized the way in which we investigate the mechanisms of processes on the nanoscale,” says Don C. Lamb, Professor for Biophysical Chemistry in the Faculty of Chemistry and Pharmacy at the LMU. In particular, current fluorescence spectroscopy and microscopy techniques enables researchers to investigate individual, dynamic biomolecules in cells, membranes, surfaces and solutions.

To discover more about the structure, dynamics and behavior of such complex biomolecules, scientists tag them with fluorescent dyes at specific sites, enabling them to measure distances and track changes over time. Depending on the complexity of the research question being addressed, one to several different fluorophores may be required. As the experiments increase in intricacy, the evaluation and interpretation of the data obtained in this manner become extremely laborious, time-consuming and susceptible to misinterpretation.

Artificial data analyst: Deep-LASI

In the journal Nature Communications, Lamb’s research group has now presented a software package which, through the use of machine learning, will automate, accelerate, and increase the reliability of the evaluation of such datasets. Deep-LASI (Deep-Learning Assisted Single-molecule Imaging analysis) is a neural-network-based software for the analysis of single-, two- and three-color single-molecule time trajectories. It has the ability to automatically and quickly extract the relevant information in the data while filtering out background, noise, photobleaching and other artifacts. “This is a real game-changer,” says Lamb. “Out of the raw data, Deep-LASI generates precise information about the quality of the measurement, the various states of the molecules and their dynamics – without prior knowledge or assumptions about the system being investigated.”

To demonstrate this, the researchers tested Deep-LASI on multiple use cases. This involved pitting the software against humans and against conventional analytical programs that were trained on the same data but do not incorporate deep neural networks (so-called hidden Markov model analyses). In this way, the scientists were able to demonstrate that their algorithm works with exceptional reliability and precision and outperforms both the human and automated competition.

Unbiased and capable of learning

One of the major advantages of neural networks is their ability to register and analyze complicated temporal dependencies and complex patterns in the data in a fraction of the time it would take humans to do so. This permits increased accuracy and the identification of subtle transitions or states that either cannot be recognized with conventional analytical methods or only with difficulty. Furthermore, deep learning models can learn from large data volumes and are unbiased compared to human observers, which reduces systematic errors. In their study, however, the authors also caution that it can be a challenge to reconstruct how the neural networks make decisions. “Unknown prejudices can arise that are inherent to the network itself,” explains Simon Wanninger, the doctoral researcher who programmed Deep-LASI. Although neural networks are extremely accurate, it can be difficult to understand which underlying features and mechanisms influence their predictions. As such, we cannot entrust the machines with the totally unsupervised interpretation of the data.

Future of single-molecule investigation

To test their new method and illustrate the possibilities of the technique, Lamb’s research team used an L-shaped DNA origami structure developed by Lamb’s colleague, Professor Philip Tinnefeld. The kinetic properties of this structure are highly tunable and well known. Previously published single-molecule data on heat-shock proteins served as an additional reference. Based on these systems, the team successfully demonstrated the reliability of Deep-LASI for analyzing the dynamics and structure of biomolecules.

The authors of the study are convinced that AI-based software like Deep-LASI will be essential tools in the future of nano research. According to Don C. Lamb: “We fully expect that deep-learning approaches employed in single-molecule investigations will greatly facilitate and accelerate analytics and become indispensable in future.”

Simon Wanninger, Pooyeh Asadiatouei, Johann Bohlen, Clemens-Bässem Salem, Philip Tinnefeld, Evelyn Ploetz & Don C. Lamb: Deep-LASI: deep-learning assisted, single-molecule imaging analysis of multi-color DNA origami structures. Nature Communications, 2023.