NEW YORK – Proteomics firm Biognosys has advanced its limited proteolysis (LiP) chemoproteomics workflow to enable better analysis of protein-small molecule interactions in human samples.
In a paper published in August in Nature Communications, researchers from the Schlieren, Switzerland-based company and ETH Zurich, where the technique was first developed, combined LiP measurements with machine learning to improve the approach's ability to identify small molecule drug targets as well as to provide structural information on those targets.
The improvements make the technique better suited for work in complex organisms like humans and, consequently, for drug development research, said Oliver Rinner, Biognosys' CEO and an author on the study.
Originally developed in the lab of ETH Zurich Professor Paola Picotti, who is a member of Biognosys' scientific advisory board, the LiP approach analyzes changes in the peptides produced through proteolytic digestion to identify protein binders of small molecules of interest and generate structural information about those proteins and their binding sites.
The method uses digestion with a broadly specific protease followed by a standard mass spec-based proteomics workflow. In the initial digestion step, the protease will cleave proteins only at sites that are accessible, left exposed by whatever structural conformation it happens to be in at the time of analysis. When a sample of interest is treated with, for instance, a drug, the proteins that bind to this molecule will undergo a structural change, and this will be reflected in changes where the protease is able to cleave the protein.
By following this initial digestion step with standard trypsin digestion and mass spec analysis, researchers can compare the peptides generated in treated and untreated samples and, based on changes in the peptides produced, determine which proteins had their structures altered by the treatment in question.
Biognosys licensed the LiP technology from ETH Zurich in 2017. However, Rinner said that the initial version of the technique was best suited to analyses in relatively simple organisms like microbes. For instance, one of the Picotti lab's major publications using the technique was a 2018 paper in Cell presenting a map of protein-metabolite interactions in Escherichia coli.
Most of Biognosys' potential customers for the technique come from pharma, Rinner said, adding that for these companies "only mammalian systems are relevant." As such, in collaboration with Picotti, the company had to “bring the technology to a new level," he said.
To do this, the company and its ETH Zurich collaborators developed a machine learning approach that uses a number of features including drug dose-response curves; how frequently proteins are identified as targets in experiments where they are not expected targets; and the correlation of dose-response data from different peptides from the same protein, to score proteins according to how likely they are to be true drug targets.
In the Nature Communications paper the researchers provided an example of the improvement this approach provided, comparing analyses of the drug rapamycin, which is known to interact only with the protein FKBP1A.
Using the original workflow to analyze rapamycin-protein binding in HeLa cell digests and live HeLa cells, the researchers identified 52 potential protein targets in digest and 37 in the live cells, indicating that the method did not provide the specificity needed to distinguish between true positive hits and likely false positives.
Performing the experiment using the machine learning approach, the researchers found that the top five scoring peptides all came from FKBP1A, making it the top-scoring candidate and demonstrating the improved utility of the machine learning-based technique. They followed this with an analysis of the compound FK506, which also targets FKBP1A, and again found that the five top-scoring peptides came from FKBP1A.
Rinner said that in addition to the target binding information, the approach also appeared useful for providing structural information on the protein targets based on what peptides are impacted by a molecule's binding.
He added that the structural analysis provided by the approach could potentially replace or supplement nuclear magnetic resonance (NMR) or hydrogen-deuterium exchange (HDX) analyses often used in drug development to collect low-resolution structural information on potential targets.
"We really feel that this is moving into a structural proteomics technology aside from just the element of identifying compound binding," Rinner said, adding that the company has received significant interest from its pharma customers in using the approach to generate this sort of structural information.
At the American Association for Cancer Research virtual meeting in April, Biognosys presented a poster detailing work done with Cambridge, Massachusetts-based pharma firm Cedilla Therapeutics on using LiP to analyze protein binding sites and binding-induced conformational changes to a pair of drug compounds, including the BRD4 inhibitor JQ1. The researchers found that the approach could determine small molecule-protein binding at a resolution down to five to 10 amino acids and that the structural data they generated matched well to data generated using HDX, NMR, and X-ray.
Looking ahead, Rinner said that continued advances on the mass spectrometry side of things should also improve the technique's performance.
"The limiting factor is absolutely the [mass spec] sensitivity," he said, noting that finding and measuring the altered peptides remains one of the major challenges, particularly given the fact that the peptides whose digestion patterns are changed by a compound's binding may not be the most easily detectable peptides by mass spec.
"What we have seen is that the deeper [the mass spec goes], the better the results you get," he said. "So the main thing we can do from the mass spec side is continue to get more sensitive so we can pick up more primary or secondary LiP peptides and more peptides from the same protein, which will increase the confidence [in the results] further. It's directly connected to the increase in sensitivity provided by new mass specs."