Automatic Assignment of Biacore SPR and Octet BLI Kinetic Binding Curve Profiles with Artificial Intelligence.
SLAS2024, Boston, MA, USA
February 5, 2024
Kinetic binding curve profiles contain key information that is used to demonstrate target engagement for small and large molecule drug candidates and to determine whether the drug candidates meet drug discovery project criteria. As part of the hit characterization and assessment of molecules for Amgen’s therapeutic projects, thousands of affinity values and binding profiles have been collected for protein/protein and small molecule/protein interactions using the SPR and Octet instruments. Unfortunately, assessing such binding profiles requires expert time and knowledge, limiting the number of maximum processable analytes. To overcome this limit, Amgen and Genedata have collaborated to generate an AI-based classifier that automatically assigns binding profiles to Biacore SPR and Octet BLI kinetic binding curves. The classifier considers the shape of the kinetic binding curve and automatically assigns its binding profile to one of four classifications: “no binding”, “kinetic”, “steady state” or “unclear”. The classifier first analyzes the kinetic binding curves to determine if there is no binding. If binding is detected, then it determines if it can be classified as either a kinetic or steady state fit. If the binding does not clearly fall into either one of these classifications, it is then described as “unclear” which is left for the user to define. The classifier was trained and validated on a ground truth dataset comprised of real-world kinetic binding data from SPR multi-cycle kinetic samples, SPR single-cycle kinetics samples, and BLI multi-cycle kinetic samples, where each binding curve profile in the ground truth dataset was assigned by Amgen expert users. The classifier has now been tested on datasets that were not part of the original ground truth. Out of 500 new binding curve profiles, the classifier was 90% accurate across all four categories distributed between ~80% clear (no binding, kinetic, or steady-state) and ~20% unclear. In the 80% of the dataset that was categorized as either no binding, kinetic, or steady-state, the classifier achieved 95% accuracy--mimicking the performance of Amgen expert classification as determined by comparing three domain expert votes on the same data. Overall, this workflow confers a twofold benefit: (1) to the end user, by automating the assignment of binding profiles, significantly reducing the manual effort on the part of the scientists and allowing them to focus on the ~20% unclear cases, and (2) to the organization, standardize binding curve profiling process across many scientists, thereby helping to ensure FAIR data and consistent use of result parameters.