P07-07
Scaling up Binding Free Energy Calculations: Integrating Free Energy Perturbation (FEP) and Active Learning to Prioritize Compound Designs
Preferred Networks
( * E-mail: ytamura@preferred.jp )
Binding free energy calculation measures the change in free energy when a compound binds with its target protein. It’s a commonly used technique for designing drug candidates with strong affinity. P-FEP is one of the relative free energy perturbation (RBFEP) implementations [1], which calculates the binding free energy differences between compounds within a group sharing a common scaffold. The performance of P-FEP has been demonstrated on a number of benchmark sets. One of the challenges with RBFEP is its significant computational load, requiring the use of high-performance computing systems. However, even with such resources, applying FEP to thousands or tens of thousands of compounds within a finite time remains a significant issue.
As a potential solution to this problem, the pool-based active learning has been proposed [2]. This study demonstrated the effectiveness of P-FEP combined with the pool-based active learning protocol targeting proteins involved in the treatment of disease. The targets were selected based on whether the 3D structures were available, compounds had common or similar scaffold, and their activities were diverse.
Before applying the active learning, we calculated the binding free energy of the compounds whose SAR has been available. And good correlation between experimental values and calculated values were confirmed. Initially in the active learning, we prepared a compound pool using machine learning-based and chemical reaction-based methods. Then we repeated sampling and labeling based on the active learning strategy using the Gaussian process regression model. Utilizing our approach, we were able to narrow down compounds with strong binding affinity. In this presentation, we will discuss the effectiveness of our approach in prioritizing compounds for synthesis and evaluation from a large compound pool.
[1] https://tech.preferred.jp/ja/blog/pfep-launch/
[2] Thompson J. et al. Optimizing active learning for free energy calculations. Artif. Intell. Life Sci. 2022,2,100050.