Active Learning FEP+ Best Practices

FEP+ has demonstrated a high level of accuracy in predicting the binding potency of ligands, providing the driving force for lead optimization. Due to the relatively large computing cost of FEP+ and the ever-expanding chemical space chemists want to explore in lead optimization, brute force application of FEP+ on all design ideas may sometimes not be practical and more efficient method is needed to rapidly explore the large chemical space for the identification of novel potent molecules. Combination of FEP+ and machine learning through active learning FEP+ proves to be an efficient method enabling the rapid exploration of large compound libraries and faster identification of potent molecules. The purpose of this document is to describe how active learning FEP+ is deployed in internal Schrodinger drug discovery projects.

Before applying active learning FEP+ in a drug discovery project, the standard target analysis and retrospective validation of FEP+ should be conducted, and active learning FEP+ should only be applied when the FEP+ model is validated. See FEP+ Best Practices for more information.

Active learning FEP+ consists of the following steps:

Enumeration of the compound library—The list of compounds, usually in the order of 10 K to 100 K, can come from designs from medicinal chemists, (reaction-based) enumeration of R-group libraries (e.g., Pathfinder, AutoDesigner), AI generated molecules (e.g., REINVENT), or a combination of these.
FEP+ amenability assessment—The compounds to be scored by active learning FEP+ should be similar enough to the reference compounds that a single edge FEP+ calculation between the reference compound and each of the idea compounds can yield accurate predictions. Compounds not satisfying the similarity criteria should be removed from the list. Functionality to assess the FEP+ amenability is available through LiveDesign.
Select an initial set of 1000 molecules randomly to perform single edge FEP+ simulations with the reference molecule. The simulation time for these FEP+ scanning jobs can be shortened to 1–2 ns to get enough throughput provided that the 1–2 ns FEP+ protocol has been validated on a small number of compounds to generate similar results as the default 5 ns protocol.
Train a neural network model based on the FEP+ results of the initial set of molecules. Use DeepAutoQSAR to train a regression model for 4 hours of model search time.
Select the top 500 molecules with best potency based on the neural network model, and perform single edge FEP+ calculations on these molecules.

Visually inspect the compounds with best potency by single edge FEP+ and run FEP+ calculations on these compounds with full cycle closure. If full cycle closure FEP confirms their strong binding, synthesis these molecules and test their activity. Users can skip the following steps if potent and property satisfactory molecules are identified in this step.
Retrain the neural network model using the FEP+ data on all compounds with FEP+ predicted binding potencies, and reselect the top 500 molecules based on the updated neural network model to perform FEP+ calculations.
Repeat step 6 until the FEP+ scores of the top compounds selected by neural network model are similar to the neural network scores.
Perform FEP+ calculations on the top molecules with closed cycles and longer simulation time (5–10 ns in practice). Check the free energy convergence of these calculations and perform additional simulations (including, extending the simulation time, adding protein side chains in the REST region, custom core, and other debugging methods outlined in the FEP+ Best Practices) if needed to reach convergence. Synthesize these molecules and test their activity. If the experimental binding potency differs significantly from the prediction, analyze the possible reasons for the incorrect prediction, and adjust the FEP+ protocol to repeat the entire process.

A workflow streamlining many of these steps is available as an optional enhancement to LiveDesign.