Evaluating Large Ligand Libraries with Active Learning Glide

Tutorial Created with Software Release: 2026-3
Topics: Hit Discovery, Machine Learning, Small Molecule Drug Discovery, Virtual Screening
Products Used: AL-Glide

Tutorial files

1.7 GB
This tutorial is written for use with a 3-button mouse with a scroll wheel.
Words found in the Glossary of Terms are shown like this: Workspacethe 3D display area in the center of the main window, where molecular structures are displayedthe 3D display area in the center of the main window, where molecular structures are displayed
Abstract:

This tutorial demonstrates how to screen large ligand libraries efficiently using Active Learning Glide.

Tutorial Content
  1. Introduction
  2. Creating Projects and Importing Structures
  3. Setting up Active Learning Glide Docking
  4. Analyzing Active Learning Glide Results
  5. Conclusion and References
  6. Glossary of Terms

1. Introduction

Modern virtual screening libraries have exploded from the millions into the billions of compounds. With a publication indicating that screening larger ligand libraries leads to finding more binders and better binders for protein targets, there is an increased interest in screening ligand libraries of 1 million to 1+ billion ligands. Virtual enumeration tools make accessing virtual libraries of that size, and larger, also easy for researchers.

While larger libraries increase the chance of finding top-scoring hits, exhaustively docking them is not scalable and tractable. Active Learning Glide (AL-Glide) addresses this bottleneck by pairing physics-based Glide Docking with Schrödinger’s DeepAutoQSAR machine learning framework in an iterative loop:

  • Initial docking – The workflow selects a small, diverse subset of compounds and docks them with Glide SP.
  • Model training – A DeepAutoQSAR ML model is trained using the docking scores to predict scores for the full library.
  • Biased selection – ML predictions are used to select promising or uncertain compounds for the next round.
  • Iteration – This process repeats docking, training, and selection until the docking budget is reached.
  • Final rescoring – The top-ranked ligands predicted by the final ML model are then docked with full Glide SP to obtain final results.

Because ML evaluation is roughly 1000 times faster than docking (~20 ms per ligand), this approach lets researchers screen ultra-large libraries (≥ 500K compounds, scaling to 1B) in a single campaign rather than many smaller ones.

This tutorial uses a random selection of 1 million compounds from the Enamine REAL Database. You do not need to run LigPrep for the entire library prior to running AL-Glide, as only a percentage of the top-ranked compounds identified by the ML model will be docked. Only these top-ranked ligands need to be prepared using LigPrep, and this takes place as a part of the AL-Glide workflow once the entire library has been ranked using the final interaction of the ML model.

Further, to save time, the 5WIU structure has been prepared using the Protein Preparation Workflow panel and included in the tutorial files. Note that the binding site waters were retained during preparation as they were shown to be important for reproducing the known binding pose. These preparation steps are a necessary part of a virtual screen and must be done before docking. Please see the Introduction to Structure Preparation and Visualization tutorial for instructions on using the Protein Preparation Workflow. Additionally, a receptor grid file for 5WIU structure is included in the tutorial files. This grid was generated using the Receptor Grid Generation panel without any constraints.

In this tutorial, you will learn how to screen large ligand libraries efficiently using Active Learning Glide. Note that for structure-based screening of libraries even larger than 1 billion compounds, we recommend using Generative Glide.

2. Creating Projects and Importing Structures

  1. Open Maestro and create a new project named AL-Glide_D4.prj for this tutorial.
  2. Download the tutorial zip file including input files and reference outputs here: https://www.schrodinger.com/sites/default/files/s3/release/current/Tutorials/zip/al-glide_d4-enamine.zip
  3. After downloading the zip file, unzip the contents in your Working Directorythe location that files are saved for ease of access throughout the tutorial.

Figure 1-1. Importing the prepared receptor-ligand structure.

  1. Go to File > Import Structures.
  2. Navigate to and choose 5WIU-prepared.maegz from the tutorial files in your Working Directorythe location that files are saved.

3. Setting up Active Learning Glide Docking

In this section, you will set up Active Learning Glide docking calculations. Note that you will proceed with the default training options of using 50,000 ligands per iteration, standard training time of 4 hours per iteration, and 3 iterations of training. We strongly recommend using 3 iterations of training for optimal results.

The following interactive tutorial guides you through setting up the Active Learning Glide docking calculations in Maestro.

Depending on the size of your ligand library of interest and compute resources, you may want to alter the default training options. The recommended maximum sample size is 100,000 per iteration. Benchmarks show the enrichment score is effectively the same and the recovery rate is the same for larger samples sizes. While it is common to assume that as library sizes grow so should training sets, we would advise against scaling training sets beyond 100,000. Further, using more time than the standard could improve the models if they are not already well fit.

Note

Active Learning Glide can also be run in pilot mode. This is a way to test the effectiveness of Active Learning Glide on a pilot library (default size is 50K) instead of the whole library. One iteration of training on a 5K subset of ligands is performed and the whole pilot library is docked in full. A performance report is generated to aid analysis, but it is not the same in-depth report generated from a full Active Learning Glide calculation. The docking results from the pilot library are reused in full Active Learning Glide run, so as not to duplicate calculations. You may hover over the option in the panel to learn more.

Before scaling up to 100 million or even a billion compounds – a process that can take days or weeks – it is critical to run a pilot screen and have a medicinal chemist examine the scaffolds you obtain. This is also an appropriate stage to refine your library one more time, for example by applying additional filters to eliminate any remaining undesirable high-scoring chemotypes if they are in the top 10% of your pilot.

4. Analyzing Active Learning Glide Results

The goal of virtual screening is often to identify diverse hit scaffolds with strong activity and high potency. There are many ways to narrow down the results post screening – such as volume-based clustering, fingerprint-based clustering, similarity filtering, or forwarding compounds directly into advanced screening methods like Glide WS and AL-ABFEP+.

In this section, you will evaluate the results of the Active Learning Glide calculation using the Hit Analyzer panel. It is a powerful tool for analyzing very large datasets (can efficiently process ~1M compounds without performance delays).

The Hit Analyzer panel provides an interface for searching through the results of a screening based on customizable filters. Note that while we’ll walk you through some of the available filters in this section, when working with your own data, you will need to determine which filters and cutoffs are helpful to understand your screening results and determine next steps.

An AL-Glide run generates a set of outputs that let you trace the entire campaign from raw selection through the final list of hits. Some important output files are described here briefly. For each AL-Glide active learning cycle, a dedicated glide_active_learning_screen_iter_N folder is created, capturing that round’s four internal stages: the compounds selected for docking, their Glide docking results and any failures, the cumulative training data fed to the machine learning model (which grows larger with each iteration), and the resulting trained model along with its test-set performance metrics.

Once all iterations complete, the top-ranked compounds are re-docked with and compiled into the glide_active_learning_screen_RescoreNode_iter_3_rescore.vsdb – a screening database containing the final 3D poses, complete Glide scoring terms, and precomputed molecular properties (MW, AlogP, PSA, HBD/HBA, shape similarity, etc.) for every hit. This .vsdb file is exactly what you load into the Hit Analyzer panel for filtering and triage.

Finally, the glide_active_learning_screen_report.html ties everything together as a self-contained, interactive dashboard, letting you click through each stage per iteration to inspect model accuracy and hit enrichment. It gives you a transparent record of how the active learning loop converged on its final compound selection.

The following interactive tutorial guides you through the steps for analyzing the results of your calculation using the Hit Analyzer panel.

5. Conclusion and References

In this tutorial, you set up an Active Learning Glide calculation in Maestro GUI and learned how this workflow enables efficient screening of ultra-large libraries by combining machine learning with traditional ligand docking.

6. Glossary of Terms

Entries - a simplified view of the Project Table that allows you to perform basic operations such as selection and inclusion

included - the entry is represented in the Workspace, the circle in the In column is blue

Project Table - displays the contents of a project and is also an interface for performing operations on selected entries, viewing properties, and organizing structures and data

Scratch Project - a temporary project in which work is not saved, closing a scratch project removes all current work and begins a new scratch project

selected - (1) the atoms are chosen in the Workspace. These atoms are referred to as "the selection" or "the atom selection". Workspace operations are performed on the selected atoms. (2) The entry is chosen in the Entries (and Project Table) and the row for the entry is highlighted. Project operations are performed on all selected entries

Working Directory - the location that files are saved

Workspace - the 3D display area in the center of the main window, where molecular structures are displayed