Screening Ultra-Large Libraries with Generative Glide

Tutorial Created with Software Release: 2026-2
Topics: Hit Discovery, Machine Learning, Small Molecule Drug Discovery, Virtual Screening
Products Used: Glide

Tutorial files

1 GB
Abstract:

In this tutorial, you will learn how to perform targeted structure-based screening on ultra-large libraries using Generative Glide.

Tutorial Content
  1. Introduction
  2. Prerequisites for performing virtual screening with Generative Glide
  3. Setting up Generative Glide
  4. Analyzing the results of a virtual screen using Generative Glide
  5. Conclusion and References

1. Introduction

Chemical vendor library sizes have grown into the billions, allowing easy synthetic access to large swaths of chemical space. Together with the observations made by Schoichet et al that rigid-receptor docking is a powerful tool to identify binders, this has incentivized applying docking to ever larger library sizes. While optimized for speed, classical Glide docking is not fast enough to handle library sizes in the millions with reasonable cost. Workflows like Active-Learning Glide (AL-Glide) expand the library sizes reachable with Glide to approximately 100M compounds. Beyond this point, it becomes increasingly infeasible to exhaustively screen a database by enumerating every compound even with AL-Glide.

Generative Glide uses a more targeted approach, using a modern machine learning architecture as sketched out in Figure 1 below to perform a more targeted exploration of the chemical space relevant for each specific project. To achieve this, a generative model – the prior – is trained to produce SMILES strings representing compounds which are likely to be in the chemical space covered by a given library. For popular vendor libraries such as Enamine REAL space or Wuxi Galaxi, pre-trained prior models are available from Schrödinger. The prior is used to instantiate multiple agent models, which are trained using reinforcement learning so that the molecules they generate also fit additional, project-specific criteria. These criteria are evaluated by a scorer model, which is trained to predict docking scores for the specific target based on a pilot screening dataset. The scorer also evaluates compounds based on how well they fit in a user-specified physicochemical property space and encourages the agents to generate diverse compounds. After several iterations of agent refinement, the final set of generated compounds are passed through a Bloom filter to confirm whether they are directly purchasable and docked using the docking model to obtain the final hit list.

A Bloom filter is a compact probabilistic data structure for fast membership testing. It may yield false positives but never false negatives. For molecular libraries, a Bloom filter built from known molecules enables rapid in-library queries: candidate molecules can be quickly checked for possible presence, with only positives requiring exact verification against the full actual library contents.

Figure 1: A schematic overview of the Generative Glide workflow.

Options for training the supervised model

By default, Generative Glide trains an ensemble of supervised scoring models on a fixed-size dataset as a compromise between scoring accuracy and speed.

If you are willing to trade a small accuracy hit for a significant speedup of the supervised model training stage, you can enable the option to train a single model.

Alternatively, if you want to have a supervised model that is really tailored to the chemical space of interest, you can enable the Active Learning option to train a single scoring model in three iterations.

In this tutorial, you will set up a virtual screen using Generative Glide on an ultra-large library from the WuXi GalaXi space to find small molecule ligands for A2A, a well-established drug target.

2. Prerequisites for performing virtual screening with Generative Glide

To screen a library with Generative Glide, you will need a glide docking grid for your target as well as the library you wish to screen against.

For comprehensive guidance and best practice recommendations on designing ligand libraries for virtual screening, see the corresponding online certification course or learning path.

Generative screening of popular vendor libraries

Currently, we recommend using Generative Glide on vendor libraries such as Enamine REAL Space, WuXi Galaxi, or their subsets. For these libraries, Schrödinger provides up-to-date prior models and bloom filters if you have a Generative Glide license.

These files are not bundled with this tutorial due to their size (~20 GB). Please reach out to your Application Scientist or help@schrodinger.com to get access.

As usual, you should prepare your protein, generate a receptor grid, and validate the docking model with any known actives and decoys. For this tutorial, you can find a ready-to-go docking grid for A2A (using the 3EML structure) in the zip archive. Check the Introduction to Structure Preparation and Visualization and Structure-Based Virtual Screening using Glide tutorials for more details or click through the interactive tutorials below for a quick reminder on how to obtain a receptor grid for your own system:

3. Setting up Generative Glide

Warning: Beta product ahead

This tutorial covers a product which is currently in open beta testing. During this stage, we are refining both its backend and graphical user interface and performing additional scientific validation. Please be aware that specific panel features may not be fully functional at this point and the images and files in this tutorial may not exactly match what you see due to the rapid pace of development.

3. Analyzing the results of a virtual screen using Generative Glide

Once your Generative Glide job is complete, you can analyze the results. This process has three steps: First, you should check various performance metrics for the training of the scorer model (and, if using a custom library, for the prior). Then, you can dive into exploring the chemical matter on a high level, checking the diversity of generated compounds and identifying promising scaffolds. Finally, you can decide on the compounds you wish to bring forward into the next stage of the screening funnel and extract their structures.

3.1 Get a first look at the screening results and performance using the Generative Glide HTML report

  1. Find and open the HTML report in the a2a_wuxi_part2_tutorial/a2a_wuxi_part2_tutorial_agent folder in the tutorial files.
  2. Scroll through the individual plots in the report to understand how your Generative Glide screening went.

Note

In addition to this overview report, you can find additional metrics and plots for the individual models and stages in their respective subdirectories. This information can be helpful for technical or debugging purposes, but should not be generally necessary.

The principal component analysis plot gives you an intuition for how well the chemical space explored by the generative models is aligned with the chemical space covered by the training data for the prior training. This plot is only relevant if you are using a custom library rather than using one of the pre-trained priors provided by Schrödinger.

If the ‘Prior Samples’ cloud is much smaller or larger than the ‘Prior training population’ cloud, something has probably gone wrong with training the prior.

The Tanimoto similarity histogram shows how similar the compounds generated by the prior are to their closest neighbors in the chemical space covered during prior training. As with the previous plot, you should only need to check this plot if using a custom library.

A distribution centered at lower similarity (e.g., 0.3–0.4) indicates that the Prior has successfully generalized beyond simple memorization, enabling the exploration of novel chemical regions while maintaining the underlying logic of the training data.

The Supervised model training section shows key statistics for the model trained to predict docking scores based on your receptor grid.

The initial plot in this section shows the correlation between the true and predicted docking scores. You can find the R² correlation coefficient in the text.

The supervised model in the tutorial example files was only trained for four hours rather than the default of 12 hours, leading to fairly low correlation coefficients on the hold-out data set and shortening the run time for the entire workflow by 8 hours.

Reducing the training time like this can be a helpful approach if you’re pressed for time, as the supervised model does not need to be very predictive to guide the agents towards generating good dockers. As you will see below, correlation coefficients of 0.2-0.3 as seen here are already sufficient to significantly enrich good dockers in the hit list.

Still, we generally recommend training the supervised model for the default 12 hours if possible.

The Generative Agent training section shows some key metrics for the screening, such as the total number of generated (unique) compounds, and which fraction of generated compounds was confirmed to be in-library using the Bloom filter.

The plots below the statistics compare the distributions between the compounds generated by the prior and the agents. There’s a plot for each property filter you specified, plus one for the docking scores.

The differences in the distributions in the Generative Agent training section show how the agents are generating compounds that are sampling a sub-region of chemical space defined by your docking model and physicochemical properties compared to the prior (which represents the entire screening library).

The Generated Compound Analysis section gives you an overview of the compounds which pass all filters: that is, all compounds in the right property space which are confirmed to be in the vendor library and successfully dock to the receptor grid.

In the first plot of the section, you can see how the ML model predicting docking scores has biased the agents to generate better-scoring compounds.

The Bemis-Murcko Scaffold report shows the five most frequently-generated scaffolds. Note that none of them occurs in more than 30 compounds. Over 12000 scaffolds are present in the final compound list.

These common scaffolds seem to be well-compatible with the pocket and might provide room for optimization through decoration.

The interactive molecule visualization allows you to see some representative generated compounds at a glance.

The Enriched Substructure Analysis section lists the substructures most frequently found in the generated compounds. Frequently occurring substructures can highlight key functionality for interacting with the receptor.

You can see that one of the two highly similar substructures highlighted in red is present in over half of the generated compounds. To prevent the agents from focusing too much on this part of chemical space, you could re-run Generative Glide and exclude these substructures by following the instructions below.

If you want to see how this exclusion affects the results of the screening, you can find the HTML report file and output .vsdb files for a run where these two substructures are excluded in the tutorial files. These results are also what you will continue analyzing in the next section.

Finally, the Workflow Summary section shows you some technical details for the run, such as a breakdown of the times for the individual parts of the workflow which can be helpful for troubleshooting.

3.2 Looking into the hits in more detail with the Hit Analyzer

So far, you’ve only had a birds-eye view of the screening results via the HTML report. To help you decide which of the generated compounds to bring forward to the next stage of the screening (e.g. GlideWS or AB-FEP+), we recommend using the Hit Analyzer.

In this section, you will load the results into the Hit Analyzer in Maestro and identify 1000 diverse compounds to bring forward into more accurate scoring methods. Note that rather than using the results discussed in the previous section, we will use the output of a second Generative Glide run, where the two most enriched substructures (present in ~50% of generated compounds) were excluded to further diversify the explored chemical space. Additionally, we have limited the number of compounds in the virtual screening results database to 10000 to keep the size of the tutorial files manageable.

Note

There is no one-size-fits-all approach for analyzing virtual screening results, as the relevant properties and criteria will be unique to each project. This tutorial only scratches the surface of what you can do with the Hit Analyzer and the Schrödinger suite in general. For a more comprehensive overview of building screening pipelines and analyzing screening results, check out the Virtual screening with integrated physics & machine learning online course.

Follow the instructions here to load the results into the Hit Analyzer. Experiment with the different filters to get a feel for the chemical space explored by the generated compounds.

In addition to viewing the 2D structures in the Hit Analyzer, also take a look at the distributions of the physicochemical properties of the generated compounds, and export them to view the docked poses.

Distributions of physicochemical properties across all (black) and the filtered (blue) compounds.

Looking at the property distributions (see image above), you can see that the full ranges for the various properties you specified when setting up the Generative Glide screen are well-represented in the final docked compounds. Note that the methods used to calculate these properties during the generative and scoring stages and after the docking are different, so you may see compounds whose properties fall outside the ranges you specified when setting up the screening.

Finally, when inspecting the generated compounds and their docked poses, you can see that many of the generated compounds have extended heterocycles which can mimic A2A’s natural ligand adenosine.

Generally, we recommend directly re-scoring the outputs of Generative Glide with Glide WS or AB-FEP+ rather than performing property-based filtering at this stage.

4. Conclusion and References

In this tutorial, you learned how to set up and run Generative Glide to screen an ultra-large vendor library using generative models rather than exhaustive enumeration. In addition to providing a receptor grid, you specified the target space of physicochemical properties. You analyzed the results of the screening by looking through the generated HTML report and learned how to identify and exclude over-enriched substructures and liabilities from the generated compounds. Finally, you used the Hit Analyzer to view property distributions of the generated compounds, extracted a diverse set of compounds and checked their docked poses.

The next steps from this point would be to bring as many of the docked compounds as possible into the next stage of the screening funnel. Depending on the number of compounds and available budget, we recommend using either GlideWS or AB-FEP+ to re-score the compounds from this screening. See the further learning session below for the corresponding tutorials.

For further reading: