Clustering Fragment Screening Hits With a Self-Organizing Map

 


Clustering Fragment Screening Hits With a Self-Organizing Map (SOM)



In a paper, "Fragment binding to the Nsp3 macrodomain of SARS-CoV-2 identified through crystallographic screening and computational docking", published last year by scientists from UCSF and the Diamond Light Source, the authors reported more than 200 structures of fragments bound to the Nsp3 macrodomain of SARS-CoV-2.   I wanted to dig into the supporting data from this paper and compare fragments that were binding to the same site.  I couldn't find a good tool to do this, so I decided to write one.  After writing the code, I thought that others might find the methodology instructive.   I have a Jupyter notebook with the code on GitHub.  There is also a link to run the notebook on Google Colab.  Here's a quick outline of the workflow.  

1.  Download the structures from the PDB 
The supporting information for the paper has a list of the PDB ids for the structures.  In addition, I was able to obtain an Excel spreadsheet (in the GitHub repo) with the SMILES for all of the ligands from Jaime Fraser and Galen Correy at UCSF.   Prody made it easy to programmatically download a lot of structures. 

2.  Align the structures
To facilitate the comparison, I needed to align all of the structures to a common reference.  After speaking with Galen Correy and Stefan Gabhauer from UCSF,  I chose 5RS7 as a reference structure and aligned the remaining structures to it.  There's actually a bit more to this.  Many of the structures have multiple copies of the protein.  I did a few tricks with Prody to separate the chains before aligning.  Please note that this notebook exploits a feature that isn't in Prody versions before 2.0.  If you're running this on your own machine, please do "pip install -U prody" before trying to run the notebook. 

3.  Assign bond orders to the ligands
It astounds me that the PDB has been in existence for more than 50 years and we still don't have a reasonable means of representing bond orders in ligands.  I've ranted about this before, so I won't do it again.  Anyway, I used a method similar to the one described in this blog post, which exploits the AllChem.AssignBondOrdersFromTemplate method in the RDKit.  

4. Identify structures where multiple ligands are bound 
In a few cases, there are multiple ligands bound in the same structure.  To avoid confusing the clustering algorithm, I decided to separate these into individual ligands.  This is pretty easy to do using Chem.GetMolFrags from the RDKit.  

5.  Calculate the center of each ligand
There are a number of ways to do the clustering. I opted for something simple and represented each molecule as its centroid.  

6. Cluster the fragment structures
Readers of this blog probably know about my fondness for Self-Organizing Maps (SOMs), which I wrote about here and here.  For this application,  I like the fact that the SOM not only clusters the data but also provides a "map" indicating which clusters are adjacent.  I described each molecule with the x,y,z coordinates of its centroid and clustered.   The output of the SOM is shown on the right in the figure above.   It's pretty easy to see from this figure that cluster 0 is the largest cluster and that clusters 2,4,6,7 and 9 are adjacent to cluster 0.  

7.   (Extra) Clustering with HDBSCAN
As I was doing this work, I came across a post on clustering geographic data using Hierarchical Density-based Spatial Clustering of Applications with Noise.  The data we're dealing with here consists of coordinates representing the centers of molecules so it's sort of like geographic data.  Anyway, I thought I'd give it a try.  One of the nice things about HDBSCAN is that it doesn't try to put all of the molecules into clusters.  Singletons and outliers are labeled as cluster -1.   This part of the notebook isn't very well developed, but I thought I'd put it out there for people to play with.   I'll revisit HDBSCAN in a future post. 

Resources
The code on GitHub is here
The code on Google Colab is here

Acknowledgments
I'd like to thank Stefan Gahbauer, Brian Shoichet, Jaime Fraser, and Galen Correy for their help with the data. Also, big thanks to James Krieger and She Zhang for help with Prody.

Comments

Post a Comment

Popular posts from this blog

AI in Drug Discovery 2023 - A Highly Opinionated Literature Review (Part I)

Image
Here’s the first part of my review of some interesting machine learning (ML) papers I read in 2023.  As with the previous editions , this shouldn’t be considered a comprehensive review.  The papers covered here reflect my research interests and biases, and I’ve certainly overlooked areas that others consider vital.  This post is pretty long, so I've split it into three parts, with parts II and III to be posted in the next couple of weeks.    I. Docking, protein structure prediction, and benchmarking II. Large Language Models, active learning, federated learning, generative models, and explainable AI III. Review articles 2023 was a bit of a mixed bag for AI in drug discovery.  Several groups reported that the deep learning methods for protein-ligand docking weren’t quite what they were initially cracked up to be.  AlphaFold2 became pervasive, and people started to investigate, with mixed success, the utility of predicted protein structures.  There were reports of significant advanc
Read more

Generative Molecular Design Isn't As Easy As People Make It Look

Image
I was taken aback by a recent CNBC article entitled “ Generative AI will be designing new drugs all on its own in the near future ”.  I should know better than to pay attention to AI articles in the popular press, but I feel that even scientists working in drug discovery may have a skewed perception of what generative AI can and can’t do.  To understand exactly what’s involved, it might be instructive to walk through a typical generative molecular design workflow and point out a few things.  First, these programs are far from autonomous.  Even when presented with a well-defined problem, generative algorithms produce a tremendous amount of nonsense.  Second, domain expertise is essential when sifting through the molecules produced by a generative algorithm.  Without a significant medicinal chemistry background, one can’t make sense of the results.  Third, while a few nuggets exist in the generative modeling output, a lot of work and good old-fashioned cheminformatics are required to ext
Read more

AI in Drug Discovery - A Highly Opinionated Literature Review (Part II)

Image
Picking up where we left off in Part I , this post covers several other ML in drug discovery topics that interested me in 2023.  Some areas, like large language models, are new, and most of the work is at the proof-of-concept stage.  Others, like active learning, are more mature, and several groups are starting to explore nuances of the methods.   Here’s the structure of Part II.  4. Large Language Models 5. Active Learning 6. Federated Learning 7. Generative Models 8. Explainable AI 9. Other Stuff 4. Large Language Models The emergence of GPT-4 and ChatGPT brought considerable attention to large language models (LLMs) in 2023.  In November and December, several large pharmas held “AI Day” presentations featuring LLM applications for clinical trial data analysis. Many of these groups demonstrated the ability of LLMs to ingest large bodies of unstructured clinical data and subsequently generate tables and reports based on natural language queries.  Aside from some very brief demos on co
Read more