Filtering Chemical Libraries


This post was partially motivated by a recent post from Karl Leswing describing how to use the DeepChem package to do virtual screening on a large database.  As part of the tutorial, Karl used the HIV sample file that is part of the DeepChem distribution to build a model.  This model was then used to select compounds from the ZINC database.  The tutorial is nice and the methodology is explained in a manner that is easy to follow.  The problem is that the molecules selected by the model are not what I would consider "drug-like".





The molecules reported in Karl's post have aryl sulfonic acid groups.  In fact, molecule D has 2 sulfonic acids AND a thiol.  To someone who has spent a bit of time looking through HTS data, these molecules scream false positive.  If an enterprising young computational chemist were to take this set of hits to an experienced medicinal chemistry colleague, my guess is that he/she would simply shake his/her head and say something like "that's nice, please come back when you have something real".

This brings us to the topic of this post, filtering chemical libraries.  Over the last 25 years, many people in the cheminformatics community (including me), have put together sets of substructure queries designed to identify compounds which are likely to produce artifacts in biochemical or cellular assays.  In the late 1990s, Mark Murcko and I put together a set of filters that we called REOS (Rapid Elimination of Swill).  This wasn't the first and certainly wasn't the last such effort.  Some examples from the literature include work by groups at

Like many other things in life, this topic has not been without controversy. Some people feel that filters are an essential way of avoiding problematic compounds and sifting through screening hits.  Others feel that filters are overzealous and may lead to the elimination of potentially valuable therapeutics.  My friend Peter Kenney considers the PAINS filters to be an abomination only rivaled by ligand efficiency and data binning.  Most people fall somewhere between these two extremes.

The controversy came to a boil with a 2017 paper by Tropsha and coworkers that questioned recent guidelines on the use of PAINS filters by the editors of the Journal of Medicinal Chemistry (JMC).  In the interest of full disclosure, I am a member of the JMC Editorial Advisory Board and helped to draft those guidelines. It should be noted that the JMC guidelines do not state that a paper will be rejected solely based on the presence of PAINS substructures.  The guidelines simply state that screening hits containing PAINS substructures should be supported by additional experimental evidence (e.g. SAR, structural information, orthogonal assays etc.).  In truth, any paper reporting hits from virtual screening should demonstrate the purity and the validity of the hits. Tropsha argued that the PAINS filters are somewhat arbitrary and have limited experimental validation.  He also pointed out that a number of FDA approved drugs contain PAINS filters.  The editors of JMC and a number of related ACS journals responded with a detailed editorial supporting the original guidelines.

There is a lot more I could say about structural alerts, but that's enough of a preamble.  After all, this blog is called Practical Cheminformatics, so let's get to it.  I'm sure that many people may want to run a set of in-silico filters on their compounds, but simply don't know where to start.  Fear not, you've come to the right place.

The ChEMBL database now contains a database table that features a large collection of substructure filters.  The table is called "structural_alerts" and, in ChEMBL 23, contains more than a thousand structural alerts from 8 different alert sets.  Of course, this may not be a lot of use to you if you're not a database geek and a cheminformatics geek.  Again, fear not, I've pulled the structural alerts from the database, cleaned up a few that misbehaved, and wrapped this up in an easy to use Python script.

The script, rd_filters.py, is available on my GitHub site.  The GitHub site has some examples, but let's take a look at what happens if we run one of the ChEMBL structural filter sets on the HIV dataset from DeepChem.  The file is deepchem/examples/hiv/HIV.csv.  First off, rd_filters.py expects a SMILES file and HIV.csv is a csv file.  No worries, we can fix this.

awk -F',' 'NR > 1{printf("%s MOL%04d\n",$1,NR-1)}'

In case you're not an awk geek, this uses a comma as a delimiter, skips the first line (which is a header), prints the first token on the line (the SMILES string), prints MOL followed by the line number padded with 4 zeros.  I subtract 1 from the line number (NR) because we skipped the first line.  Note that the file HIV.smi is also on the GitHub site.

Ok, now that we have a SMILES file, the fun can begin.  If we want to run HIV.smi with the default filters and the default Inpharmatica structural alerts, we will get output that looks like this.

rd_filters.py filter --in HIV.smi --prefix out

 using 8 cores
Using alerts from Inpharmatica
Wrote SMILES for molecules passing filters to out.smi
Wrote detailed data to out.csv
18499 of 41913 passed filters 44.1%
Elapsed time 22.87 seconds

There are a few things to note here.

  • First off, the script runs in parallel across multiple cores.  For the Python geeks in the crowd, this is due to the magic that is Pool.map.  By default, it uses all of the available cores, but this can be controlled with the --np flag (see the README.md for more information).  It's pretty fast, 22 seconds for 42K compounds on my MacBook Pro.
  • We used the structural alerts from the Inpharmatica set in the ChEMBL table, we can modify the configuration file rules.json to change to a different set of structural alerts. 
  • The script outputs 2 files, one with the SMILES for the compounds which pass all of the alerts and property filters, another with detailed filtering results.
  • Only 44% of the compounds in the input file passed the filters.  This is not good. 
Let's take a look and see if we can figure out what happened.

awk -F',' '{print $4}' out.csv | sort | uniq -c | sort -rn | head

 23450 OK
2643 Filter9_metal > 0
2401 Filter44_michael_acceptor2 > 0
1760 Filter82_pyridinium > 0
1296 Filter14_thio_oxopyrylium_salt > 0
1086 Filter20_hydrazine > 0
 930 Filter39_imine > 0
 894 Filter41_12_dicarbonyl > 0
 697 Filter5_azo > 0
 614 Filter1_2_halo_ether > 0

 6% of the compounds were eliminated by a single rule "Filter9_metal", let's look at the SMARTS for this alert.
 [!#1!#6!#7!#8!#9!#15!#16!#17]~[*,#1]

 This filter is designed to identify compounds with atom types outside those typically found in organic molecules
 H,C,N,O,F,P,S,Cl.  As you can see in the figure below, this filter does indeed pick up a lot of silly molecules that probably shouldn't be in anyone's screening collection.
 * Copper complexes
 * Zinc complexes
 * Boronates
 * Cobalt complexes
 * Arsenates



 Of course, the astute reader will also note that the filter tosses compounds with bromine, something I wouldn't do (the disulfide in MOL0049 is another story).  This highlights the fact that these sorts of things can't just be used blindly, you have to run the filters, take a look at what is being eliminated and adjust as necessary.

 To sum all of this up.

  •  Most screening collections contain a lot of molecules that probably shouldn't be there.
  •  Structural alerts provide a means of identifying and potentially eliminating some of these potentially problematic molecules.
  • The rd_filters.py script provides a convenient way of applying a number of different sets of structural alerts to a set of molecules. Please give it a try. 
  • A lot of these filters are somewhat subjective and based on people's experience.  4 of the 8 rule sets in ChEMBL will reject the sulfonic acids shown at the beginning of this post, the other four will not.

 Another thing we haven't talked about is, which of the 8 ChEMBL structural alert sets is the best one to use.  That's a discussion for another day.

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Comments

Post a Comment

Popular posts from this blog

We Need Better Benchmarks for Machine Learning in Drug Discovery

Image
Most papers describing new methods for machine learning (ML) in drug discovery report some sort of benchmark comparing their algorithm and/or molecular representation with the current state of the art.  In the past, I’ve written extensively about statistics and how methods should be compared .  In this post, I’d like to focus instead on the datasets we use to benchmark and compare methods.  Many papers I’ve read recently use the MoleculeNet dataset, released by the Pande group at Stanford in 2017, as the “standard” benchmark.   This is a mistake.  In this post, I’d like to use the MoleculeNet dataset to point out flaws in several widely used benchmarks.  Beyond this, I’d like to propose some alternate strategies that could be used to improve benchmarking efforts and help the field to move forward.   To begin, let’s examine the MoleculeNet benchmark, which to date, has been cited more than 1,800 times.   The MoleculeNet collection consists of 16 datasets divided into 4 categories.  Qua
Read more

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

Getting Real with Molecular Property Prediction

Image
Introduction If you believe everything you read in the popular press, this AI business is easy. Just ask ChatGPT, and the perfect solution magically appears. Unfortunately, that's not the reality. In this post, I'll walk through a predictive modeling example and demonstrate that there are still a lot of subtleties to consider. In addition, I want to show that data is critical to building good machine learning (ML) models. If you don't have the appropriate data, a simple empirical approach may be better than an ML model.  A recent paper from Cheng Fang and coworkers at Biogen presents prospective evaluations of machine learning models on several ADME endpoints.  As part of this evaluation, the Biogen team released a large dataset of measured in vitro assay values for several thousand commercially available compounds.  One component of this dataset is 2,173 solubility values measured at pH 6.8 using chemiluminescent nitrogen detection (CLND), a technique currently consider
Read more