Using Counterfactuals to Understand Machine Learning Models


While machine learning (ML) models have become integral to many drug discovery efforts, most of these models are "black boxes" that don't explain their predictions.  There are several reasons we would like to be able to explain a prediction. 

  • Provide scientific insights that will guide the design of new compounds. 
  • Instill confidence among team members.  As I've said before, a computational chemist only has two jobs; to convince someone to do an experiment and to convince someone not to do an experiment.  These jobs are much easier when you can explain the "why" behind a prediction. 
  • Debugging and improving models.  Improving a model is easier if you can understand the rationale behind a prediction.  
As I wrote in a previous post,  identifying and highlighting the molecular features that drive an ML prediction can be difficult.  One recent promising approach is the counterfactuals method published by Andrew White's group at the University of Rochester.  Counterfactuals have their origins in the assessment of credit risk.  In many jurisdictions, a bank must provide an explanation for denying someone credit.  The bank can't simply say, "we won't give you a credit card because our neural network doesn't like you".   One way of explaining a credit risk prediction is to identify two people with similar characteristics, where one had their credit approved, and another had their credit declined by a black box ML model.  The slight differences between these individuals (e.g., the number of credit cards) can then be used to explain the prediction.  

We can apply a similar approach to explaining quantitative structure-activity relationships (QSARs).  Let's consider an ML model trained to predict aqueous solubility.  If we have a molecule predicted to be insoluble, we could gain insights into this prediction by examining predictions for similar molecules predicted to be soluble and noticing where the differences are.  For instance, we can generate analogs of Palbociclib, a molecule predicted to be insoluble, and see that replacing a methyl ketone with a carboxylic acid leads to a molecule predicted to be soluble.  We can then infer that the methyl ketone may be one cause of poor solubility.  As a bonus, we get the structure of an analog predicted to overcome that liability. 


Let's take a look at how we would implement something like this.  I've assembled a demo Jupyter notebook available on GitHub for those who like code.

1. Generate a set of analogs for a molecule of interest.  
There are several ways to computationally generate analogs for a molecule.  In the original publication, Andrew White's group used the STONED SELFIES method to generate a set of analogs.  As I pointed out in an earlier post, STONED SELFIES tend to produce a lot of very strange molecules.  For instance, if we start with Palbociclib, STONED SELFIES generates these analogs, which I don't find realistic or useful.  
The later version of the White group's exmol method uses a reaction-based approach that generates more realistic molecules but sometimes fails to generate a sufficient number of analogs.  As an alternative, I used the Chemically Reasonable Mutations (CReM) approach published by Pavel Polishchuk.  I wrote a post in 2020 describing how CReM can be used.  In short, CReM captures the chemical environments around functional groups and swaps functional groups in similar environments to create new molecules.  This is an oversimplification; please check out Pavel's paper for a complete explanation. 

2. Generate predictions and similarities for the analogs
We begin the counterfactual process by generating a set of analogs, calculating their predicted solubility, and their similarity to Palbociclib, our reference molecule.   Note that there are a few molecules in the set below that I should have eliminated using functional group filters.  


3. Sort the analogs by similarity to the reference molecule
Since Palbociclib is predicted to be insoluble (labeled as false), we want to examine similar molecules predicted to be soluble (labeled as true).  We can do this by selecting only the molecules where the prediction is "true" and sorting by similarity.  


4. Examine detailed comparisons
To simplify comparisons, I added a routine that will align pairs of molecules and highlight their differences.  For more on this, please check out the code. 

Counterfactuals provide an intuitive understanding of machine learning predictions and can suggest ways to overcome liabilities.  Please give the code a spin and see if you find it helpful. Andrew White's group has an excellent Perspective article in JCTC that describes counterfactuals and several other approaches to explainable ML for molecules.   











Comments

  1. AsharMay 3, 2023 at 8:55 AM

    Fascinating article and thanks for such a clear explanantion of this topic. I love your blog! While in the Chemoinformatics QSAR world, using Counterfactual is a cool new way to analyse 'black box' ML models and I believe has parallels to the traditional 'open box' Matched Pair Analysis, the broader connections between Causal Inference methodologies and ML methodologies is becoming ever deeper.
    Here as you so well put it, Causal Inference is being used to explain ML QSAR models; there are scanarios where ML models are being made more "transportable" using Causal Inference methodologies and even more interestingly ML models are enhacing Causal Inference itself by providing "valid" counterfactual inferences. These ideas are coming out from Epidemiology and Econometrics fields.

    ReplyDelete
  2. Noel O'BoyleMay 4, 2023 at 8:07 AM

    Just on nomenclature, aren't these "activity cliffs", but for predicted properties? I like the idea of retaining that nomenclature to link it to the existing literature.

    ReplyDelete
    Replies
    1. Pat WaltersMay 5, 2023 at 7:53 AM

      Thanks, Noel. I don't know if "activity cliffs" are exactly what we're talking about here. In this case, we're referring to a structural change that flips the label from one category to the other. That change doesn't necessarily have to be huge. In the case of activity cliffs, we have a small change in chemical structure that leads to a large change in an activity or property. I see a subtle distinction, but others may disagree.

      Delete

Post a Comment

Popular posts from this blog

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 c...
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...
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.  Asid...
Read more