Predicting Properties of Small Molecules with Kernel-Based...
Transcript of Predicting Properties of Small Molecules with Kernel-Based...
![Page 1: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/1.jpg)
Predicting Properties of Small Molecules with
Kernel- Based Machine Learning Methods
Klaus-Robert Müller, Matthias Rupp, Katja Hansen,
Gisbert Schneider et al.
Predicting Properties of Small Molecules with
Based Machine Learning Methods
Müller, Matthias Rupp, Katja Hansen, Timon Schroeter,
![Page 2: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/2.jpg)
Machine Learning in a nutshell
Typical scenario: learning from dataTypical scenario: learning from data
• given data set X and labels Y (generated
• LEARN/INFER underlying unknown mapping
Y = f(X)
Example: distinguish toxic and non-toxic compoundsExample: distinguish toxic and non-toxic compounds
BUT: how to do this optimally with good performance
? f
generated by some joint probabilty distribution p(x,y))
mapping
Y = f(X)
compounds, metabolically stable compounds ...compounds, metabolically stable compounds ...
performance on unseen data?
![Page 3: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/3.jpg)
Gaussian Processes
![Page 4: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/4.jpg)
Gaussian Process in 2-dim
![Page 5: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/5.jpg)
Gaussian Process in 3-dim
![Page 6: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/6.jpg)
Gaussian Process in 5-dim
![Page 7: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/7.jpg)
Gaussian Process in 100-dim
![Page 8: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/8.jpg)
And here is the GP
![Page 9: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/9.jpg)
Covariance Functions
![Page 10: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/10.jpg)
Gaussian Process Models
![Page 11: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/11.jpg)
Bayes Theorem
![Page 12: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/12.jpg)
GP Training
![Page 13: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/13.jpg)
Prediction with GPs
![Page 14: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/14.jpg)
GP Learning – a cartoon
![Page 15: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/15.jpg)
GP the Movie
![Page 16: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/16.jpg)
Application: predict chemical endpointsendpoints from descriptors
![Page 17: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/17.jpg)
Data available: solubilty ( physico
[Schwaighofer et al. JCIM 2007, Schroeter et al,
physico -chemical property)
et al. JCIM 2007, Schroeter et al, ChemMedChem 2007]
![Page 18: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/18.jpg)
Issues: Multiple Measurements
![Page 19: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/19.jpg)
Fitness for Purpose
Log s
![Page 20: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/20.jpg)
Descriptors
ML model ML model ML model can tell which descriptors are importantML model can tell which descriptors are important
![Page 21: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/21.jpg)
Results Solubility Schering in House (Schering in House ( at pH 7)
5-fold CV5-fold CV
![Page 22: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/22.jpg)
Results Solubility Huuskonen
![Page 23: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/23.jpg)
Results LogD (at pH 7)
[Schroeter et al 2008][Schroeter et al 2008]
![Page 24: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/24.jpg)
GPs for Classification
![Page 25: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/25.jpg)
Measuring Metabolic Stability (bio
[Schwaighofer
(bio -chemical property)
Schwaighofer et al, J Comp Mol Des 2008]
![Page 26: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/26.jpg)
Measuring Metabolic Stability: DetailedDetailed set-up
![Page 27: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/27.jpg)
Quantifying Performance
![Page 28: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/28.jpg)
Model Performance
![Page 29: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/29.jpg)
Predicting biological properties
![Page 30: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/30.jpg)
Virtual Screening: Optimization CriteriaCriteria
![Page 31: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/31.jpg)
The Study
![Page 32: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/32.jpg)
Choosing the Kernel
![Page 33: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/33.jpg)
PPAR-Gamma Data Set
kPCA
![Page 34: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/34.jpg)
Results
![Page 35: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/35.jpg)
Virtual Screening: cherry picking
Uncertainty ofprediction
Highly certain prediction& strong binding
Higher pKi values indicate stronger binding
![Page 36: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/36.jpg)
Detailed Results
Fraction of inactives in top 20
![Page 37: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/37.jpg)
Results: prospective validation
![Page 38: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/38.jpg)
Best Hit: a natural product
[Rupp et al., ChemMedChem 2009, Steri et al.,
Cynodon dactylon
et al., Bioorg Med Chem Lett 2010]
![Page 39: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/39.jpg)
Misc RemarksRemarks
![Page 40: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/40.jpg)
Explaining single Predictions
![Page 41: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/41.jpg)
Ranking or Regression
![Page 42: Predicting Properties of Small Molecules with Kernel-Based ...helper.ipam.ucla.edu/publications/ccsws2/ccsws2_9829.pdf · Kernel-Based Machine Learning Methods Klaus-Robert Müller,](https://reader034.fdocuments.us/reader034/viewer/2022043000/5f750c1b34bf9f1e7679f60b/html5/thumbnails/42.jpg)
Conclusion
• GPs and SVM have been applied
• CYP, hERG, metabolic stability, toxicity
• ranking, explaining, error bars
• Kernel holds the key
•Machine Learning Methods are universal
in many practical applications
toxicity, log p, log d, solubility, mutagenicity
universal tools and useful