predicting activity cliffs - can machine learning handle special cases?
TRANSCRIPT
Predicting Activity Cliffs - Can We Use Machine Learning for Special Cases?
Rajarshi GuhaNIH Center for Translational Therapeutics
August 4, 2011Joint Statistical Meeting, Miami Beach
Outline
• Structure-activity landscapes• Characterization• Prediction
Structure Activity Relationships
• Similar molecules will have similar activities• Small changes in structure will lead to small
changes in activity• One implication is that SAR’s are additive• This is the basis for QSAR modeling
Martin, Y.C. et al., J. Med. Chem., 2002, 45, 4350–4358
Exceptions Are Easy to Find
Tran, J.A. et al., Bioorg. Med. Chem. Lett., 2007, 15, 5166–5176
Ki = 39.0 nM Ki = 1.8 nM
Ki = 10.0 nM Ki = 1.0 nM
Structure Activity Landscapes
• Rugged gorges or rolling hills?– Small structural changes associated with large
activity changes represent steep slopes in the landscape
– But traditionally, QSAR assumes gentle slopes – Machine learning is not very good for special cases
Maggiora, G.M., J. Chem. Inf. Model., 2006, 46, 1535–1535
Characterizing the Landscape
• A cliff can be numerically characterized• Structure Activity Landscape Index (SALI)
• Cliffs are characterized by elements of the matrix with very large values
Guha, R.; Van Drie, J.H., J. Chem. Inf. Model., 2008, 48, 646–658
Visualizing SALI Values
• The SALI graph– Compounds are nodes– Nodes i,j are connected if SALI(i,j) > X– Only display connected nodes
What Can We Do With SALI’s?
• SALI characterizes cliffs & non-cliffs• For a given molecular representation, SALI’s
gives us an idea of thesmoothness of the SAR landscape
• Models try and encodethis landscape
• Use the landscape to guidedescriptor or model selection
Descriptor Space Smoothness
• Edge count of the SALI graph for varying cutoffs• Measures smoothness of the descriptor space• Can reduce this to a single number (AUC)
Feature Selection Using SALI
• Instead of fingerprints, we use molecular descriptors
• SALI denominator now uses Euclidean distance
• 2D & 3D random descriptor sets– None are really good– Too rough, or– Too flat
2D
3D
Measuring Model Quality
• A QSAR model should easily encode the “rolling hills”• A good model captures the most significant cliffs• Can be formalized as
How many of the edge orderings of a SALI graph does the model predict correctly?
• Define S (X ), representing the number of edges correctly predicted for a SALI network at a threshold X
• Repeat for varying X and obtain the SALI curve
SALI Curves
Predicting the Landscape
• Rather than predicting activity directly, we can try to predict the SAR landscape
• Implies that we attempt to directly predict cliffs– Observations are now pairs of molecules
• A more complex problem– Choice of features is trickier– Still face the problem of cliffs as outliers– Somewhat similar to predicting activity differences
Scheiber et al, Statistical Analysis and Data Mining, 2009, 2, 115-122
Motivation
• Predicting activity cliffs corresponds to extending the SAR landscape
• Identify whether a new molecule will perform better or worse compared to the specific molecules in the dataset
• Can be useful for guiding lead optimization, but not necessarily useful for lead hopping
Predicting Cliffs
• Dependent variable are pairwise SALI values, calculated using fingerprints
• Independent variables are molecular descriptors – but considered pairwise– Absolute difference of descriptor pairs, or– Geometric mean of descriptor pairs– …
• Develop a model to correlate pairwise descriptors to pairwise SALI values
A Test Case
• We first consider the Cavalli CoMFA dataset of 30 molecules with pIC50’s
• Evaluate topological and physicochemical descriptors
• Developed random forest models– On the original observed
values (30 obs)– On the SALI values
(435 observations)
Cavalli, A. et al, J Med Chem, 2002, 45, 3844-3853
Double Counting Structures?
• The dependent and independent variables both encode structure.
• But pretty low correlations between individual pairwise descriptors and the SALI values
Model Summaries
• All models explain similar % of variance of their respective datasets
• Using geometric mean as the descriptor aggregation function seems to perform best
• SALI models are more robust due to larger size of the dataset
Original pIC50RMSE = 0.97
SALI, AbsDiffRMSE = 1.10
SALI, GeoMeanRMSE = 1.04
Test Case 2
• Considered the Holloway docking dataset, 32 molecules with pIC50’s and Einter
• Similar strategy as before• Need to transform SALI values • Descriptors show minimal
correlation
Holloway, M.K. et al, J Med Chem, 1995, 38, 305-317
Model Summaries
• The SALI models perform much poorer in terms of % of variance explained
• Descriptor aggregation method does not seem to have much effect
• The SALI models appear to perform decently on the cliffs – but misses the most significant
Original pIC50RMSE = 1.05
SALI, AbsDiffRMSE = 0.48
SALI, GeoMeanRMSE = 0.48
Model Summaries
• With untransformed SALI values, models perform similarly in terms of % of variance explained
• The most significant cliffs correspond to stereoisomers
Original pIC50RMSE = 1.05
SALI, AbsDiffRMSE = 9.76
SALI, GeoMeanRMSE = 10.01
Test Case 3
• 38 adenosine receptor antagonists with reported Ki values; use 35 for training and 3 for testing
• Random forest model on the SALI values performed reasonable well (RMSE = 7.51, R2=0.62)
• Upper end ofSALI rangeis better predicted
Kalla, R.V. et al, J. Med. Chem., 2006, 48, 1984-2008
Test Case 3
• For any given hold out molecule, range of error in SALI prediction is large
• Suggests that some form of domain applicability metric would be useful
• The dataset does not containing really big cliffs
• Generally, performance is poorer for smaller cliffs
Model Caveats
• Models based on SALI values are dependent on their being an SAR in the original activity data
• Scrambling results for these models are poorer than the original models but aren’t as random as expected
Conclusions
• SALI is the first step in characterizing the SAR landscape
• Allows us to directly analyze the landscape, as opposed to individual molecules
• Being able to predict the landscape could serve as a useful way to extend an SAR landscape
Acknowledgements
• John Van Drie• Gerry Maggiora• Mic Lajiness• Jurgen Bajorath