ALL Metrics
-
Views
-
Downloads
Get PDF
Get XML
Cite
Export
Track
Research Note
Revised

Understanding covariate shift in model performance

[version 3; peer review: 2 approved]
PUBLISHED 17 Oct 2016
Author details Author details
OPEN PEER REVIEW
REVIEWER STATUS

This article is included in the Cheminformatics gateway.

Abstract

Three (3) different methods (logistic regression, covariate shift and k-NN) were applied to five (5) internal datasets and one (1) external, publically available dataset where covariate shift existed. In all cases, k-NN’s performance was inferior to either logistic regression or covariate shift. Surprisingly, there was no obvious advantage for using covariate shift to reweight the training data in the examined datasets.

Keywords

covariate shift, model building, ChEMBL, logistic regression, k-NN

Revised Amendments from Version 2

In response to reviewer feedback, Figure 1 has been replaced with a new version.

See the authors' detailed response to the review by Robert Sheridan

Introduction

A common prerequisite in supervised learning algorithms is that the training and prediction data arise from the same distribution and are independently and identically distributed (iid)1. Intuitively this is justified, as one should not expect to learn a classifier on one distribution of examples and apply it to accurately predict labels of examples drawn from a different distribution. Covariate shift is a machine learning technique that can be utilized in supervised learning when the training and prediction distributions are known to differ, but the concept being learned remains stationary. While standard machine learning classifiers are trained and then used to predict on arbitrary compounds, covariate shifted classifiers must be trained specifically for each prediction dataset. This is because covariate shifted classifiers weight the training distribution to be more similar to the prediction distribution. A recent book provides an excellent overview of the current state of the art in covariate shift methods2.

Covariate shift frequently occurs during the drug discovery process where learning systems are built to predict physiochemical properties of interest. Initially a chemistry team may focus on a particular chemical series, and information from this series is used to train a learning system. As the project progresses, the chemistry team may refocus their efforts on a new, structurally distinct series. The accuracy of prospective computational predictions on the new series may be compromised as these molecules originate from a distribution that is distinct from the molecular set used to train the learning tool.

For example one may wish to build a learning system to predict hERG activity (unwanted cardiovascular toxicity). Initially the computational tool is trained using series A but must now predict on series B. The concept “binding to hERG” is fixed, however the area of interest has transitioned from chemical series A to chemical series B. The feature vectors describing these two sets are likely related but potentially different; and as such, their covariates have shifted. Put more mathematically, the probability of observing a feature vector from the prediction set is different from the probability of observing a feature vector from the training set. That is, the training and prediction sets are non-iid. A well-constructed learning system will recognize that predictions on series B are outside the “domain of applicability” of the model and predict with low confidence. The covariate-shift method attempts to adjust the domain of applicability so that it is more aligned with the prediction set. It is analogous to a nearest neighbor classifier but employs distributions rather than individual examples. Covariate shifted classifiers weight examples from the training set to create a distribution that is more aligned with the prediction set. This weighted data set is then used to train the classifier, resulting in a covariate shifted classifier. As such, covariate shift is applied at the distribution level whereas nearest neighbor methods are applied at the example level. Once a training set has been shifted, it can be used by any machine learning algorithm.

Covariate shift methods typically reweight instances in the training data so that the distribution of training instances is more closely aligned with the distribution of instances in the prediction set. This is accomplished by providing more weighting during model building to an instance in the training set that are similar to an instance in the prediction set. It has been shown3 that the appropriate importance weighting factor w(x) for each instance “x” in the training set is:

w(x)=pp(x)pt(x)(1)

where pt (x) is the probability of seeing instance x in the training set and pp (x) is the probability of seeing x in the prediction set. It is important to note that only the feature vector values (not their labels) are used in reweighting. The importance weighting scheme is intuitively understandable. If the probability of seeing a particular instance from the training set in the prediction is very small, then this instance should carry little weight during the training process and consequently have little effect on the decision function.

Figure 1 plots two Gaussian distributions and w(x). If instances from the blue distribution are used for training a classifier to predict on an instance from the green distribution then the red curve gives the importance of each instance. Note the increased importance for instances from the training distribution overlapping with high-density regions of the prediction distribution.

fdfd821b-3e35-4795-9bef-5d010a3f8cfa_figure1.gif

Figure 1. Train, prediction and importance.

Methods

For our experiments, we use a logistic regression classifier where each training instance is weighed by its importance w(x). For the calculation of w(x) we use the Kullback-Leibler Importance Estimation Procedure (KLIEP) method developed by Sugiyama4. The KLIEP method is based on the Kullback-Leibler divergence theorem and attempts to find weights to minimize the divergence from ptrain(x) to ppredict(x). Briefly, the importance is modeled as a linear function:

w(x)=i=1bαiφi(x)(2)

The αi are the weights to be learned and φi the basis functions. The importance weight from Equation 1 can be rearranged and used to estimate the probability of observing a feature vector in the predictive set.

p^p(x)=w(x)pt(x)(3)

The KL divergence from pp(x) to its estimate p^p(x) can then be expressed as:

KL[pp(x)||p^p(x)]=pp(x)log(pp(x)pt(x)w(x))dx

After algebraic manipulation, removing terms independent of w^(x) and adding constraints to ensure proper normalization, a final objective function to be maximized can be derived as (see 4 for details):

maximize{αl}l=1b[j=1nplog(l=1bαlφl(xj))]subjectto:j=1ntl=1bαlφl(xj)=1andα1,α2,,αb0

The resulting problem is convex and can be solved using standard optimization techniques. The result is an expression for w(x) that allows calculating weights for a training instance x. These weights can then be incorporated when training a classifier to obtain a covariate shifted version of the classifier.

Toy example

To demonstrate the use of covariate shift methods, we repeated a simple toy experiment as detailed in 3. Figure 2 graphically displays the results we obtained.

fdfd821b-3e35-4795-9bef-5d010a3f8cfa_figure2.gif

Figure 2. Classification using covariate shift.

The red training points are drawn from two (2) two-dimensional Gaussian distributions representing a class 1 and a class 2. The green prediction points are drawn from a slightly rotated version of the training distributions. The red line plots the classifier obtained when training on only the training points; the green line plots the classifier trained on both the training and prediction points (the optimal classifier in this case). The blue line plots the classifier trained on the training data that was weighted by the importance factor as estimated by the KLIEP method. Note how the blue line is shifted towards the optimal classifier, demonstrating the effect of the KLIEP algorithm and covariate shift.

Experiments

Dataset 1.The beta secretase IC50 data derived from the ChEMBL database.
Units are in nM.

Using the Python programming language we implemented the KLIEP method for determining weights for use in covariate shift5. In principle, covariate shift is applicable to any classifier that allows weighting of input instances (e.g. support vector machines and random forest). For this study we wanted to isolate the effects of covariate shift and therefore selected a classifier without adjustable parameters and used logistic regression (LR). Logistic regression is a classification technique analogous to linear regression and is applicable when the dependent variable is categorical6. We combined logistic regression with KLIEP and applied it to five different in-house ADME (absorption, distribution, metabolism and excretion) assays and one external dataset (beta secretase). The cutoff values for determining the binary categories for the compounds in each dataset are listed in Table 1. Due to inherent noise in the assays we discard data where the assay values are between the positive and negative cutoffs listed in the Table 1. We compare KLIEP+Logistic Regression (KL+LR) to Logistic Regression and a k-NN (using Tanimoto similarity) classifier (k=5).

Table 1. Proprietary Assays Utilized for Covariate Shift Analysis.

Data SetPositive CutoffNegative Cutoff
hERGIC50 <10uMIC50 > 15uM
Human Liver Microsome (HLM)stable > 60% remainunstable < 30% remain
Rat Liver Microsome (RLM)stable > 60% remainunstable < 30% remain
Solubility (water)insoluble < 10uMsoluble > 200uM
Solubility (DMSO)insoluble < 10uMsoluble > 50uM

Legend: The cutoff values for determining the binary categories (actives or inactives) for the compounds in each dataset are listed.

For each dataset the molecules were sorted by compound registration date. The first 75% of the data comprised the master training set while the remainder formed the master prediction set. Temporal ordering of the data represents the evolving coverage of chemical space by drug discovery projects and consequently captures the natural “shifting” of the covariates. Classifier performance statistics are generated by performing twenty different runs, each on a random 80% of the master files. Performance statistics for each classification task are then obtained by averaging the results of the twenty individual folds. In all cases, OpenEye7 path fingerprints are used as feature vectors. We experimented with different fingerprints provided by OpenEye (MACCS 166 bit structural keys and circular fingerprints) and found that they had no significant effect on the outcome.

To ensure the data was amenable to covariate shift we generated classifiers separating “training” from “prediction” data. Figure 3 shows performance of LR on this separation task. For each dataset we are able to compute highly accurate classifiers. This indicates that the training and prediction data are drawn from different distributions and hence are appropriate for covariate shift methods. This is a necessary condition for covariate shift but does not imply model improvement over unweighted data.

fdfd821b-3e35-4795-9bef-5d010a3f8cfa_figure3.gif

Figure 3. Classification of train versus test.

Figure 4 compares the performance of KL+LR, LR and k-NN on the five (5) datasets. One can see from the graph that KL+LR failed to provide any statistical improvement over standard LR.

fdfd821b-3e35-4795-9bef-5d010a3f8cfa_figure4.gif

Figure 4. Performance of KL+LR, LR and k-NN.

We extended the study to include an external dataset provided by ChEMBL8,9 such that others could use their own fingerprints and independently support or refute our claims. We chose the beta secretase IC50 data as it is a well established biochemical screen, highly accurate and contains > 7000 data points crossing multiple orders of magnitude, which are publically available. Using OpenEye path fingerprints and K-Means clustering we clustered the dataset into two clusters, A and B. Under cross-validation, a logistic regression classifier was able to separate the two clusters with a high level of accuracy (90%) indicating that the clustered dataset would be appropriate for application of the covariate shift algorithm. Ten random subsets of molecules from cluster A were used to train a logistic regression classifier using covariate shift which was then used to predict on molecules from cluster B. The performance of the shifted classifier was compared to an unshifted classifier trained and tested on the same clustered datasets and random splits. The process was repeated by training on molecules from cluster B and predicting on molecules from cluster A. Analogous to the internal datasets, as measured by overall classifier accuracy, there was no statistical advantage for application of covariate shift (Shifted Accuracy: 82.95% +/- 1.6%; Unshifted Accuracy 82.73% +/- 1.2%).

A possible explanation for the failure of the covariate shift method to provide a boost in predictive performance could be that the calculated importance weights are all similar. This would cause each training example to exert the same influence on the decision function and thus the importance weighting would have no effect. This was not the case. Figure 5 plots the cumulative distribution function of the importance weight for the training set compound. The plot demonstrates that weights are distributed across a range of classifier performance.

fdfd821b-3e35-4795-9bef-5d010a3f8cfa_figure5.gif

Figure 5. Cumulative distribution function.

Conclusions

We have applied the KLIEP method to five (5) internal data sets and one (1) external data set where covariate shift was evident. Although KL+LR was an advantage over k-NN, there is no statistical advantage of reweighting the training dataset. We are surprised with this outcome and are currently exploring other datasets where application of covariate shift may improve the predictions.

Data availability

F1000Research: Dataset 1. The beta secretase IC50 data derived from the ChEMBL database, 10.5256/f1000research.8317.d11788210

Comments on this article Comments (1)

Version 3
VERSION 3 PUBLISHED 17 Oct 2016
Revised
Version 2
VERSION 2 PUBLISHED 17 Jun 2016
Revised
Discussion is closed on this version, please comment on the latest version above.
  • Author Response 17 Oct 2016
    Georgia McGaughey, Modeling & Informatics, Vertex Pharmaceuticals, Boston, USA
    17 Oct 2016
    Author Response
    New version of Figure 1 has been submitted. The AUCs of the prediction and training sets are now the same and the label for the y axis has been removed ... Continue reading
  • Discussion is closed on this version, please comment on the latest version above.
Author details Author details
Competing interests
Grant information
Copyright
Download
 
Export To
metrics
Views Downloads
F1000Research - -
PubMed Central
Data from PMC are received and updated monthly.
- -
Citations
CITE
how to cite this article
McGaughey G, Walters WP and Goldman B. Understanding covariate shift in model performance [version 3; peer review: 2 approved]. F1000Research 2016, 5(Chem Inf Sci):597 (https://doi.org/10.12688/f1000research.8317.3)
NOTE: If applicable, it is important to ensure the information in square brackets after the title is included in all citations of this article.
track
receive updates on this article
Track an article to receive email alerts on any updates to this article.

Open Peer Review

Current Reviewer Status: ?
Key to Reviewer Statuses VIEW
ApprovedThe paper is scientifically sound in its current form and only minor, if any, improvements are suggested
Approved with reservations A number of small changes, sometimes more significant revisions are required to address specific details and improve the papers academic merit.
Not approvedFundamental flaws in the paper seriously undermine the findings and conclusions
Version 3
VERSION 3
PUBLISHED 17 Oct 2016
Revised
Views
24
Cite
Reviewer Report 18 Oct 2016
Robert Sheridan, Cheminformatics Department, Merck Research Laboratories, Rahway, NJ, USA 
Approved
VIEWS 24
The only issue I am still having trouble with is equation 1.
 
w(x)=Pp(x)/Pt(x)
 
where the P’s represent probabilities for molecule x. Given that the training and test sets are distinct points (molecules) in a ... Continue reading
CITE
CITE
HOW TO CITE THIS REPORT
Sheridan R. Reviewer Report For: Understanding covariate shift in model performance [version 3; peer review: 2 approved]. F1000Research 2016, 5(Chem Inf Sci):597 (https://doi.org/10.5256/f1000research.10228.r17055)
NOTE: it is important to ensure the information in square brackets after the title is included in all citations of this article.
Version 2
VERSION 2
PUBLISHED 17 Jun 2016
Revised
Views
22
Cite
Reviewer Report 21 Jun 2016
Martin Vogt, Department of Life Science Informatics, Bonn-Aachen International Center for Information Technology (B-IT), LIMES (Life & Medical Sciences Institute) Program Unit Chemical Biology and Medicinal Chemistry, University of Bonn, Bonn, Germany 
Approved
VIEWS 22
The revised paper has been significantly improved and addresses a number of issues raised in the original reviews. However, a few issues remain that should be addressed.

1. The description of Figure 1 is inconsistent with the ... Continue reading
CITE
CITE
HOW TO CITE THIS REPORT
Vogt M. Reviewer Report For: Understanding covariate shift in model performance [version 3; peer review: 2 approved]. F1000Research 2016, 5(Chem Inf Sci):597 (https://doi.org/10.5256/f1000research.9428.r14444)
NOTE: it is important to ensure the information in square brackets after the title is included in all citations of this article.
Views
31
Cite
Reviewer Report 20 Jun 2016
Robert Sheridan, Cheminformatics Department, Merck Research Laboratories, Rahway, NJ, USA 
Approved with Reservations
VIEWS 31
Having read the revised paper it seems better, but I’m still somewhat puzzled about a few things.
 
  1. I am getting the feeling that the K-NN method is meant as a baseline control method since
... Continue reading
CITE
CITE
HOW TO CITE THIS REPORT
Sheridan R. Reviewer Report For: Understanding covariate shift in model performance [version 3; peer review: 2 approved]. F1000Research 2016, 5(Chem Inf Sci):597 (https://doi.org/10.5256/f1000research.9428.r14472)
NOTE: it is important to ensure the information in square brackets after the title is included in all citations of this article.
  • Author Response 22 Jun 2016
    Georgia McGaughey, Modeling & Informatics, Vertex Pharmaceuticals, Boston, USA
    22 Jun 2016
    Author Response
    I'm responding directly to the questions you posed, in that order.

    1.  Agreed that we could have added such a sentence to the paper.

    2.  We didn't examine ... Continue reading
  • Reviewer Response 02 Aug 2016
    Martin Vogt, Department of Life Science Informatics, Bonn-Aachen International Center for Information Technology (B-IT), LIMES (Life & Medical Sciences Institute) Program Unit Chemical Biology and Medicinal Chemistry, University of Bonn, Bonn, Germany
    02 Aug 2016
    Reviewer Response
    Concerning Figure 1:
    According to equation (1) w(x) = p_p(x)/p_t(x) there is a certain symmetry between the importance weight w(x) and the training pdf p_t(x).

    This means that in ... Continue reading
  • Author Response 03 Aug 2016
    Georgia McGaughey, Modeling & Informatics, Vertex Pharmaceuticals, Boston, USA
    03 Aug 2016
    Author Response
    The figure is really meant to be an illustration of the importance weight - and not mathematically accurate with respect to pdfs. In actuality, the green and blue curve look ... Continue reading
COMMENTS ON THIS REPORT
  • Author Response 22 Jun 2016
    Georgia McGaughey, Modeling & Informatics, Vertex Pharmaceuticals, Boston, USA
    22 Jun 2016
    Author Response
    I'm responding directly to the questions you posed, in that order.

    1.  Agreed that we could have added such a sentence to the paper.

    2.  We didn't examine ... Continue reading
  • Reviewer Response 02 Aug 2016
    Martin Vogt, Department of Life Science Informatics, Bonn-Aachen International Center for Information Technology (B-IT), LIMES (Life & Medical Sciences Institute) Program Unit Chemical Biology and Medicinal Chemistry, University of Bonn, Bonn, Germany
    02 Aug 2016
    Reviewer Response
    Concerning Figure 1:
    According to equation (1) w(x) = p_p(x)/p_t(x) there is a certain symmetry between the importance weight w(x) and the training pdf p_t(x).

    This means that in ... Continue reading
  • Author Response 03 Aug 2016
    Georgia McGaughey, Modeling & Informatics, Vertex Pharmaceuticals, Boston, USA
    03 Aug 2016
    Author Response
    The figure is really meant to be an illustration of the importance weight - and not mathematically accurate with respect to pdfs. In actuality, the green and blue curve look ... Continue reading
Version 1
VERSION 1
PUBLISHED 07 Apr 2016
Views
29
Cite
Reviewer Report 25 Apr 2016
Martin Vogt, Department of Life Science Informatics, Bonn-Aachen International Center for Information Technology (B-IT), LIMES (Life & Medical Sciences Institute) Program Unit Chemical Biology and Medicinal Chemistry, University of Bonn, Bonn, Germany 
Approved with Reservations
VIEWS 29
The study investigates the influence of accounting for covariate shift in classification performance using logistic regression models. Overall, this short paper is very well and clearly written, however the method section should be expanded (see below). Although no increase in ... Continue reading
CITE
CITE
HOW TO CITE THIS REPORT
Vogt M. Reviewer Report For: Understanding covariate shift in model performance [version 3; peer review: 2 approved]. F1000Research 2016, 5(Chem Inf Sci):597 (https://doi.org/10.5256/f1000research.8943.r13393)
NOTE: it is important to ensure the information in square brackets after the title is included in all citations of this article.
Views
31
Cite
Reviewer Report 08 Apr 2016
Robert Sheridan, Cheminformatics Department, Merck Research Laboratories, Rahway, NJ, USA 
Approved with Reservations
VIEWS 31
This is potentially an important negative result in QSAR, however I think some revision is necessary because some aspects are unclear.

The title “Understanding covariate shift…” is a little weak. One could say “Failure of covariant shift to improve model performance…”

It ... Continue reading
CITE
CITE
HOW TO CITE THIS REPORT
Sheridan R. Reviewer Report For: Understanding covariate shift in model performance [version 3; peer review: 2 approved]. F1000Research 2016, 5(Chem Inf Sci):597 (https://doi.org/10.5256/f1000research.8943.r13266)
NOTE: it is important to ensure the information in square brackets after the title is included in all citations of this article.

Comments on this article Comments (1)

Version 3
VERSION 3 PUBLISHED 17 Oct 2016
Revised
Version 2
VERSION 2 PUBLISHED 17 Jun 2016
Revised
Discussion is closed on this version, please comment on the latest version above.
  • Author Response 17 Oct 2016
    Georgia McGaughey, Modeling & Informatics, Vertex Pharmaceuticals, Boston, USA
    17 Oct 2016
    Author Response
    New version of Figure 1 has been submitted. The AUCs of the prediction and training sets are now the same and the label for the y axis has been removed ... Continue reading
  • Discussion is closed on this version, please comment on the latest version above.
Alongside their report, reviewers assign a status to the article:
Approved - the paper is scientifically sound in its current form and only minor, if any, improvements are suggested
Approved with reservations - A number of small changes, sometimes more significant revisions are required to address specific details and improve the papers academic merit.
Not approved - fundamental flaws in the paper seriously undermine the findings and conclusions
Sign In
If you've forgotten your password, please enter your email address below and we'll send you instructions on how to reset your password.

The email address should be the one you originally registered with F1000.

Email address not valid, please try again

You registered with F1000 via Google, so we cannot reset your password.

To sign in, please click here.

If you still need help with your Google account password, please click here.

You registered with F1000 via Facebook, so we cannot reset your password.

To sign in, please click here.

If you still need help with your Facebook account password, please click here.

Code not correct, please try again
Email us for further assistance.
Server error, please try again.