Let's say I have created a QSAR model on some training dataset, and validated it on a test set.
Next, I'd like to create a model that gives me a 1.0 (or nearby) if the molecule I present to the QSAR model looks similar to the training set, but a 0.0 (or nearby) if it looks dissimilar.
Currently, I was thinking about training a one class Support Vector Machine on the training set.
Broad picture: I'd like to get a warning in case the QSAR model is presented with something that looks too far from anything it was trained on. The warning would mean: "be careful with this prediction".
Dear Mr Berenger, you can use the method of Williams plot (you plot standardized residuals vs leverages), you can see the outliers.
Best Regards
Dear Mr Berenger
If you had created the model you can use AmbitDiscovery software to explore if the molecules of the test set are in the aplicability domain of the model. That software permits you to use diferent metrics to solve the problems. Other way is using the Williams´ graph with leverage values, we describe bot methods in several papers of our group.
best regards
You might just want to check if your unknow molecule is extrapolating that is if the values for your predictors are within the ranges of the predictors used when modelling. This becomes more complicated when considering combinations of predictors. Looking at residuals may not reveal that you are looking at extrapolated predictions.
Another way of doing this is to create a low dimensional summary using principal components analysis (with say 95+% explained variance) and projecting your new observations in this space. It should more or less fall in the bulk of your training data.
Have a look at the paper:
Article Comparison of Different Approaches to Define the Applicabili...
additionally you can find a matlab toolbox (The Applicability Domain toolbox) that is a collection of MATLAB modules for calculating the Applicability Domain (AD) of a model for a set of test samples and verify if they fall inside or outside the model AD. The toolbox implements a set of AD approaches based on several strategies, such as: bounding box, bounding box on PCs, convex hull, leverage, distance to centroid, k Nearest Neighbours (kNN) approach with fixed k, k Nearest Neighbours (kNN) approach with variable k, Probability Density Function-based methods. A graphical user interface (GUI), which allows an easy AD evaluation and analysis of results, is also provided.
http://michem.disat.unimib.it/chm/download/adinfo.htm
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training