P53: Prediction of Activity.

There are about 2000 P53 assays  in the ChEMBL DB. Among all we have chosen the reporter assay described in the data description section: National Center for Biotechnology Information. PubChem BioAssay Database; AID=504706. The assay contains 20k records in the ChEMBL database, corresponding to the active records. Fortunately, records of inactive compounds are also stored in the PubChem DB, which have been incorporated to the compound list to built the predictive models.  There is a total of 300k compounds assayed in the P53 luc reporter assay indicated.

P53 assays in the ChEMBL DB ranked by the amount of compounds assayed. Top 13 assays.

Activity distribution histogram of the P53 luc reporter assay. Inset zooming into the active records.

For these 300k compounds there are 3.6M records in the ChEMBL DB corresponding to more than 200K assays upon more than 5k targets.

Most representative assays carried out with the P53 compounds.

ChEMBL activity score distribution histogram for the 3.6M assays. Inset zooming into the active records.

Many of these compounds have been also tested in growth inhibition assays with tumor cell lines. In the charts below we study the correlation of  the activity enhancing P53 expression vs the activity on each cell line. Note that many drugs inhibit cell growth through mechanisms not involving P53, but most of the molecules with P53 activity also inhibit growth in most tumor cell lines.

Correlation between activity at P53 luc reporter and cell growth inhibition in tumor cell lines (1).

Correlation between activity at P53 luc reporter and cell growth inhibition in tumor cell lines (2).

Calculating R2 and slope for linear regressions performed for all assays, and adding extra variables, as the number of unique molecules per assay and the average potency at target we can select the best correlating targets.

Dual selection of best correlating targets

Best correlating targets.

Targets showing best correlations with the enhancement of P53 expression assay

For model evaluation, data have been splitted in validation and test sets The validation set has been labelled for classification (Binary and multiple) and regression and used to generate the model. The test set is tested against the model and the predictions compared to the previously hidden activity columns. Accuracy of predictions is then assessed by the classical machine learning evaluation parameters in the charts below.

Validations set are passed through decision trees (Rpart: binary classification) random forests (classification and regression), and Naive Bayes (binary and multiple classification) algorithms. Each algorithm is tested against different combinations of variables: RPart is used with all possible variables, while Naive bayes is tested against a reduced set of variables and all variables (using classifiers from e1071-default- and NaiveBayes libraries). For each combination of classification algorithm and variable set, a confusion matrix and a bar diagram with most meaningful statistical quality descriptors is charted below. If regression, there is a correlation plot instead.

Confusion matrixes for classification algorithms. If there is no indication, NB uses e1071 library. NBLib refers to the naivebayes library.

Statisticalqualifiers for each algorithm. If there is no indication, NB uses e1071 library. NB package refers to the naivebayes library.

Random forest regression prediction compared to P53 actual values.

As a result of this evaluation, the random forest classifier will be the core predictor on the whole ChEMBL DB. NB classifiers provide a very high false positives rate and will be discarded. The random forest regression trends to underestimate potency for less potent compounds in particular, but despite this alert , it is still a valid predictor and shows an almost null population in the false positives section. The decision tree classifier will be employed to add elements to the random forest selection as it is the only classifier that uses the whole set of variables.

Selection of active compounds.

We use a panel where predictions through different models are interactively compared to allow quality control of the selection.   As most compounds include multiple experimental events recorded in the database, and there is a prediction for each event, we need to calculate an index of activity for each compound. For regressions, this is the average P53 score (combinedP53Score), which uses actual values when they exist, and predictions when they don’t. Classification will use the ratio between the number of  times a compound has been predicted active vs the total number of events (countRatio). In this example we will use binomial random forests and Rpart classification.

Left chart in the dashboard below visualizes the random forests binomial countRatio for each unique molecule-target pair vs the number of occasions that this pair has a result in the DB. Marking the most active (either actual or predicted) in a convenient number of occasions in this chart causes the marked compounds to appear in the top right plot that compares the average P53 potency (actual and predicted) with the number of records for each molecule. Marking the desired compounds in this plot raises the down right chart with the Rpart classification countRatio.

Selection of active compounds may be as stringent or inclusive as we want to make it by applying the desired thresholds to selections performed on each dashboard section.

Compound selection dashboard.

Correlation with anti tumor activity.

In the ChEMBL DB there are 73 cell growth screens carried out on tumor cell lines with 40k records each. Next two plots compare the predicted and actual P53 scores with the chemblScores for each anti-tumor screen. It can be observed the existence of a population of compounds whose activity in the P53 reporter assay correlates with inhibition of cell growth in tumor cell lines.

Correlation between activity enhancing P53 expression and inhibition of tumor cell lines growth. Actual and predicted values.

Correlation between activity enhancing P53 expression and inhibition of tumor cell lines growth. Actual values.

Most potent P53 Compounds

Some of the most potent structures…