020 激酶的相似性：配体结构

Aim of this talktorial

The aim of this talktorial is to investigate kinase similarity through ligand profiling data (ChEMBL29). In the context of drug design, the following assumption is often made: if a compound was tested active on two different kinases, it is suspected that these two kinases may have some degree of similarity. We will use this assumption in this talktorial. The concept of kinase promiscuity is also covered.

Contents in Theory

• Kinase dataset
• Bioactivity data
• Kinase similarity descriptor: Ligand profile
  • Kinase similarity
  • Kinase promiscuity
• From similarity matrix to distance matrix

Contents in Practical

• Define the kinases of interest
• Retrieve the data
• Preprocess the data
  • Hit or non-hit
• Kinase promiscuity
• Kinase similarity
  • Visualize similarity as kinase matrix
  • Save kinase similarity matrix
• Kinase distance matrix
  • Save kinase distance matrix

References

• Kinase dataset: Molecules (2021), 26(3), 629
• ChEMBL database
  • Website: https://www.ebi.ac.uk/chembl/
  • Paper: Nucleic Acid Research (2017), 45(D1), D945-D954
• KinMap
  • Website: http://www.kinhub.org/kinmap/
  • Paper: BMC Bioinformatics (2017), 18(1), 16

Theory

Kinase dataset

We use the kinase selection as defined in Talktorial T023.

Bioactivity data

In order to measure kinase similarity through ligand profiling data, bioactivity data is retrieved from the well-known ChEMBL database and the query focuses on human kinases. Luckily, a curated version of ChEMBL29 is already freely available through the Openkinome organization, see https://github.com/openkinome/kinodata. For more details on querying the ChEMBL database, please refer to Talktorial T001.

In drug design, it is common to binarize the activity of a compound against a target of interest as a "hit" or "non-hit". Practically speaking, this is done using a cutoff value for measured activity. If the activity is greater than the cutoff, the compound is labeled as active (hit), and inactive (non-hit) otherwise.

Figure 1: Number of ChEMBL29 bioactivities per kinase - as collected in kinodata - mapped onto the Manning kinome tree using KinMap. The figure has been generated in Talktorial T023.

Kinase similarity descriptor: Ligand profile

As a measure of similarity, we use ligand profiling data in this talktorial.

Kinase similarity

We use the following metric as similarity between two kinases $K_i$ and $K_j$ :

$$\text{similarity}(K_i,K_j)=\frac{\#\text{ of compounds that were tested active on both }{K}_i\text{ and }{K}_j}{\#\text{ of compounds that were tested on both }{K}_i\text{ and }{K}_j}.$$

Assuming that only one compound was tested on two kinases, and that the compound was tested as active for one and inactive for the other, then the similarity between these two kinases would be zero.

Kinase promiscuity

Computing the similarity between a kinase and itself may be interpreted as kinase promiscuity, where the similarity described above would therefore represent the fraction of active compounds over all tested compounds.

From similarity matrix to distance matrix

As discussed in Talktorial T024, we convert the similarity matrix to a distance matrix.

Practical

Define the kinases of interest

Let's load the kinase selection as defined in Talktorial T023.

Retrieve the data

We retrieve a pre-curated version of a kinase subset of ChEMBL29 freely available at Openkinome, see https://github.com/openkinome/kinodata/releases/tag/v0.3.

Preprocess the data

We look at the type of activity and the associated units.

Let's keep the entries which have pIC50 values only.

The DataFrame contains many columns that won't be necessary for the rest of the notebook which are therefore removed. Only relevant information is kept, namely the canonical SMILES of the compound (compound_structures.canonical_smiles), the measured activity (activities.standard_value) and the UniProt ID of the kinase (UniprotID). These columns are renamed for readability.

NA values are dropped.

We only keep the data for the query kinases:

Let's look at example data (which corresponds to the first row in the kinase selection DataFrame):

Some compounds have been tested several times against a target, as shown below.

We have a look at those compounds.

In this example (demo mode), the first molecule is gefitinib, a known FDA-approved drug against EGFR.

As a simple workaround — since we prefer to have one activity value per compound-kinase pair — we keep the value for which the compound has the best activity value, i.e., the highest pIC50 value.

Hit or non-hit

Finally, we binarize the pIC50 values to obtain hit or non-hit using a cutoff. We use a pIC50 cutoff of $6.3$, similarly to the cutoff used in Molecules (2021), 26(3), 629.

Kinase promiscuity

We now look at the kinase promiscuity.

For a given kinase, three values are computed:

1. the total number of measured compounds against the given kinase,
2. the number of active compounds against the kinase, and
3. the fraction of active compounds, i.e., the ratio of active compounds over the total number of measured compounds per kinase.

Let's see what information we get for the first kinase in our dataset:

Let's create a table from these values for all kinases:

Let's beautify the table:

From the table, we notice that CDK2 is the least (in yellow) and BRAF the most (in red) promiscuous kinase.

Kinase similarity

We now investigate how we can use the similarity measure discussed in the Theory part to compare kinases.

Let's look at the values and similarity between two kinases.

Visualize similarity as kinase matrix

Let's first look at the non-reduced fraction of number of active compound against total number of compounds to have an idea of the counts.

Note that the total number of tested compounds as well as the number of active compounds on two kinases vary largely.

• For the p110a-ErbB2 pair, there are none.
• For p110a and [BRAF, CDK2, LCK, MET and p38a], there are less than $13$ commonly tested compounds.
• In contrast, the EGFR-ErbB2 pair has $1170$ commonly tested compounds, of which $662$ were active on both.

Now let's look at the similarity, in this case, the reduced fraction:

Note that the diagonal contains the previously discussed promiscuity values.

As mentioned above, no compounds were measured on both ErbB2 and p110a and therefore creates a np.nan entry which can be problematic for algorithmic reason.

As a simple workaround, we will fill the NA values with zero.

Save kinase similarity matrix

Kinase distance matrix

The similarity matrix $SM$ is converted to a pseudo-distance matrix (all entries of the similarity matrix are between $0$ and $1$):

Finally, we set the diagonal values to $0$ and we obtain the kinase distance matrix:

Save kinase distance matrix

Discussion

In this talktorial, we investigate how activity data can be used as a measure of similarity between kinases. The fraction of compounds tested as actives over the total number of measured compounds is a way of accessing the similarity. Moreover, using the same rationale, the promiscuity of a kinase can be quantified using the ratio of active compounds over measured compounds.

When working with these data, we have to keep in mind that some kinases have much higher coverage with respect to the number of compounds that were tested against them, leading to an imbalance in information content. This cannot be inferred from the calculated fraction. For example, the pairs EGFR-KDR and EGFR-p38a both have a profile similarity of $0.35$. However, the first was calculated based on $893$ tested compounds, whereas the latter on $52$ only.

The kinase distance matrix above will be reloaded in Talktorial T028, where we compare kinase similarities from different perspectives, including the ligand profile perspective we have talked about in this talktorial.

Reference

https://github.com/volkamerlab/teachopencadd

Reprint statement

Original title: Kinase similarity: Ligand profile

Authors:

Talia B. Kimber, 2021, Volkamer lab, Charité
Dominique Sydow, 2021, Volkamer lab, Charité
Andrea Volkamer, 2021, Volkamer lab, Charité