# Creating a high fidelity deepchem dataset from experimental data
# Creating a high fidelity DeepChem dataset from experimental data
%% Cell type:markdown id: tags:
Suppose you were given data collected by an experimental collaborator. You would like to use this data to construct a deepchem model.
Suppose you were given data collected by an experimental collaborator. You would like to use this data to construct a DeepChem model.
*How do you transform this data into a dataset capable of creating a useful deepchem model?*
*How do you transform this data into a dataset capable of creating a useful DeepChem model?*
%% Cell type:markdown id: tags:
Building models from novel data can present several challenges. Perhaps the data was not recorded in a convenient manner. Additionally, perhaps the data contains noise. This is a common occurance with, for example, biological assays due to the large number of external variables and the difficulty and cost associated with collecting multiple samples. This is a problem because you do not want your model to fit to this noise.
Hence, there are two primary challenges:
* Parsing data
* De-noising data
In this tutorial, will walk through an example of curating a dataset from an excel spreadsheet of experimental drug measurements.
%% Cell type:markdown id: tags:
## Parsing data
%% Cell type:markdown id: tags:
In order to read in the data, we will use the pandas data analysis library.
In order to convert the drug names into smiles strings, we will use pubchempy.
%% Cell type:code id: tags:
``` python
importos
importpandasaspd
frompubchempyimportget_cids,get_compounds
```
%% Cell type:markdown id: tags:
Pandas is magic but it doesn't automatically know where to find your data of interest. You likely will have to look at it first using a GUI.
We will now look at a screenshot of this dataset as rendered by LibreOffice.
We see the data of interest is on the second sheet, and contained in columns "TA ID", "N #1 (%)", and "N #2 (%)".
Additionally, it appears much of this spreadsheet was formatted for human readability (multicolumn headers, column labels with spaces and symbols, etc.). This makes the creation of a neat dataframe object harder. For this reason we will cut everything that is unnecesary or inconvenient.
However, some of these drug names have variables spaces and symbols (·, (±), etc.), and names that may not be readable by pubchempy.
For this task, we will do a bit of hacking via regular expressions. Also, we notice that all ions are written in a shortened form that will need to be expanded. For this reason we use a dictionary, mapping the shortened ion names to versions recognizable to pubchempy.
Unfortunately you may have several corner cases that will require more hacking.
Hooray, we have mapped each drug name to its corresponding smiles code.
Now, we need to look at the data and remove as much noise as possible.
%% Cell type:markdown id: tags:
## De-noising data
%% Cell type:markdown id: tags:
In machine learning, we know that there is no free lunch. You will need to spend time analyzing and understanding your data in order to frame your problem and determine the appropriate model framework. Treatment of your data will depend on the conclusions you gather from this process.
Questions to ask yourself:
* What are you trying to accomplish?
* What is your assay?
* What is the structure of the data?
* Does the data make sense?
* What has been tried previously?
For this project (respectively):
* I would like to build a model capable of predicting the affinity of an arbitrary small molecule drug to a particular ion channel protein
* For an input drug, data describing channel inhibition
* A few hundred drugs, with n=2
* Will need to look more closely at the dataset*
* Nothing on this particular protein
%% Cell type:markdown id: tags:
*This will involve plotting, so we will import matplotlib and seaborn. We will also need to look at molecular structures, so we will import rdkit.
%% Cell type:code id: tags:
``` python
importmatplotlib.pyplotasplt
%matplotlibinline
importseabornassns
sns.set_style('white')
```
%% Cell type:code id: tags:
``` python
fromrdkitimportChem
fromrdkit.ChemimportAllChem
fromrdkit.ChemimportDraw,PyMol,rdFMCS
fromrdkit.Chem.DrawimportIPythonConsole
fromrdkitimportrdBase
```
%% Cell type:code id: tags:
``` python
# i will use numpy on occasion for manipulating arrays
importnumpyasnp
```
%% Cell type:markdown id: tags:
Our goal is to build a small molecule model, so let's make sure our molecules are all small. This can be approximated by the length of each smiles string.
<PIL.Image.Image image mode=RGBA size=1200x200 at 0x7F7C65299E90>
%% Cell type:markdown id: tags:
As suspected, these are not small molecules, so we will remove them from the dataset. The argument here is that these molecules could register as inhibitors simply because they are large. They are more likely to sterically blocks the channel, rather than diffuse inside and bind (which is what we are interested in).
The lesson here is to remove data that does not fit your use case.
Then, let's examine the distribution of n1 and n2.
%% Cell type:code id: tags:
``` python
df=smiles_data.dropna(axis=0,how='any')
```
%% Cell type:code id: tags:
``` python
# seaborn jointplot will allow us to compare n1 and n2, and plot each marginal
sns.jointplot('n1','n2',data=smiles_data)
```
%% Output
<seaborn.axisgrid.JointGrid at 0x7f7c6187c3d0>
%% Cell type:markdown id: tags:
We see that most of the data is contained in the gaussian-ish blob centered a bit below zero. We see that there are a few clearly active datapoints located in the bottom left, and one on the top right. These are all distinguished from the majority of the data. How do we handle the data in the blob? I suspect that they are not active, and just noise, but let's be sure.
We see that most of the data is contained in the gaussian-ish blob centered a bit below zero. We see that there are a few clearly active datapoints located in the bottom left, and one on the top right. These are all distinguished from the majority of the data. How do we handle the data in the blob?
Because n1 and n2 represent the same measurement, ideally they would be of the same value. This plot should be tightly aligned to the diagonal, and the pearson correlation coefficient would be 1. We see this is not the case. This helps gives us an idea of the error of our assay.
Because n1 and n2 represent the same measurement, ideally they would be of the same value. This plot should be tightly aligned to the diagonal, and the pearson correlation coefficient should be 1. We see this is not the case. This helps gives us an idea of the error of our assay.
Let's look at the error more closely, plotting in the distribution of (n1-n2).
%% Cell type:code id: tags:
``` python
diff_df=df['n1']-df['n2']
sns.distplot(diff_df)
plt.xlabel('difference in n')
plt.ylabel('probability')
```
%% Output
<matplotlib.text.Text at 0x7f7c5b951150>
%% Cell type:markdown id: tags:
This looks pretty gaussian, let's get the 95% confidence interval by fitting a gaussian via scipy, and taking 2*the standard deviation
%% Cell type:code id: tags:
``` python
fromscipyimportstats
```
%% Cell type:code id: tags:
``` python
mean,std=stats.norm.fit(diff_df)
```
%% Cell type:code id: tags:
``` python
ci_95=std*2
ci_95
```
%% Output
17.803505462981775
%% Cell type:markdown id: tags:
Now, I don't trust the data outside of the confidence interval, and will therefore drop these datapoints from df.
For example, in the plot above, at least one datapoint has n1-n2 > 60. This is disconcerting.
%% Cell type:code id: tags:
``` python
noisy=diff_df[abs(diff_df)>ci_95]
df=df.drop(noisy.index)
```
%% Cell type:code id: tags:
``` python
sns.jointplot('n1','n2',data=df)
```
%% Output
<seaborn.axisgrid.JointGrid at 0x7f7c5b7bbb10>
%% Cell type:markdown id: tags:
Now that data looks much better!
So, let's average n1 and n2, and take the error bar to be ci_95.
In my case, this required domain knowledge. Having worked in this area, and having consulted with professors specializing on this channel, I am interested in compounds where the absolute value of the activity is greater than 25. This relates to the desired drug potency we would like to model.
If you are not certain how to draw the line between active and inactive, this cutoff could potentially be treated as a hyperparameter.
* Removed data that did not address the question we hope to answer (small molecules only)
* Dropped NaNs
* Determined the noise of our measurements
* Removed exceptionally noisy datapoints
* Identified actives (using domain knowledge to determine a threshold)
%% Cell type:markdown id: tags:
## Determine model type, final form of dataset, and sanity load
%% Cell type:markdown id: tags:
Now, what model framework should we use?
Given that we have 392 datapoints and 6 actives, this data will be used to build a low data one-shot classifier (10.1021/acscentsci.6b00367). If there were datasets of similar character, transfer learning could potentially be used, but this is not the case at the moment.
Let's apply logic to our dataframe in order to cast it into a binary format, suitable for classification.
Lastly, it is often advantageous to numerically transform the data in some way. For example, sometimes it is useful to normalize the data, or to zero the mean. This depends in the task at hand.
Built into Deepchem are many useful transformers, located in the deepchem.transformers.transformers base class.
Built into DeepChem are many useful transformers, located in the deepchem.transformers.transformers base class.
Because this is a classification model, and the number of actives is low, I will apply a balancing transformer. I treated this transformer as a hyperparameter when I began training models. It proved to unambiguously improve model performance.
