Goals

Please use the provided subset of this sequence->fitness dataset (representing 0.3% of the possible sequence diversity) to:

• construct 10 new sequences outside of the provided dataset with the highest predicted fitness for a subsequent round of synthesis.

• Suggest 10 compounds outside of the dataset provided that if you had experimental data on would most improve your model (and thus answer for #1).

Usage

pip install -r requirements.txt
cp "path/to/DataSet for Assignment.xlsx - Sheet1 (1).csv" .
python main.py

This takes approximately 26h on a MacBook Pro with a 3.1 GHz i7 CPU and 16 GB 1867 MHz DDR3 memory to complete.

Methods

This implementation resembles that of innov’SAR.

Steps:

1) Encoding

a) AAIndex1 features

Via PyBioMed:

This database holds more than 500 numerical indices representing various physicochemical and biochemical properties for the 20 standard amino acids and correlations between these indices are also listed.

TODO: use correlations

b) Fast Fourier Transform (FFT)

Via numpy.fft.

2) Modelling

the modelling phase, will use the experimental values of the target activity, in conjunction with these protein spectra, in order to identify a predictive model. The model is constructed by the application of standard regression approaches based on a learning step and a validation step. innov’SAR used a partial least square regression, PLS, as algorithm of regression to do the model for the predictions of the enantioselectivity of epoxide hydrolase.

This implementation differs in the following ways:

• Encode the given sequences (having length 4 instead of 9)
• Predict the given fitness values (instead of enantioselectivity of epoxide hydrolase)

3) Validation

The root mean squared error (RMSE) and the coefficient of determination (R2) are the performance parameters to assess a regression model, during the validation step.

The cross validation procedure described in the paper is implemented here with sklearn.

Results

Once the optimal model is selected, it can be used to predict the fitness of all other possible sequences, and the top 10 with the highest fitness values can be synthesized for the next round.

10 compounds outside of the provided dataset whose experimental data would most improve the model are likely those that lie far away from the provided ones in the mutation graph. These can be determined by constructing an exhaustive mutation graph of the whole space and sampling from regions that are far away both from each other and from the provided dataset.

Future Work

• Training on more than 0.3% of the data would likely improve performance.

• Training more models (e.g. SVM, Random Forest) with various hyperparameter settings might yield a model with better accuracy.

• Creating a mutation graph of the given variants and training separate classifiers on connected components might have the effect of reducing the search space for each model, thereby improving performance. This could be implemented via a 4D boolean adjacency matrix and scipy.ndimage.measurements.find_objects.

• More features can be added, such as physical features from libraries such as PyBioMed, or deep learning features such as rawMSA.