Gnina is a method of featurisation of 3D protein-ligand complexes [1] for input into convolutional neural networks. This repo is a collection of machine learning algorithms built on top of gnina. It can be used as is, but for testing to work properly a local installation is required:

cd gnina_tensorflow
pip install -e .
python3 -m pytest

The following are required for all or part of gnina_tensorflow:

libmolgrid (https://github.com/gnina/libmolgrid)
scikit-learn (https://github.com/scikit-learn/scikit-learn)
joblib (https://github.com/joblib/joblib)
tensorflow 2 (https://github.com/tensorflow/tensorflow)

Please note: 3D convolutions on the data format provided by gnina do not work when running TensorFlow on a CPU. Hopefully this will change in a future release of TensorFlow.

Previous work using gnina by this author can be found in Ref. [3]

There are two virtual screening architectures: the original implementation [1], and DenseFS [2]. There is a random forest, is trained for the same task but uses low dimensional encodings of gnina grids as input (mainly used as a diagnostic for a paper which is in the works). These encodings are generated by an autoencoder, which aims to reduce the dimensionality of the original gnina input.

Virtual screening is tasked with discriminating between active compounds and decoy compounds - that is, molecules that bind vs do not bind to a given protein target. An example of how to train a model on the small training set provided is shown below:

cd gnina_tensorflow
python3 classifier/gnina_tensorflow.py --data_root data --train data/small_chembl_test.types --batch_size 16 --iterations 100 --save_path classifier_example --densefs --inference_on_training_set

An autoencoder aims to learn a meaningful low-dimensional representation of a high-dimensional input. In this instance, gnina inputs which are similar in some sense (perhaps the proteins are the same and the ligands only differ by the location of a functional group on an aromatic ring) should be converted to encodings (vectors) which are a low distance apart; the inverse should hold true for dissimilar inputs.

Dimensionality reduction is an important part of my ongoing work with gnina; an example is show below:

cd gnina_tensorflow
python3 autoencoder/gnina_autoencoder.py --data_root data --train data/small_chembl_test.types --batch_size 1 --iterations 1000 --save_path autoencoder_example --encoding_size 200 --optimiser adamax --learning_rate 0.001 --final_activation sigmoid --dimension 18.0 --resolution 1.0 --save_encodings

Random forests are used to test the quality of the encodings generated by the autoencoder. The autoencoder is first trained on, for example, the DUD-E dataset [4]; the resulting encodings are used to train a random forest to discriminate between actives and decoys. The same trained autoencoder is then used to generate encodings for a test set - like the ChEMBL validation set from Ref. [3] - which are used to validate the model. If good performance is achieved on the test set, we can deduce that the encodings are good representations of the original inputs:

cd gnina_tensorflow
python3 classifier/random_forest.py rf_example --train_dir data/small_chembl_test_encodings

[1] M Ragoza, J Hochuli, E Idrobo, J Sunseri, DR Koes. (2017). Protein–Ligand Scoring with Convolutional Neural Networks J. Chem. Inf. Model. 57, 4, 942–957

[2] Imrie, F.; Bradley, A. R.; van der Schaar, M.; Deane, C. M. (2018). Protein Family-Specific Models Using Deep Neural Networks and Transfer Learning Improve Virtual Screening and Highlight the Need for More Data J. Chem. Inf. Model. 58, 2319−2330.

[3] Scantlebury, J.; Brown, N.; von Delft, F.; Deane, C. M. (2020). Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions J. Chem. Inf. Model. 60, 3722-3730.

[4] Mysinger, M. M.; Carchia, M.; Irwin, J. J.; Shoichet, B. K. (2012). Directory of Useful Decoys, Enhanced (DUD-E): Better Ligands and Decoys for Better Benchmarking J. Med. Chem. 2012, 55, 14, 6582–6594.