Protein Secondary Structure Prediction

Note: This repository and its contents support the coursework of the INM701 module at City, University of London.

The output of the code for $k$-nearest neighbours and random forests is reproducible using the seed set in each notebook (which is 50). Unfortunately, we required a GPU in order to train the neural network models, which requires optimising millions of parameters. It is difficult to achieve reproducible results through a GPU, and so we use the 'tf.config.experimental.enable_op_determinism()' command from TensorFlow (tf). This command is experimental, and does not seem to work on older GPUs. An alternative approach would be repeatedly training models a number of times, and calculating the mean (and possibly the standard error) of the accuracies obtained.

Library Versions: numpy == 1.24.3, sklearn == 1.2.2, tensorflow == 2.13.0

GPU for TensorFlow: NVIDIA P100 16GB VRAM (Kaggle GPU)

Background

Protein structure prediction is an important area of bioinformatics and more specifically structural biology. By understanding the 3-D structure of a protein from its amino acid sequence, researchers and scientists can understand the biological function of proteins and how proteins can interact with one another. In recent years, powerful tools such as DeepMind's AlphaFold have been able to predict over 200 million protein structures to tackle the protein folding problem. We investigate the use of simple machine learning algorithms to predict the secondary structure of proteins, and how these compare in performance against each other. We also determine how oversampling can affect the predictive performance of our algorithms.

Specifically, we look at using (for classification) the:

• Non-parametric $k$-nearest neighbours
• Non-parametric random forests
• Parametric neural networks

Dataset

The dataset used is a subset of the RCSB Protein Data Bank, obtained jointly using various programs by @alfrandom and @kirkdco. The dataset can be found here. After some processing, we filtered the dataset by selecting proteins that had length in range [16, 100]. The filtered dataset containing the protein sequences for this study can be found here.

$k$-nearest neighbours

$k$-nearest neighbours ($k$-NN) is a non-parametric machine learning algorithm, which is also a type of lazy learning algorithm. The algorithm classifies data by assigning labels to observations based on the majority class amongst their $k$ closest observations. We construct models based on $k$-NN, and the related work can be found in the k-nearest-neighbours folder.

Random Forests

Random forests is a non-parametric machine learning algorithm and a type of ensemble method. The random forests algorithm works by constructing a multitude of decision trees during the training phase. We construct models based on random forests, and the related work can be found in the random-forests folder.

Neural Networks

Artificial neural networks are parametric machine learning algorithms that try to mimic the way information is processed in the brain. The idea is to create models of biological neurons that have connections between nodes, with weights attached. These weights are to be optimised through the use of methods such as gradient descent with backpropagation. We construct models based on neural networks, and the related work can be found in the neural-networks folder.