Despite multiple efforts over the last three decades, an effective vaccine against the human immunodeficiency virus (HIV) is still not available. One of the major challenges for vaccine design is that HIV mutates and replicates at a high rate, with the resulting diversity enabling it to escape host immune responses. An ideal vaccine would elicit antibodies that target parts of the viral proteins where mutations severely compromise the virus’ fitness. In order to guide such vaccine design, a systematic characterization of the fitness landscape (a mapping from the viral strain sequence to its fitness or viability) is required.

Experimentally determining the complex fitness landscape is infeasible (due to the prohibitively large number of experiments), and computational approaches based on the statistical analysis of available viral sequence data are emerging as alternative strategies. A promising approach is to use unsupervised machine learning algorithms to infer a probability model from publicly-available viral sequences obtained from infected patients.

The goal of this project is to investigate the use of restricted Boltzmann machines (RBM), a simple form of feedforward neural network (FNN), to infer a probability model for HIV protein sequences, which could ultimately lead to new vaccine strategies for HIV.

We will adopt a systematic approach where we will: (i) consider a simplified toy sequence model and develop an RBM learning (inference) framework; (ii) Test the inference framework with synthetically generated data, sampled from a predefined ground truth model; (iii) make the necessary adaptations in order to apply our learning framework to HIV sequence data (obtained from a publicly-available database); (iv) Evaluate the accuracy of the learned model (fitness landscape) with available fitness experimental data.

The project is best suited to students with an interest in data science (big data), neural networks, machine learning and their applications. A solid background in probability and statistics, as well as MATLAB programming skills, are desirable.