Major Histocompatibility Complex (MHC) alleles and peptides are sequences of amino acids. Peptides are shorter units of proteins, and MHC alleles are proteins essential for the immune system’s adaptivity as they help it recognize foreign substances.
According to the WHO, cancer is the second leading cause of death worldwide. To help this situation, research is now focused on personalized cancer vaccines. As it is difficult for the immune system to distinguish between a healthy and a cancerous cell, personalized cancer vaccines intend to externally synthesize a peptide that, when passed into the patient, helps the immune system identify the cancerous cells. This identification is made by forming a bond between the injected peptide and cancerous cells in the body. A proper analysis is required to choose the right peptide that can trigger an appropriate immune response.
Thus, one of the major steps in synthesizing personalized vaccines is computationally predicting whether the given peptide will bind with the patient’s MHC allele.
Major barriers in creating personalized cancer vaccines are a lack of understanding among the scientific community about how exactly the peptide binding takes place and the need for clinical tests of different molecules.
Introduction to MHCAttnNet
A recent bioinformatics research paper, (Menaka Rajapakse et al) introduces MHCAttnNet. MHCAttnNet is a deep learning model that uses Bi-LSTMs (Bi-directional long short-term memory) to predict the MHC-peptide bindings more accurately than existing methods. It uses an attention-based deep neural model.
This research is one of the first working towards understanding the underlying binding mechanism to make a prediction. This model is unique because it predicts the binding more accurately as well as highlights the subsequence of amino-acids that is likely to be important in making a prediction.
How it works:
MHCAttnNet uses the attention mechanism, a technique to highlight the important subsequence of the amino-acid sequence of the peptides and MHC alleles used to make the binding prediction. It focuses on the frequency of a particular subsequence of the allele getting highlighted with a particular amino-acid of the peptide. This helps in understanding the relationship between the peptide and allele subsequence. In an experiment, the computational model has predicted the number of trigrams of amino-acids of the MHC allele potentially suitable for predicting the binding, corresponding to an amino-acid of a (given) peptide, is reasonably around 3% of the total possible trigrams. So, this helps in “sequence reduction” which further reduces the work and expense of the vaccines’ clinical trials.
MHCAttnNet is a deep learning model that uses attention mechanisms to highlight the most likely subsequence of the peptides and MHC alleles, thus, predicting the MHC-Peptide bindings. This research will help in the reduction of manual work and the cost of clinical testing.