Modeling the sequence dependence of differential antibody binding in the immune response to infectious disease

Abstract

Past studies have shown that incubation of human serum samples on high density peptide arrays followed by measurement of total antibody bound to each peptide sequence allows detection and discrimination of humoral immune responses to a variety of infectious diseases. This is true even though these arrays consist of peptides with near-random amino acid sequences that were not designed to mimic biological antigens. This “immunosignature” approach, is based on a statistical evaluation of the binding pattern for each sample but it ignores the information contained in the amino acid sequences that the antibodies are binding to. Here, similar array-based antibody profiles are instead used to train a neural network to model the sequence dependence of molecular recognition involved in the immune response of each sample. The binding profiles used resulted from incubating serum from 5 infectious disease cohorts (Hepatitis B and C, Dengue Fever, West Nile Virus and Chagas disease) and an uninfected cohort with 122,926 peptide sequences on an array. These sequences were selected quasi-randomly to represent an even but sparse sample of the entire possible combinatorial sequence space (~1012). This very sparse sampling of combinatorial sequence space was sufficient to capture a statistically accurate representation of the humoral immune response across the entire space. Processing array data using the neural network not only captures the disease-specific sequence-binding information but aggregates binding information with respect to sequence, removing sequence-independent noise and improving the accuracy of array-based classification of disease compared with the raw binding data. Because the neural network model is trained on all samples simultaneously, a highly condensed representation of the differential information between samples resides in the output layer of the model, and the column vectors from this layer can be used to represent each sample for classification or unsupervised clustering applications.

Predicting Monoclonal Antibody Binding Sequences from a Sparse Sampling of All Possible Sequences

“…To build a general rela�onship between sequence and mAb binding from the array data, a fully connected neural network was employed with 2 hidden layers and 250 nodes per layer. This was similar in structure to networks used previously for describing serum IgG binding 19 or binding to isolated proteins 18 . Two addi�ons were made to the sequence representa�on beyond what has been done previously.…”

“…Also, the N-and C-termini were marked with start and end tokens "(" and ")", respec�vely. The sequences, including the three addi�onal characters, were then input into the neural network as one-hot encoded vectors, similar to previous work 18,19 . Note that only the pep�des between 5 and 11 residues in length were used in the training (there were 121715 unique pep�des in this length range) because the focus was on iden�fying con�nuous epitopes which are generally in that length range and because there is a tendency for the neural network to overes�mate the binding of longer pep�des, simply due to the larger number of elements to assign value to, poten�ally biasing the model.…”

“…Further, any bias that may exist in amino acid composi�on of a par�cular posi�on or any effects from the fact that the sequences are not all the same length are effec�vely washed out. This was clearly the most significant means of improving the predic�ve power of the model, and it is an approach that can easily be applied to other situa�ons, such as whole sera for epitope predic�ons associated with an immune response to disease 19 .…”

“…Previous work from this lab has demonstrated that it is possible to use binding of target samples to high density arrays in order to generate sequence-binding rela�onships capable of predic�ng with high sta�s�cal accuracy the binding values of sequences that are not present in the training set. This has been done both for binding of purified proteins 18 and for binding of serum samples from cohorts of individuals with 5 different infec�ous diseases or uninfected individuals 19 . In the later case, the sequence-binding rela�onship not only maintained the ability to accurately classify disease cohorts, it was also able to remove sequence independent noise from the measured binding data and improve the ability to classify cohorts.…”