Identification of Protein–Protein Interaction (PPI) Sites on the Influenza A (H1N1) Viral Genome Using Gradient Boosting and Artificial Neural Network (ANN) Models

Identification of protein–protein interaction
(PPI) sites
is crucial for understanding molecular recognition. Experimental identification
of PPI is expensive, time-consuming, and laborious. A large number
of computational methods addressed this problem. However, no computations
specifically addressed PPI site prediction for the frequently mutating
influenza A virus (IAV) genome that invades human hosts. For the first
time, we report the prediction of PPI sites on the IAV genome (protein
sequences). The method was benchmarked across various machine-learning
models, optimizing class imbalance and unlabeled data types. The best-performing
models were (i) the gradient boosting model, augmented with minority
class oversampling and positive unlabeled (PU) learning and (ii) the
protein-specific bidirectional encoder representations from transformers
(Prot-BERT) combined with an artificial neural network (ANN) (termed
the Prot-BERT-ANN model), adjusted with class weight correction and
threshold tuning. The models were trained on two types of interaction
site data sets: one obtained from diverse protein families (Train-1)
(17995 amino acid sites) with known interaction sites from protein
structures and the other from the IAV consensus protein sequences
(3322 amino acid sites) with experimentally annotated PPI sites on
the conserved regions of the proteins (Train-2). External validation
was performed on two test data sets: (i) from six IAV proteins, M1,
NS1, NEP, NP, PB1, and PB2, reported to interact with host factors,
with experimentally annotated PPI sites on the nonconserved region
of the proteins (Test-1), and (ii) the SARS-CoV-2 spike protein sequence
(195 amino acid sites) (Test-2). Blind prediction was performed on
three IAV protein sequencesNA, HA, and M2curated from
the Human Viral Interaction Database (HVIDB). The prediction aimed
to decipher the effect of amino acid substitutions on the protein–protein
interaction sites of the viral genome. The gradient boosting method
with oversampling and PU learning, trained on the Train-2 data set,
consistently performed better on both external validation data sets.
The recall values obtained from the predictions on the Test-1 data
set were compared with the published D-SCRIPT (a neural language-based
model) results. The gradient boosting model showed a higher average
recall value (0.53 ± 0.04) for six IAV proteins compared to the
D-SCRIPT results (0.18 ± 0.19). The gradient boosting prediction
for the experimentally reported PPI sites on the SARS-CoV-2 spike
protein (Test-2 data set) was 55% accurate, despite Test-2 being independent
of Train-2. The results indicated the generalizability and interpretability
of the gradient boosting model for IAV PPI site predictions. The effects
of amino acid substitutions on PPI sites were demonstrated on five
Matrix 1 (M1) protein sequences. This approach could be used to identify
the PPI sites on newly emerging viral strains (e.g., influenza virus,
SARS-CoV-2, etc.) with potential applications for drug design, improvement
of drug binding, or drug repurposing, subject to further validation.