On deep-learning tools used in protein-peptide residues identification...and on how AI approaches permeate structural bioinformatics

 

Using CNNs to predict peptide-protein binding interfaces:

PepCNN deep learning tool for predicting peptide binding residues in proteins using sequence, structural, and language model features

 I'd like to explore an article found regarding the use of deep learning tools used in protein structural bioinformatics. The challenge of determining, which residues from a protein would bind to a certain ligand, is in a way a key challenge in research. This is, partly, because the number of proteins as well as the number of ligands is endless, but obtaining these results via experiments proves expensive. Or, the effect of peptides as drugs resides ultimatively in their ability to bind certain proteins, therefore knowing what binds to what and how does it happen is of utmost importance.

The original article could be found on Link, here being presented a short summary using good old ChatGPT assisted text editing. While some of the points may be left out due to excessive details, I'd like the readers to see, that the tools provided by Deep Learning, especially CNN, prove quite useful in tackling structural bionformatics, especially there where big data is involved. Following, between "...", some insides from the article:

...

 In the abstract, the author discusses the crucial role of protein–peptide interactions in cellular processes and their connection to diseases like cancer. They highlight the challenges of experimental methods in studying these interactions and introduce PepCNN, a new deep learning-based prediction model. PepCNN utilizes structural and sequence-based information, outperforming existing methods in terms of accuracy. The software and datasets for PepCNN are available publicly at https://github.com/abelavit/PepCNN.git, offering a promising tool for researchers in genomics and drug discovery.

Introduction highlights the critical role of protein–peptide interactions in cellular functions and their implication in diseases. Traditional experimental approaches face limitations, leading to the introduction of computational methods, but they also have challenges. Two main categories of computational methods exist: structure-based and sequence-based. The former, like PepSite and SPRINT-Str, focus on structural attributes, while the latter, like PepBind and PepNN-Seq, use machine learning with features such as amino acid sequences. Recent advancements include deep learning technologies like Convolutional Neural Network (CNN) and Transformer-based models, such as Visual and PepNN-Seq, which have shown promise in predicting protein–peptide interactions. These models leverage image-like representations and sequence embeddings, showcasing the evolving landscape of computational proteomics.

There is discussed the potential of deep learning algorithms in addressing complex challenges in protein science and structural biology. Deep learning models, inspired by human cognitive processes, have shown superiority over traditional machine learning frameworks. In the context of proteomics, these algorithms, particularly Convolutional Neural Networks (CNNs), are effective in handling large and complex data, including protein structures. The integration of pre-trained contextualized language models designed for protein biology has further enriched computational tools in the field.

The article introduces PepCNN, an innovative model that combines protein sequence embeddings from a protein language model with CNN. This approach, utilizing sequence-based features from ProtT5-XL-UniRef50 and traditional features like Position Specific Scoring Matrices (PSSMs) and structure-based attributes, sets a new benchmark in predictive performance. PepCNN outperforms existing methods, including PepBCL, PepNN-Seq, PepBind, SPRINT-Str, and SPPPred, promising advancements in drug discovery, disease mechanism understanding, and bioinformatics computational approaches.

The proposed work for predicting binding and non-binding residues includes feature extraction for proteins, residue-specific feature extraction, CNN model training using 80% of the training set, and model evaluation on the remaining 20% for validation. The flow diagram, created with Inkscape software, illustrates these sequential steps in the prediction process.

3D structure visualization of three proteins (pdbID: 1dpuA, pdbID: 2bugA, and pdbID: 1uj0A) illustrating the binding (in magenta) and non-binding (in gray) residues using the PyMol software⁴². The experimental output (true binding residues) of the proteins are located in the top part (A, C, and E) and its corresponding predicted binding residues by our method PepCNN are located in the bottom part (B, D, and F).

...

While not yet close to the experiment, results could be viewed as promissing. Noteworthy, the region where a peptide would bind is identified right, i.e. these examples show that the predicted binding site would superpose with the experimental site. Moreover, while not completely overlapping, the Exp. and Teor. do share some resemblance, therefore one might argue that if not exactly predicting the right residues used in binding, the tool at hand could already limit the search-space.

While these tools only appear, it is important to say that it's expected for the performance and accuracy to increase. Especially, when these tools are to be developed for different purposes, such as protein-protein and ligand-ligand interactions. It could also be assumed, that networks containing different joined architectures could prove even better at finding binding residues.