Compact Assessment of Molecular Surface Complementarities Enhances Neural Network-Aided Prediction of Key Binding Residues

Journal of Chemical Information and Modeling, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 21, 2025

Predicting interactions between proteins is fundamental for understanding the mechanisms underlying cellular processes, since protein–protein complexes are crucial in physiological conditions but also many diseases, example by seeding aggregates formation. Despite advancements made so far, performance of docking protocols deeply dependent on their capability to identify binding regions. From this, importance developing low-cost and computationally efficient methods this field. We present an integrated novel protocol mainly based compact modeling protein surface patches via sets orthogonal polynomials regions high shape/electrostatic complementarity. By incorporating both hydrophilic hydrophobic contributions, we define new matrices, which serve as effective inputs training a neural network. In work, propose Neural Network (NN)-based architecture, Core Interacting Residues (CIRNet), achieves terms Area Under Receiver Operating Characteristic Curve (ROC AUC) approximately 0.87 identifying pairs core interacting residues balanced data set. blind search residues, CIRNet distinguishes them from random decoys with ROC AUC 0.72. test enhance algorithms filtering proposed poses, addressing one still open problems computational biology. Notably, when applied top ten models three widely used servers, improves outcomes, significantly reducing average RMSD selected poses native state. Compared another state-of-the-art tool rescaling more efficiently identified worst generated servers under consideration achieved superior two cases.

Language: Английский

---

Deep learning in GPCR drug discovery: benchmarking the path to accurate peptide binding

Briefings in Bioinformatics, Journal Year: 2025, Volume and Issue: 26(2)

Published: March 1, 2025

Abstract Deep learning (DL) methods have drastically advanced structure-based drug discovery by directly predicting protein structures from sequences. Recently, these become increasingly accurate in complexes formed multiple chains. We evaluated advancements to predict and accurately model the largest receptor family its cognate peptide hormones. benchmarked DL tools, including AlphaFold 2.3 (AF2), 3 (AF3), Chai-1, NeuralPLexer, RoseTTAFold-AllAtom, Peptriever, ESMFold, D-SCRIPT, interactions between G protein-coupled receptors (GPCRs) their endogenous ligands. Our results showed that structure-aware models outperformed language binding classification, with top-performing achieving an area under curve of 0.86 on a benchmark set 124 ligands 1240 decoys. Rescoring predicted local further improved principal ligand among decoy peptides, whereas DL-based approaches did not. explored competitive tournament approach for modeling peptides simultaneously single GPCR, which accelerates performance but reduces true-positive recovery. When evaluating poses 67 recent complexes, AF2 reproduced correct modes nearly all cases (94%), surpassing those both AF3 Chai-1. Confidence scores correlate structural mode accuracy, provides guide interpreting interface predictions. These demonstrated can reliably rediscover binders, aid discovery, selection optimal tools GPCR-targeted therapies. To this end, we provided practical selecting best specific applications independent benchmarking future evaluation.

Language: Английский

---

Protein A-like Peptide Design Based on Diffusion and ESM2 Models

Molecules, Journal Year: 2024, Volume and Issue: 29(20), P. 4965 - 4965

Published: Oct. 21, 2024

Proteins are the foundation of life, and designing functional proteins remains a key challenge in biotechnology. Before development AlphaFold2, focus design was primarily on structure-centric approaches such as using well-known open-source software Rosetta3. Following deep-learning techniques for protein gained prominence. This study proposes new method to generate diffusion model ESM2 language model. Diffusion models, which widely used image natural generation, here design, facilitating controlled generation sequences. The model, trained basis large-scale sequence data, provides deep understanding context sequence, thus improving model's ability biologically relevant proteins. In this study, we Protein A-like peptide object, combined sequences from minimal input verified their biological activities through experiments BLI affinity test. conclusion, developed that novel strategy meet challenges generic generation.

Language: Английский

---

EuDockScore: Euclidean graph neural networks for scoring protein-protein interfaces

Bioinformatics, Journal Year: 2024, Volume and Issue: 40(11)

Published: Oct. 21, 2024

Abstract Motivation Protein–protein interactions are essential for a variety of biological phenomena including mediating biochemical reactions, cell signaling, and the immune response. Proteins seek to form interfaces which reduce overall system energy. Although determination single polypeptide chain protein structures has been revolutionized by deep learning techniques, complex prediction still not perfected. Additionally, experimentally determining is incredibly resource time expensive. An alternative technique computational docking, takes solved individual proteins produce candidate (decoys). Decoys then scored using mathematical function that assess quality system, known as scoring functions. Beyond functions critical component assessing produced many generative models. Scoring models also used final filtering in those generate antibody binders, perform docking. Results In this work, we present improved protein–protein utilizes cutting-edge Euclidean graph neural network architectures, interfaces. These docking score EuDockScore, EuDockScore-Ab with latter being antibody–antigen dock specific. Finally, provided EuDockScore-AFM model trained on outputs from AlphaFold-Multimer (AFM) proves useful reranking large numbers AFM outputs. Availability implementation The code these available at https://gitlab.com/mcfeemat/eudockscore.

Language: Английский

---

Comprehensive Evaluation of AlphaFold-Multimer, AlphaFold3 and ColabFold, and Scoring Functions in Predicting Protein-Peptide Complex Structures

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 11, 2024

Abstract Determining the three-dimensional structures of protein-peptide complexes is crucial for elucidating biological processes and designing peptide-based drugs. Protein-peptide docking has become essential predicting complex structures. AlphaFold-Multimer, ColabFold AlphaFold3 provided groundbreaking tools to enhance accuracy. This study evaluates these three using Template-Based (TB) Template-Free (TF) methods. AlphaFold-Multimer excels in TB predictions performs moderately TF scenarios prediction pool, but outperforms first-ranked models. demonstrates versatility both settings. generates high-quality more proteins, medium accuracy not as good a large model pool. We also assessed performance various scoring functions ranking predicted While function built AlphaFold best performance, some other functions, e.g., FoldX-Stability HADDOCK-mdscore, provide complementary values. The findings suggest potential enhancing targeting AlphaFold-based by combining multiple or consensus approach from many

Language: Английский

---

ProAffinity-GNN: A Novel Approach to Structure-based Protein-Protein Binding Affinity Prediction via a Curated Dataset and Graph Neural Networks

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: March 15, 2024

Abstract Protein-protein interactions (PPIs) are crucial for understanding biological processes and disease mechanisms, contributing significantly to advances in protein engineering drug discovery. The accurate determination of binding affinities, essential decoding PPIs, faces challenges due the substantial time financial costs involved experimental theoretical methods. This situation underscores urgent need more effective precise methodologies predicting affinity. Despite abundance research on PPI modeling, field quantitative affinity prediction remains underexplored, mainly a lack comprehensive data. study seeks address these needs by manually curating pairwise interaction labels all available 3D structures proteins complexes, with experimentally determined creating largest dataset structure-based date. Subsequently, we introduce “ProAffinity-GNN”, novel deep learning framework using language model graph neural network (GNN) improve accuracy protein-protein affinities. evaluation results across several benchmark test sets demonstrate that ProAffinity-GNN not only outperforms existing models terms but also shows strong generalization capabilities.

Language: Английский

---

PhosHSGN: Deep Neural Networks Combining Sequence and Protein Spatial Information to Improve Protein Phosphorylation Site Prediction

IEEE Access, Journal Year: 2024, Volume and Issue: 12, P. 100611 - 100627

Published: Jan. 1, 2024

Phosphorylation site prediction is one of the key processes in protein post-transcriptional modification. It an important research direction field bioinformatics and great significance for understanding function signal transduction. Since it time-consuming error-prone to perform determination through experiments, application artificial intelligence very necessary. This study introduces a novel deep neural network named PhosHSGN designed identify examine post-translational modification (PTM) sites. The model predicts phosphorylation by extracting local sequence incorporating global spatial information. To effectively combine information prediction, graph introduced with residuals. integrates Alphafold structure module construct residue contact graph. Additionally, pre-trained language employed generate base extraction embeddings. Simultaneously, incorporates one-dimensional residual explore proteins. Experimental data were collected from PhosphoSitePlus, UniProt, GPS 5.0, Phospho.ELM. Comparing experimental results Phosidn other state-of-the-art models on different datasets reveals that outperforms sequence-based methods all metrics sensitivity 96.18%, accuracy 93.72%, Mcc value 84.19% dataset S/T. On Y, F1 score was 94.42% AUC 96.19%.

Language: Английский

---

ProAffinity-GNN: A Novel Approach to Structure-Based Protein–Protein Binding Affinity Prediction via a Curated Data Set and Graph Neural Networks

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 18, 2024

Protein-protein interactions (PPIs) are crucial for understanding biological processes and disease mechanisms, contributing significantly to advances in protein engineering drug discovery. The accurate determination of binding affinities, essential decoding PPIs, faces challenges due the substantial time financial costs involved experimental theoretical methods. This situation underscores urgent need more effective precise methodologies predicting affinity. Despite abundance research on PPI modeling, field quantitative affinity prediction remains underexplored, mainly a lack comprehensive data. study seeks address these needs by manually curating pairwise interaction labels available 3D structures complexes, with experimentally determined creating largest data set structure-based date. Subsequently, we introduce ProAffinity-GNN, novel deep learning framework using language model graph neural network (GNN) improve accuracy protein-protein affinities. evaluation results across several benchmark test sets an additional case demonstrate that ProAffinity-GNN not only outperforms existing models terms but also shows strong generalization capabilities.

Language: Английский

---