ATPint: Identification of ATP Binding Residues of a Protein from Its Primary Sequence

Welcome to the official repository for ATPint, a web server for predicting ATP-interacting residues in proteins from their primary sequence using Support Vector Machine (SVM) models. This resource is designed to support researchers in protein function annotation, structural biology, and computational drug discovery.

Web Server: https://webs.iiitd.edu.in/raghava/atpint/

zenodo link ; https://doi.org/10.5281/zenodo.20120263

Citation

Chauhan, J. S., Mishra, N. K., & Raghava, G. P. S. (2009). Identification of ATP binding residues of a protein from its primary sequence. BMC Bioinformatics, 10, 434. https://doi.org/10.1186/1471-2105-10-434

About the Tool

ATPint is a web-based server for predicting residues in a protein that interact with the ATP ligand, using only the primary amino acid sequence as input. Adenosine-5'-triphosphate (ATP) is a critical coenzyme involved in membrane transport, muscle contraction, cellular motility, kinase-mediated phosphorylation, and regulation of metabolic processes. Identifying ATP-interacting residues experimentally is costly and time-consuming; ATPint provides a fast computational alternative.

The method was developed using:

• 168 non-redundant ATP-binding protein (ABP) chains from PDB
• SVM models trained on binary sequence patterns and PSI-BLAST PSSM profiles
• Seven physico-chemical property scales as additional features
• Five-fold cross-validation for all performance evaluations

Key Features

Dataset

• 360 ATP-binding protein chains extracted from the SuperSite encyclopedia
• 267 non-redundant PDB chains after CD-HIT filtering (≤40% pairwise identity)
• 168 final non-redundant ABP chains verified using Ligand Protein Contact (LPC) software
• 3,056 unique overlapping windows of length 17 from 3,082 ATP-interacting residues

SVM Model Performance Summary

Model	Threshold	Sensitivity (%)	Specificity (%)	Accuracy (%)	MCC	AUC
Single sequence (binary)	0.0	65.53	66.97	66.25	0.33	0.725
PSSM profile (evolutionary)	0.1	70.01	80.39	75.20	0.51	0.823
Physico-chemical properties	0.0	63.18	62.95	63.06	0.26	—

Window Size Optimization (Single Sequence Model)

Window Size	Accuracy (%)	MCC
7	62.73	0.25
9	64.33	0.29
13	64.11	0.28
17	66.25	0.33
23	66.08	0.32
25	66.79	0.34

Window size 17 selected as optimal — highest balance between sensitivity and specificity.

Key Amino Acid Composition Findings

Residues significantly enriched in ATP-interacting regions (p < 0.05):

• Gly (p = 0.00186) — critical for P-loop/phosphate-binding loop
• Arg (p = 0.00862), Lys (p = 0.00254), His (p = 0.07941) — positively charged residues preferred for ATP interaction
• Leu (p = 0.00511), Pro (p = 0.00011), Ala (p = 0.00049), Val (p = 0.02253) — non-polar/hydrophobic residues also significant

Overview

ATPint provides:

• Prediction of ATP-interacting residues from protein primary sequence
• Best performance using PSSM-based SVM model (MCC 0.51, AUC 0.823)
• User-selectable threshold (range: −1.0 to +1.0; default: 0.2)
• Graphical output displaying predicted ATP-interacting and non-interacting residues in distinct colors
• Web server built using CGI-Perl 5.8.4

Methods Summary

Dataset Preparation

ATP-binding protein chains were extracted from the SuperSite encyclopedia and filtered for redundancy using CD-HIT (≤40% identity cutoff). The LPC software was used to verify genuine ATP contacts and remove false positives, yielding a final set of 168 non-redundant ABP chains. Each protein residue was labeled as ATP-interacting (positive) or non-interacting (negative) based on LPC assignments.

Pattern Generation

Overlapping sliding windows of length 17 were generated from each protein chain. The central residue of each window determines its label (positive or negative). Terminal residues were padded with dummy "X" residues to generate M patterns from a sequence of length M. This yielded 3,056 unique positive (ATP-interacting) patterns.

SVM Model 1 — Binary Sequence Encoding

Each window of length N is represented as a vector of N × 21 binary units (20 amino acid types + 1 dummy "X"). Implemented using SVM_light with RBF kernel (best parameters: g=0.1, c=2, j=3). Achieves MCC 0.33 and AUC 0.725 at window size 17.

SVM Model 2 — Evolutionary Information (PSSM)

PSI-BLAST profiles (PSSM) were generated by searching each protein against SWISS-PROT with E-value ≤ 0.001. Each element of the 21 × M PSSM matrix (normalized before input) encodes amino acid conservation at each position. Best model parameters: g=0.01, c=4, j=1. Achieves MCC 0.51 and AUC 0.823 — a significant improvement over the binary sequence model.

SVM Model 3 — Physico-chemical Properties

Seven residue-level property scales were used as SVM input features:

• Hydrophobicity — Fauchère & Pliska scale
• Beta-sheet propensity — Chou & Fasman scale
• Polarity — Grantham R scale
• Solvation potential — Jones et al. scale
• Residue interface propensity — Jones & Thornton scale
• Net charge — Klein et al. scale
• Average accessible surface area — Janin et al. scale

All values normalized before input. Maximum MCC achieved: 0.26 — lower than the sequence-based model.

Evaluation — Why BLAST and Motif Methods Fail

• BLAST: Only 71 of 168 ABP chains showed any similarity to other chains; sensitivity on matched chains was 44% with 43.37% probability of correct prediction — unsuitable for this task.
• Motif search (FingerPRINTScan): Motifs found in only 54 of 168 chains; covered only ~11% of ATP-interacting residues; no common motif shared across all ABPs. Motif-based approaches are insufficient for comprehensive ATP residue prediction.

Improvements Over Existing Methods

• First computational method specifically designed to predict ATP-interacting residues (not just ATP-binding sites) from primary sequence alone
• Demonstrates for the first time that evolutionary information (PSSM) significantly improves ATP-interacting residue prediction — analogous to its benefit in secondary structure and RNA-binding site prediction
• Outperforms BLAST and motif-based approaches on the same dataset
• Provides residue-level predictions rather than protein-level or binding-site-level annotations
• Publicly accessible web server with graphical output and tunable threshold

Limitations

• Moderate accuracy (~75% with PSSM) reflects the inherent complexity of ATP–protein interaction prediction from sequence alone
• Dataset is relatively small (168 non-redundant ABP chains); larger datasets may improve future models
• Physico-chemical property-based model underperforms sequence and PSSM models
• BLAST-based prediction entirely unsuitable for 97 of 168 ABP chains in the dataset
• Structural information (3D coordinates) not used — methods that use known structure would have higher accuracy but are inapplicable to uncharacterized proteins

Applications

• Functional annotation of ATP-binding proteins in post-genomic proteomics
• Identification of ATP-interacting residues in newly sequenced proteins
• Support for kinase substrate and phosphorylation site research
• Guiding mutagenesis experiments to validate ATP-binding residues
• Understanding protein–ligand interaction mechanisms in enzymes, motor proteins, and transporters

Contact & Authors

Prof. Gajendra P. S. Raghava**

IIIT DELHI raghava@iiitd.ac.in Institute of Microbial Technology, Chandigarh, India raghava@imtech.res.in https://webs.iiitd.edu.in/raghava/atpint/

License

This article is distributed under the Creative Commons Attribution License (CC BY 2.0) © 2009 Chauhan et al; licensee BioMed Central Ltd.