INPS

INPS predicts the impact of non-synonymous single-point mutations on protein stability by estimating changes in thermodynamic free energy (ΔΔG) from sequence data to support protein engineering and variant interpretation.

Key Features:

• Sequence-Based SVM Regression: Employs Support Vector Machine (SVM) regression to predict ΔΔG from protein sequence, removing dependence on structural information.
• Thermodynamic ΔΔG Prediction: Estimates changes in thermodynamic free energy resulting from single-point amino acid substitutions.
• Performance and Validation: Validated against non-redundant datasets and performs comparably to state-of-the-art structure-based methods.
• p53 Variant Evaluation: Demonstrates robust performance in predicting stability changes for variants of the tumor suppressor protein p53.
• Complementarity with mCSM: Produces predictions complementary to the structure-based method mCSM, and combining INPS with mCSM improves predictive accuracy on datasets such as p53.
• INPS-MD and INPS3D Extensions: Extended into INPS-MD with INPS3D, which incorporates features derived from protein 3D structures and shows higher Pearson correlation with experimental ΔΔG (0.58 in cross-validation and 0.72 in blind tests).
• Competitive Structure-Informed Performance: Both sequence-based INPS and structure-based INPS3D perform competitively against existing methods.

Scientific Applications:

• Protein Engineering: Guides design and selection of mutations to modulate protein stability for industrial or therapeutic purposes.
• Genomic Research: Assists interpretation of Mendelian and somatic non-synonymous variants by predicting their impact on protein stability.
• Drug Development: Provides stability-based insights into disease mechanisms and potential targets for therapeutic intervention.

Methodology:

Applies machine learning—specifically SVM regression—on sequence-derived features to predict ΔΔG; INPS3D incorporates features derived from protein 3D structures; performance assessed by cross-validation and blind tests reporting Pearson correlations of 0.58 and 0.72, and the approach is intended for annotating large-scale datasets from Next Generation Sequencing.