TL;DR
This paper presents a large-scale bioinformatic analysis of nonsynonymous single nucleotide polymorphisms (nsSNPs) to predict their functional impact on proteins. By comparing various prediction methods and applying them to both disease-associated (OMIM) and common (dbSNP) variants, the authors demonstrate that prediction tools can effectively distinguish deleterious mutations. A key finding is that rare nsSNPs are predicted to be deleterious as frequently as common nsSNPs, challenging previous assumptions.
What Problem Does This Paper Address?
The human genome is replete with genetic variations, many of which are single nucleotide polymorphisms (SNPs). Nonsynonymous SNPs (nsSNPs) are particularly interesting because they alter the amino acid sequence of proteins, potentially impacting their structure and function. With the advent of genome-wide association studies (GWAS) aiming to identify genetic predispositions to common diseases, there's a growing need for efficient methods to predict the functional consequences of these nsSNPs. This is crucial for pinpointing causal variants within disease-associated genomic regions and for guiding further functional studies. The challenge lies in the sheer volume of nsSNPs and the difficulty in experimentally determining the functional impact of each one. This paper addresses the need for robust bioinformatic tools to prioritize nsSNPs for experimental investigation.
Methods
The study employs a multi-pronged bioinformatic approach, combining homology modeling with various prediction algorithms.
- Homology Modeling: To overcome the lack of experimental structures for many proteins, the researchers built 3D models using homology modeling. They first used the FUGUE program for fold recognition to identify potential structural templates from the HOMSTRAD database. Subsequently, the Modeller program was used to build structural models for over 4,700 genes.
- Prediction Methods: A suite of established and novel prediction methods were utilized to assess the impact of nsSNPs:
- Structure-based methods: These included SDM (Site Directed Mutator), I-Mutant2.0, PolyPhen, and environment substitution scores (env_score). These methods leverage structural information (either experimental or modeled) to predict changes in protein stability or function.
- Sequence-based methods: SIFT (Sorts Intolerant From Tolerant substitutions) was used, which relies on sequence conservation without requiring structural information.
- Combined/Knowledge-based methods: LS-SNP, which uses machine learning and knowledge-based rules, was also employed.
- Functional residue prediction: The Crescendo method was used to identify residues likely involved in protein function or interactions by comparing observed sequence conservation with that predicted by local environment substitution tables.
The study systematically applied these methods to large datasets of nsSNPs and compared their predictions.
Datasets
The study utilized two primary datasets of non-synonymous SNPs:
- OMIM: This dataset comprises 2,256 sequence variations found in 500 genes previously identified for Mendelian inherited disorders. These are generally rare mutations with high penetrance, often associated with monogenic diseases.
- dbSNP (version 125): This dataset contains 21,471 sequence variations from nearly 5,000 genes. The majority of these nsSNPs have a minor allele frequency (MAF) greater than 0.05 and are typically considered to have a neutral effect on protein function, though some may be associated with common diseases.
The HapMap database was also used to classify dbSNP variants by allele frequency.
Key Insights and Ideas
- Predictive Power of Tools: The study confirms that existing bioinformatic tools, when applied systematically, can effectively differentiate between mutations likely to be deleterious and those likely to be neutral. This is demonstrated by the higher proportion of deleterious predictions for OMIM variants compared to dbSNP variants.
- Rare vs. Common nsSNPs: A significant finding is that rare nsSNPs (MAF < 0.01) are predicted to be deleterious as often as common nsSNPs (MAF > 0.05). This challenges the prevailing "common-disease/common-variant" hypothesis, suggesting that rare, potentially highly impactful variants might play a more substantial role in disease susceptibility than previously assumed.
- Complementarity of Methods: The comparison of different prediction methods highlights their strengths and weaknesses. While structure-based methods offer detailed insights, sequence-based methods like SIFT are crucial for broader coverage, especially when structural information is unavailable. The study also emphasizes the value of combining predictions from multiple methods for increased confidence.
- Importance of Modeled Structures: The successful application of homology modeling to a large number of genes demonstrates its utility in extending the reach of structure-based prediction methods, enabling analysis of nsSNPs in proteins lacking experimental structures.
Key Results
- Structural Prediction Coverage: Homology modeling successfully generated structural models for a significant portion of the genes analyzed, with over 90% of genes having at least one predicted structural domain.
- Deleterious nsSNP Predictions:
- For OMIM variants, approximately 51% were predicted to be deleterious by SIFT.
- For dbSNP variants, approximately 25% were predicted to be deleterious by SIFT.
- When considering multiple prediction methods, the overlap for deleterious predictions increased, with a small percentage of nsSNPs being flagged by all methods.
- Allele Frequency and Deleteriousness: Analysis of dbSNP variants showed that while 76% had MAF > 0.05 and 8% had MAF < 0.01, the proportion of rare alleles predicted to be deleterious by four or more methods was 12%, compared to 72% for common alleles. This suggests a higher relative burden of deleterious mutations in rare variants.
- Functional Residue Predictions: Crescendo identified functional residues, with OMIM mutations showing a slightly higher proportion of positive Z-scores (indicating functional residues) compared to dbSNP mutations.
Limitations and Future Work
- Structural Model Accuracy: The accuracy of homology models can vary, and the study acknowledges that careful analysis of each model can improve predictions, but this is time-consuming.
- Data Sparsity: Information on functional residues and protein-protein interactions is often sparse and inconsistent, making comprehensive functional prediction challenging.
- Validation of dbSNP Data: The validity of many nsSNPs in dbSNP remains unclear, as only a fraction have been validated in large sequencing projects. The authors stress the need for accurate estimates of nsSNP validity.
- Rare Allele Interpretation: While rare alleles are predicted to be deleterious as often as common ones, the interpretation of their role in common diseases is complex and may be confounded by biases in prediction datasets.
- Future Work: The authors suggest that the automated analysis and model-building pipeline described can be readily applied to new sequence variations associated with disease. Future work could involve improving data extraction from biological databases, predicting multi-protein complexes, and analyzing interactions between genes. Further detailed modeling and structural analysis of specific genes are also planned.
Summary updated