• Type 2 diabetes is characterized by insulin resistance in peripheral tissues and dysregulated insulin secretion by pancreatic \(\beta\)-cells.
• Pathogenesis is heterogeneous.

• Our data set contains, 6424 SNPs, for 248 individuals, with 159 controls and 89 cases of T2D.

• We want to remove any artifacts which may increase Type I or Type II errors. Examples of these artifacts include (Anderson 2010):
  • Differences in population structure
  • Differences in DNA quality or handling procedures
  • Missing genotype rate
• Quality assurance (QA) relates to the post-production review of quality (Laurie 2010). Two categories of QA:
  • Subject-level
  • SNP-level

• Using a range of QA measures at the subject and SNP level we end up dropping 8 subjects and 493 SNPs (table below can include double counting).

• A logistic regression models the log-odds of observing a given phenotype for person \(i\) as a function of a baseline probability (the intercept), an additive measure of their genotype \(g_i\), and a vector of potential confounders \(x_i\).

\[ \begin{align} \log\Bigg(\frac{P(\text{Phenotype}_i)}{1-P(\text{Phenotype}_i)} \Bigg) &= \beta_0 + \beta_1 g_i + \gamma^Tx_i \end{align} \]

• Under the null hypothesis of no genotypic effect, \(H_0: \hspace{2mm} \beta_1=0\).
• We control for two potential confounders: gender and population structure.

• Our baseline \(g_i\) assumes an additive effect

• \(\beta_1\) represents the genotype relative risk (GRR)

• For example SNP rs2843403 has CC (112), TC (108), and TT (28), so:

\[ \begin{align} g_i^{\text{additive}} &= \begin{cases} 0 & \text{if } \text{CC} \\ 1 & \text{if } \text{TC} \\ 2 & \text{if } \text{TT} \end{cases} \hspace{3mm} \text{ and } \hspace{3mm} g_i^{\text{dominant}} = \begin{cases} 0 & \text{if } \text{CC} \\ 1 & \text{if } \text{TC} \\ 1 & \text{if } \text{TT} \end{cases} \hspace{3mm} \end{align} \]

• Family-wise error rate measures the number of false positives.
• Adjustments: "conservative" Bonferroni or Monte Carlo simulations
• Permutation testing which shuffles the phenotype labels:
  • Use we 5000 runs
  • Generates empirical distribution of p-values

• The figures below shows the estimated effect size for the two SNPs of interest. While the p-value measures statistical significance, it does not measure biological significance.

• rs10008252 \(\to\) SNORD65 (upstream gene variant)
  • Found on chromosome 4
  • No association to a phenotype
  • SNORD: Small nucleolar RNA whose primary function is to guide chemical modifications of other RNAs
• rs1045971 \(\to\) C16orf46 (aka FLJ32702 gene)
  • Found on chromosome 16
  • Codes for an uncharacterized protein
  • Microarray data shows expression most prevalent on testis tissue ADHD study
  • FLJ32702 interacts with FAT atypical cadherin 3 protein

• Only one cohort
• T2D is associated with common variants \(\to\) less power
• Small sample size \(\to\) less power
• Accounting for interactions within the genome and the environment:
  • GWAS does not consider the genetic interaction between loci (epistasis)
  • Nor the interaction(s) between loci and the environment

• Increase sample size: \(>1000\)
• Incorporate NGS techniques to discover less common and rare variation
• Develop informatics methods that examine epistasis and gene vs. environment interactions
  
• Meta-analysis with multiple cohorts
• Following meta-analysis, translate findings to the clinical setting using risk assessment tools for candidate genes \(\to\) gene-specific therapies