Methodological overview of latent signal detection algorithm. (a) Learning the side-effect profile of a disease. We assessed each adverse event reported in the Adverse Event Reporting System, S_i, for significant enrichment on reports involving a type 2 diabetes mellitus (T2DM) drug (identified by the triangle indicators) vs. a background of all drugs and determined the significance by using a Fisher’s exact test. We call the result of this analysis the “disease’s side-effect profile.” (b) We then extracted reports for all possible pairs of drugs (~40,000) and scored each drug pair’s side-effect profile for its similarity to the disease’s profile. Note that we considered only the adverse events that were significantly correlated with the disease (colored boxes, i.e., S₁, S₂, and S₃) and not others (gray boxes). We ranked each drug pair according to its similarity score, and (c) clinically validated the top-ranked pairs of drugs for interaction effects on a predetermined phenotype extracted from the electronic medical records (EMRs) (e.g., random blood glucose concentration). Dx, diagnosis.

Mean baseline and post-treatment blood glucose concentrations with 95% confidence intervals for a combined analysis of data from the three independent sites. The left panel shows the analysis excluding data relating to patients with diabetes, and the right panel shows the analysis including such data. Both combination cohorts (triangles) had significantly increased random blood glucose levels: 18.5 mg/dl (1.0 mmol/l), P < 0.001 after excluding data from patients with diabetes, and 48 mg/dl (2.7 mmol/l), P < 0.001 after including these data. No significant changes in glucose levels were observed in the pravastatinonly (diamonds) cohorts. A slight decrease in glucose levels was observed in the paroxetine-only (circles) cohorts in both analyses: −3.3 mg/dl (−0.2 mmol/l), P < 0.001, after excluding data from patients with diabetes, and −4 mg/dl (−0.2 mmol/l), P < 0.001 after including these data (Table 3).

Combination treatment with paroxetine and pravastatin is associated with increase in blood glucose levels. Mean baseline and post-treatment blood glucose concentrations with 95% confidence intervals at each of the three independent replication sites are shown. Each site contained three patient cohorts, pravastatin-only (diamonds), paroxetine-only (circles), and pravastatin plus paroxetine (triangles). Site 3 contained the largest cohort of patients on combined pravastatin and paroxetine therapy (N = 109), followed by site 2 (N = 18), and site 1 (N = 8). The treatment condition was significant at each site per ANCOVA. In post hoc tests the paroxetine plus pravastatin combination cohorts at each site showed significantly greater increases in mean blood glucose levels relative to the cohorts of paroxetine-only (P = 0.008, P = 0.009, and P < 1 × 10⁻¹⁰) and pravastatin-only (P = 0.036, P = 0.002, and P < 1 × 10⁻¹⁰). The mean change in blood glucose in patients on the combined treatment was 23 mg/dl (1.3 mmol/l), 26 mg/dl (1.4 mmol/l), and 17 mg/dl (0.9 mmol/l) at sites 1, 2, and 3, respectively. The pravastatin-only cohorts had changes of 3.3 mg/dl (0.2 mmol/l), −1.3 mg/dl (−0.1 mmol/l), and −3.2 mg/dl (−0.2 mmol/l) at the three sites, respectively. The paroxetine-only cohorts had changes of 0.8 mg/dl (0.04 mmol/l), −1.5 mg/dl (−0.1 mmol/l), and −6.3 mg/dl (−0.3 mmol/l) at the three sites, respectively.

Preliminary analysis of mean and standard error of fasting blood glucose concentrations in five groups of mice (n = 10/group). Asterisks indicate significance in a multivariate linear model with covariates (***P < 0.001, **P < 0.01, *P < 0.05). (Diet control, DC) Normal diet and no treatment. (Treatment control, TC) High fat, high sucrose diet (HFSD) plus saline vehicle. (Pravastatin, PR) HFSD plus pravastatin at 1.25 mg/kg/day. (Paroxetine, PA) HFSD plus paroxetine at 1.25 mg/kg/day. (Paroxetine and pravastatin, CO) HFSD plus paroxetine and pravastatin at 1.25 mg/kg/day for each drug. Animals were kept on their respective diets for 3 months prior to drug exposure. We measured 4-hour fasting blood glucose concentration after 3 weeks of drug treatment. Fasting glucoses were 68.5 ± 31.3 mg/dl (3.8 ± 1.7 mmol/l), 74.2 ± 26.7 mg/dl (4.1 ± 1.5 mmol/l), 94.4 ± 29.6 mg/dl (5.2 ± 1.6 mmol/l), 128.1 ± 31.3 mg/dl (7.1 ± 1.7 mmol/l), 193.1 ± 23.8 mg/dl (10.7 ± 1.3) for DC, PR, PA, TC, CO groups, respectively. We tested the interaction effect of pravastatin and paroxetine using an analysis of variance with covariates and interaction terms on the log transform of the fasting glucose measurements. We included diet as a covariate in addition to the pravastatin and paroxetine exposures and the interaction term. Our preliminary analysis shows a significant effect from diet (β = −0.19 ± 0.13, P = 0.004), pravastatin (β = −0.15 ± 0.13, P = 0.02), and the strongest effect from the interaction of paroxetine and pravastatin (β = 0.33 ± 0.18, P = 5.0 × 10⁻⁴). We assessed significance using an F test.