Pair Matcher (PaM): fast model-based optimization of treatment/case-control matches

Abstract

Motivation: In clinical trials, individuals are matched using demographic criteria, paired and then randomly assigned to treatment and control groups to determine a drug's efficacy. A chief cause for the irreproducibility of results across pilot to Phase-III trials is population stratification bias caused by the uneven distribution of ancestries in the treatment and control groups.

Results: Pair Matcher (PaM) addresses stratification bias by optimizing pairing assignments a priori and/or a posteriori to the trial using both genetic and demographic criteria. Using simulated and real datasets, we show that PaM identifies ideal and near-ideal pairs that are more genetically homogeneous than those identified based on competing methods, including the commonly used principal component analysis (PCA). Homogenizing the treatment (or case) and control groups can be expected to improve the accuracy and reproducibility of the trial or genetic study. PaM's ancestral inferences also allow characterizing responders and developing a precision medicine approach to treatment.

Availability and implementation: PaM is freely available via Rhttps://github.com/eelhaik/PAM and a web-interface at http://elhaik-matcher.sheffield.ac.uk/ElhaikLab/.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

PaM_simple performances on simulated datasets. Rows show the results of eight perturbed datasets [full dataset (left), remove-1 (centre) and remove-20 (right)]. PaM_simple was applied without a threshold (dashed) and with a threshold of 7 (solid red). Columns show the number of individuals assigned to a different pair than their original counterpart per dataset (x-axis), total GD between all matched pairs and total score (maximum of 10 per pair with the three datasets having n₁ = 500, n₂ = 499, and n₃ = 480 pairs, respectively)

Fig. 2.

PaM_simple (threshold of 7) performances on 16 simulated datasets against 5 competing methods. Columns show the number of misassigned individuals, total GD and pair score for Random assignment (red), Race model 1 (cyan), Race model 2 (yellow), Race model 3 (green) and PaM (black). Results for Datasets 9–16 were identical to those of Datasets 1–8 and are no shown

Fig. 3.

IBS distance between PaM (solid) and PCA (dashed) inferred pairs

Fig. 4.

The geographical distance between individual pairs inferred by PaM and PCA. Geographic distances are calculated between pairs where both individuals are within the IBS-defined clusters (A), where individuals are in different clusters (B) and for all individuals regardless of cluster assignment (C)

Fig. 5.

Pairing accuracy for various tools across multiple datasets. Boxplots summarize the pairing accuracy of all the trials in each population dataset (Supplementary Table S6), e.g. the PCA for unmixed individuals include the three analyses (PCA2/10/20) for each of the three datasets. The order of the tools’ results per population dataset is shown in the legend. Significance was estimated for PaM using Wilcoxon rank-sum test (P-value ≤ *0.05, ** ≤ 0.01)