The (in)dependence of single-cell data inferences on model constructs.

Abstract

Recent developments in single-cell analysis have revolutionized basic research and have garnered the attention of the forensic domain. Though single-cell analysis is not new to forensics, the ways in which these data can be generated and interpreted are. Modern interpretation strategies report likelihood ratios that rely on a model of the world that is a simplification of it. It is, therefore, plausible that different reasonable models will assign noticeably different weights of evidence (WoEs) to some of these data, resulting in inconsistent reports and protracted reviews of that evidence, potentially across years. With one goal of research being to identify and understand sources of inconsistencies during early stages, we undertake a study that evaluates WoE at the limit of one single-cell electropherogram (scEPG) across three architecturally distinct probabilistic models. The three are named EESCIt (Evidentiary Evaluation of Single Cells), TD (Top-Down), and DCM (Discrete Cell Model). To do this, we performance test the three models on a set of 996 individual scEPGs and conduct one H₁-true, i.e., true contributor, and 201 H₂-true, i.e., false contributor, tests, per scEPG. With the 201,192 outcomes per model, we confirm that scEPGs well resolve the hypotheses, regardless of what model was applied. We also observe that WoEs increase, on average, by 1 for every 1000 RFU of total intensity added until a plateau near the logarithm of the inverse of the random match probability is reached at ca. 22,000 RFU. By querying WoE calibration for each model, we determine if the evidence is over- or under-stated for any one of them. We find that for WoE ≥ -1 hardly any calibration discrepancy is observed. There were rare instances, however, for which WoEs that were ≤ -1 too strongly pointed in the negative direction, though H₁ was true. This was the result of five scEPGs that not only exhibited extreme signal in stutter positions, but also carried little information in other loci. These findings show that all three models appropriately stated WoEs for scEPGs when reporting positive WoE, and the two continuous model's WoE reasonably represented the findings when WoE < -1 for most loci. To further explore, we continued with paired analyses that evaluated the agreement in WoE, per scEPG, across models. Unlike unpaired analyses, this evaluation determines if well performing models return equivalent results for the same scEPG. The paired analysis was summarized by way of intraclass correlations, which were at least 0.99997. Further, we found that 762 of 996 WoEs were within a range of 3 orders of magnitude of each other, though many of these were associated with WoEs that were large, i.e., > 9, in the first instance. When we more closely focus on scEPGs giving ranges ≥ 3, but whose WoE ≤ 9 for at least one of the models, we find there are 21 of them. When we perform a locus-by-locus investigation of these 21 and of the five scEPGs returning too strong negative WoE for true contributors we find that extreme stutter is usually the cause of the challenges. To ameliorate differences in predicting rare, though impactful, events we proffer interpretive adaptions that extend beyond manually addressing the phenomena. With the WoE being calibrated within their relevant regions across EESCIt, TD and DCM, we categorize each as meeting the pillar of legitimacy for single-cell data within their intended WoE ranges.