Fig 1.
Outline of the demographic model used for training ArchIE.
We simulate a population starting at size N0 and splitting into archaic and modern human (MH) populations at time T0. The MH population splits into a reference and target population of size N1 and N2, respectively, at time Ts. Then, at time Ta, the archaic population admixes with the target population with an associated admixture proportion m. We use data simulated from this model to train a logistic regression classifier.
Table 1.
Parameters used in training simulations.
Fig 2.
ArchIE obtains improved accuracy over related methods.
(A) Precision-Recall (PR) and (B) Receiver Operator Characteristic (ROC) curves for ArchIE (black circles), S* (red crosses), and S’ (purple triangles) in a 2% admixture scenario with a Human-Neanderthal demography. The dashed line corresponds to a false discovery rate of 20%.
Fig 3.
Relative importance of the features used as input to ArchIE.
We examined the log of the absolute value of the standardized weights associated with each of the features included in the logistic regression model underlying ArchIE. Negative values indicate standardize weights with absolute values less than 1. (A) The individual frequency spectrum mostly has small weights and lower frequency entries generally have larger weights associated with them. (B) The first three entries indicate the moments of the distance vector. The minimum distance to the reference population, skew, and variance of the distance vector have the largest weights associated with them.
Fig 4.
ArchIE is robust to misspecification in the demographic model.
We tested ArchIE on data simulated after perturbing single demographic parameters lower (left, orange) and higher (right, blue) relative to their values in the training data. Values are reported as log10 fold changes compared to the baseline model performance. We report (a, b) recall and (c,d) precision at the threshold that gives a precision of 0.8 on the unperturbed test data (P(archaic) = 0.62).
Fig 5.
Application of ArchIE to 1000 Genomes European population (CEU).
(A) Percentage of genome called archaic as a function of the threshold on the probability of archaic ancestry estimated by ArchIE. The dashed line refers to the threshold that yields a 20% FDR in simulations. (B) Mean Neanderthal match statistic (higher implies more similar to the sequenced Altai Neanderthal genome) for haplotypes inferred as archaic vs non-archaic as a function of the probability threshold. (C) Frequency of haplotypes confidently labeled as archaic near the BNC2 gene and (D) the OAS gene cluster. (E) Mean frequency of confidently archaic segments increases with B-statistic (a measure of selective constraint). Low B-statistic denotes more selectively constrained regions (standard errors estimates are obtained using a 1 Mb block jackknife).