DNA extraction and characterization of ancient samples

DNA extraction was attempted on 102 bone specimens (83 from Turin, 7 from Cuneo and 12 from Alessandria territories) from five early medieval contexts, archeologically defined as “Lombard”: Rivoli Corso Levi [19], Rivoli La Perosa [20], Mombello Monferrato [21,22], Collegno [19,23] and Centallo San Gervasio [24,25] (Table 1, where the GenBank accession numbers of their control-region sequences are reported). These specimens belong to Direzione Regionale per i Beni culturali del Piemonte, Soprintendenza per i Beni Archeologici del Piemonte e del Museo Egizio, Turin, which can be contacted to ask for access to the samples; the permit to analyze them was granted by a letter, protocol N MBAC-SBA-PE UFFPROT 00554324/05/2012, CI34.04.07/74.1 of May 24th, 2012, signed by the superintendent, Dr. Egle Micheletto. In Italy, no ethical committee is required to authorize studies on ancient samples. The experts who analyzed the specimens, CG, EB and LPB, are among the authors of this study, and an appropriate written informed consent was obtained from each of them. The collection and the analysis of their mtDNA was approved by the Ethics Committee for Clinical Experimentation at the University of Pavia, Board minutes of April 11th, 2013.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Samples analyzed by locality and their results.

https://doi.org/10.1371/journal.pone.0116801.t001

All the specimens were represented by compact bone tissue; teeth were available only for a subset of the individuals, and most of them did not appear suitable for analysis because of root damages due to fractures and/or pathologies. Most of the specimens were not washed or manipulated in any way; complete information on the handling history of each specimen is available. HVS-1 sequences of all the archaeologists and the geneticists who have been in contact with the samples were determined in order to help identify possible sources of modern contamination. DNA extraction and amplification was performed in the Ancient DNA-Palaeogenetic Laboratory in Florence. This facility is exclusively dedicated to examining ancient DNA. Laboratory rooms for Pre- and Post-PCR work are strictly separated and the work was carried out while wearing clean overalls, disposable facemasks, face shields, gloves and over-shoes. Different sets of pipettes were used for DNA extraction, PCR amplification and analysis of the PCR products. All benches and rooms were routinely treated with bleach and UV-irradiated. In order to identify potential contamination, at least one extraction or amplification blank every five samples was routinely used as negative control. The specimens were cleaned by removing the external surface with a micro-drill Marathon Multi 600 with disposable tools, UV-irradiated (254nm wavelength) in a cross-linker on each side for 45 minutes and subsequently ground into a fine powder with the micro-drill. DNA was extracted from bone powder by means of a silica-based protocol [26]. Two μl of extracted DNA was used for amplification of the HVR-I region by PCR [6 modified by using a polymerase with proofreading activity]. Three primer pairs (L15995-H16132, L16107-H16261, L16247-H 16402) were used to target a subdivided 361 bp of the HVR-I via three overlapping fragments.This process was repeated for specimens reporting positive HVR-I result using a second anatomical element and a different silica-based DNA extraction protocol [27]. A third anatomical element for 3 samples was also independently examined in the Paleogenomic Lab of the Institut de Biologia Evolutiva in Barcelona. In this second laboratory, samples were extracted with a phenol-chlorophorm-based method [28] and the HVR-I was amplified using two sets of overlapping primers: L16055-H16218 and L16185-H16378.

All PCR products were cloned using TOPO TA Cloning Kit (Invitrogen) according to the manufacturer’s instructions. Screening of white recombinant colonies was accomplished by PCR [6] and agarose gel electrophoresis. After purification of these PCR products, a volume of 3 μl was cycle-sequenced using Forward M13 universal primer following the BigDye Terminator v1.1 Cycle Sequencing Kit (Applied Biosystems) supplier’s instructions. The sequence was determined using an Applied BioSystems 3100 DNA sequencer. Different clones were sequenced for each individual from whom there were two or three extractions (S1 Fig.). HVR-I sequences were then aligned and compared across clones in order to define the consensus sequence. This consensus sequence was compared to the Cambridge Reference Sequence, CRS [29] in order to define the HVR-I mutational motif. Haplogroup assignment was carried out using the software HaploGrep [30] (http://haplogrep.uibk.ac.at) with a further check based on the mitochondrial haplogroup phylogeny in Phylotree [31] (http://www.phylotree.org).

DNA extraction and characterization of the researchers

MtDNA genotypes of all experts who handled the ancient samples were determined. After the informed consent of all donors, epithelial cells were collected from the mucosa on both sides of the oral cavity using FLOQSwabs forensic buccal swabs, DNA free (Copan flock technologies). DNA extraction as well as PCR and sequencing reaction setup involving modern samples was carried out in a laboratory that was physically separated from the laboratory where the ancient samples were analyzed. DNA was extracted using QIAmp DNA Investigator Kit (QIAGEN, Hagen, Germany) and the HVR-I was amplified by 28 cycles of PCR as reported in [32]. The entire mtDNA HVR-I region was amplified using a single primer pair named L15995-H16402. The amplification products were purified with the MinElute PCR purification Kit (Qiagen) and then sequenced directly with the same amplification primers (forward and reverse) following the BigDye Terminator v1.1 Cycle Sequencing Kit supplier’s instructions.

DNA extraction and characterization of modern samples

The 75 samples from Trino Vercellese were collected in 1994 from individuals belonging to the Association “Partecipanza dei boschi”, whose members have been transmitting their membership from generation to generation since the Middle Age, and therefore have been settled in the same archaeologically important area for at least 20–30 generations; the 89 samples from Postua, also collected in 1994, were from individuals with Piedmontese surnames (and for this reason living in the same area for many generations) with the purpose to act as controls for the previous sample. The Authors involved in the collection were AP, GM and CDG. The samples of Trino and Postua are used only for research purposes and their individual informed consents collected by the Authors above were approved by the Ethics Committee of the University of Turin. The 58 subjects from Val di Susa were unrelated on the maternal line for at least three generations. Also in this case, for all of them, appropriate written informed consent was obtained, and the study was approved by the Ethics Committee of the University of Turin. The study of mtDNA variation of all these 222 Piedmontese samples was approved by the Ethics Committee for Clinical Experimentation at the University of Pavia, Board minutes of April 11th, 2013. The samples were anonymized prior to being accessed by the authors. Details about these samples, including the GenBank accession numbers of their control-region sequences, are available in S1 Table.

Dataset of modern and ancient sequences for comparisons

Previously published and unpublished data were compiled to produce a dataset consisting of 7,249 sequences for 74 modern and 5 ancient populations, in order to test if the entire medieval sample could be considered a homogenous group. The dataset also comprised sequences from the Piedmont region that could be used in direct pairwise comparisons with our medieval samples in order to test for population/genealogical continuity in the region. In addition to the 222 modern individuals from Trino Vercellese, Postua and Val di Susa mentioned above, also 50 sequences from Turin, a putatively cosmopolitan population were included in the Approximate Bayesian Computation analysis in order to account for the potential effects of recent immigration [33] (S2 Table).

Phylogenetic and population analysis

The evolutionary relationships of medieval and modern samples were investigated through median-joining networks of control-region haplotypes constructed with the Network 4.6 software (www.fluxus-engineering.com) [34] by using the reduced median algorithm (ρ = 2), followed by the median-joining algorithm (ε = 0). Nucleotide weighting (ω) was adjusted to reflect well-known phylogenetic data: i) the C-stretch between 19182 and 19193 was down weighed; ii) all samples were grouped into four major clusters corresponding to macro-haplogroups HV, UK, JT and IWX, based onto haplogroup classification reported in S1 Table. The consensus medieval sequences were compared with a dataset of 79 modern and ancient populations from Europe, North Africa, Near East and Central Asia. Variability at the intra-population level was investigated using Arlequin 3.5.1.2 [35] by calculating haplotype diversity (or heterozygosity) and mean number of pairwise differences (MPWD) (S2 Table). A matrix of pairwise φ_ST distances (equivalent to an F_ST also considering molecular distances between alleles) was estimated using the Kimura-2p model [36] using the same software, and a Multidimensional Scaling plot was obtained from that matrix using the R MASS package [37].

Approximate Bayesian Computation analysis

We used an Approximate Bayesian Computation (ABC) framework [38] in order to investigate the genealogical and evolutionary relationships among the medieval and modern Piedmont populations. ABC utilizes simulations in order to assess the probability of data given a particular model and associated parameters (often in the absence of a tractable likelihood equation) [39]. Our ABC procedure [9] can be summarized as follows: 1 million genetic datasets are generated with the same features of the observed one (i.e. number of individuals, age of the samples, length of the sequences) by coalescent simulation for each demographic model under investigation (for a total of 16 million genealogies), drawing model parameters from the associated prior distributions. The patterns of genetic variation in the observed and simulated data are then compared via Euclidean distance using a set of summary statistics. Only the simulations generating summary statistics close to the observed ones (i.e. those associated with the shortest Euclidean distances), are then considered to estimate the posterior probabilities of the various models and parameters. For model comparison we used both the acceptance-rejection procedure (AR, [40] ) and the weighted multinomial logistic regression procedure (LR, [39]). Under the AR approach, the posterior probability of a model is obtained by considering only a certain number of “best” simulations, and then simply counting the proportion of these simulations that have been generated by each model under investigation. This method can be considered reliable only when applied to simulations (usually few) showing an excellent fit with the observed data (i.e. few hundreds, [41]). Alternatively, under the LR method, a logistic regression is fitted where the model is the categorical dependent variable and the summary statistics are the predictive variables (thus taking into account that some “best” simulations are closer to the observed data than others). The regression is local around the vector of observed summary statistics, and the probability of each model is finally evaluated in the point corresponding to the observed vector of summary statistics. The β coefficients of the regression model are estimated by maximum likelihood; the standard error of the estimates is taken as a measure of the accuracy of the method. To evaluate the stability of the so calculated models’ posterior probabilities we considered different thresholds, i.e. different number of retained simulations (100, 200, 300, 400, 500 best simulations for AR; 25000, 50000, 75000, 100000, 125000, 150000 best simulations for LR). Model parameters were estimated by a locally weighted multivariate regression [38] after a logtan transformation [42] of the 1,000 best-fitting simulations from a specific model. The models’ posterior probabilities were estimated using R scripts from http://code.google.com/p/popabc/source/browse/#svn%2Ftrunk%2Fscripts, modified by SG.

Demographic Models and Summary Statistics

We compared two main demographic models that differed with regard to the relationships between modern and ancient samples (Fig. 2). Under each model, each modern population (in turn, Trino Vercellese, Postua, Val di Susa and Turin) was independently compared with the medieval sample. In Model 1-continuity the medieval population (placed at 55 generation ago, i.e. 1375 years assuming 25 years per generation) is a direct ancestor of the modern population, and the demography follows an exponential growth that started at some point in the past (i.e. more than 55 generations ago, which is the average age of the early medieval samples). Under Model 2-discontinuity, the ancient and the modern sample descend from two different branches of a phylogeny (i.e. no genealogical relationship between them), with the medieval population not contributing to the modern gene pool, and the latter experiencing an exponential growth until the present time (Fig. 2). We also extended the comparison by including a population bottleneck mimicking the 13^th century plague epidemics that reduced the population size by an estimated one-third [43] size (Model 1-continuity plague and Model 2-discontinuity plague; S2 Fig.). The prior distributions were all uniform (log-uniform for the effective population sizes) and are detailed in S3 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Demographic models tested and associated posterior probabilities.

https://doi.org/10.1371/journal.pone.0116801.g002

We generated 1 million simulated datasets for each model by the program BayeSSC ([44], see http://iod.ucsd.edu/simplex/ssc/BayeSSc.htm); to calculate the posterior probabilities for models and parameters we used R scripts from http://code.google.com/p/popabc/source/browse/#svn%2Ftrunk%2Fscripts, modified by SG. To summarize the data we calculated the following six statistics using Arlequin ver. 3.5.1.2 [37]: the number of haplotypes for each population, the number of private polymorphic sites per population, the mean pairwise difference and gene diversity for each population, Hudson’s F_ST [45] and a measure of allele sharing, defined as the number of haplotypes of the modern sample also present in the ancient sample, scaled by the total number of haplotypes in the latter (S4 Table).

Type I Error and ROC analysis

We estimated the probability to reject the true null hypothesis (i.e Type I Error), by evaluating the proportion of cases in which 1,000 pseudo-observed datasets (PODs) randomly generated under each model were not correctly identified by the ABC analysis (both AR and LR procedures, 100 and 50,000 retained simulations in turn). The power of the model choice procedure was evaluated using a wide range of decision probability thresholds to identify the support for a specific model, i.e. 0.5, 0.6, 0.7, 0.8, 0.9. In addition, we calculated a receiver operating characteristic (ROC) curve, as in Bazin et al. (2010) [46] and Sousa et al. (2012) [47]. The method ranks the 2,000 posterior probabilities (1000 from each model) coming from the Type I Error analysis for one model (say, e.g., Model 1-continuity) from the highest to the lowest. For each of these posterior probabilities we know whether or not the data are generated by Model 1-continuity. At this point, the proportion of true and false positives is evaluated for decreasing thresholds, and the ROC curve is built as follows: first, we consider a posterior probability of 1.0 as the threshold for deciding whether to classify the data as coming from Model 1-continuity, having a proportion 0.0 of the simulated Model 1-continuity cases correctly classified (true positives), but also a proportion 0.0 of Model 2-discontinuity cases incorrectly classified (false positives). This represents the lower-left corner point of the ROC curve. Then we set 0.0 as the threshold, and we have a proportion of 1.0 of all the Model 1-continuity PODs classified correctly (true positives) and a proportion 1.0 of all the Model 2-discontinuity PODS incorrectly classified (false positives). This represents the point on the top right corner of the ROC curve. The remaining curve is then constructed by successively taking the posterior probabilities in the list from highest to lowest and plotting the proportion of Model 1-continuity cases that are correctly classified (true positives) and the proportion of Model 2-discontinuity cases that are incorrectly classified (false positives). The ideal curve is built when all the Model 1-continuity cases occur first in the list (i.e. higher posterior probabilities for the true model), followed by all the Model 2-discontinuity cases, in which case the area under the ROC curve (AUC) would be 1. The ROC analysis was performed with the method implemented in the ROCR R package [48].

Mitochondrial variation of the samples studied

Among the 28 medieval individuals sequenced we observed 18 distinct haplotypes with 23 segregating sites (S4A Table). Possible relationship between individuals presumed by archaeologists has been verified using genetic data in order to avoid that possible kinship could affect the haplotypic frequencies. Phylogenetic links between haplotypes and their distribution among the archeological sites are shown in a Median Joining Network (Fig. 3A). Comparing our ancient dataset to the modern haplotypes, the phylogenetic network reported in Fig. 3B reveals that five haplotypes are shared between medieval and modern samples. The 18 medieval haplotypes encompass almost the entire range of western Eurasian mtDNA macro-haplogroups (Table 1). Similarly to most modern European populations, haplogroup H is by far the most represented encompassing 50% of mtDNAs, while the remainders are members of I, J, T, U2e, U4 and U5a, all also commonly observed in Europeans [15], including the modern Piedmontese populations analyzed in this study (S1 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Median Joining Networks.

Node sizes are proportional to haplotype frequencies. Variable positions are indicated along links that connect haplotypes. Nucleotide changes are specified only in the case of transversions. (A) Median Joining Network for the medieval sample. Different colors represent archaeological sites where the haplotype is present (see legend). Segregating sites are shown in red. (B) Median Joining Network of the ancient (colored in red) and modern samples (Trino Vercellese in grey; Postua in black and Val di Susa in white). Major haplogroups are named according to the current nomenclature.

https://doi.org/10.1371/journal.pone.0116801.g003

In the Piedmont medieval sample, gene diversity (0.942) and Mean Number of Pairwise Differences (3.484) are similar to, but both lower than, the average found for the other 79 (ancient and contemporary) populations (mean 0.957 and 4.067 respectively, see S2 Table). Among other ancient populations, estimates of internal genetic variation appear close in the Etruscans (0.943; 2.966) [9] and lower in Bronze-age Sardinians (0.830; 1.390) [8], pre-Roman Iberians (gene diversity not given; 2.120) [51] and in a Medieval sample from Tuscany (0.860; 1.971) [6].

Accordingly, we see no reason to suspect that the early medieval, or “Lombard” individuals may represent a heterogeneous assemblage of people of different origins, even considering their vicinity in space and time and their archaeological affinities, and from this point on we felt justified in treating them as a single population. F_ST distances place Lombards in the middle of the area in which most of the populations cluster, graphically represented in S3 Fig. by a MDS plot.

We calculated two measures of genetic distance between the medieval and the modern populations from Trino Vercellese, Postua, Val di Susa and Turin, namely Hudson’s F_ST and the Allele Sharing (S4B Table). The lowest value of F_ST and the highest value of allele sharing is between medieval samples and Trino Vercellese, whereas the highest distances from the medieval samples are with Postua. Postua is the least variable sample: of the 89 individuals analyzed only 27 distinct sequences are found, and gene diversity is the lowest among the studied populations, suggesting a likely reduction in population size and thus substantial genetic drift in this population, at least along the maternal line, which may lower power if testing for continuity. The Val di Susa population shows levels of allele sharing with the Lombards similar to those observed for Trino Vercellese but a higher F_ST, a seemingly contradictory result, possibly due to the high number of mitochondrial haplotypes, making F_ST a less-than-optimal descriptor of their diversity.

Type I Error and Approximate Bayesian Computation

We compared Model 1- continuity and Model 2-discontinuity (detailed in Methods) to verify whether there is enough power in the data to discriminate between them using our ABC approach. S6 Table (top and bottom panels) shows the probability of identifying the true model when data are generated according to either Model 1-continuity or Model 2-discontinuity, both using the AR (100 simulations) and the LR (50,000 simulations) method. The power of the analysis was generally high; when the threshold was >0.9 (i.e. support for a model is assigned when its posterior probability is higher than 90%) the proportion of false positives was very low (0.013 at maximum). Furthermore, when the decision probability threshold was 0.5, the probability to recognize the true model was never lower than 96%. The receiver operating characteristic (ROC) curve analysis, shown in Fig. 4, confirmed the good performance of our analysis framework. For both the AR and the LR procedures, the ROC curve was close to the upper left corner of the plot, indicating that our ABC analysis efficiently identifies the model that generated the data. The capability to correctly predict the true model was also assessed comparing at the same time Model1, Model2, Model1plague, and Model2plague. Considering a decision probability threshold of 0.5, the power of our ABC procedure was generally high for each comparison, with a decrease of power (depending on the model that has to be recognized) for higher decisional thresholds (data not shown).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Receiver operating characteristic (ROC) curve for Model 1-continuity and Model2.

Dashed line: Acceptance Rejection approach; solid line: Logistic Regression approach.

https://doi.org/10.1371/journal.pone.0116801.g004

We then compared the fit of both models to our data, examining the medieval population alongside each of the four modern populations in separate analyses. We also repeated the analysis including a genetic bottleneck in both models (Model 1-continuity plague and Model 2-discontinuity plague) (S1 File). In all cases models without bottleneck proved to better fit the observed data than models including the bottleneck (S1 File). Fig. 2 shows the results of the comparison between Model 1-continuity and Model 2-discontinuity respectively based on the best-fitting 50,000 and 125,000 simulation experiments, under the LR procedure. We found evidence for genealogical continuity since Lombard times between our medieval sample and modern samples only when considering the population of Trino Vercellese: the posterior probability of Model 1-continuity ranged from 0.90 to 0.93. The principal component analysis (PCA) of the first 10,000 best simulations from each model (i.e. the 10,000 simulation closest to the observed dataset that are generated by each model) actually shows that the point corresponding to the observed data falls in the middle of the results obtained simulating genealogical continuity (S4A Fig.). By contrast, for all the other modern populations the best supported model was the one in which the medieval sample represents a separate branch of the genealogical tree, i.e. Model 2-discontinuity. In some cases the probability associated to Model 2-discontinuity was consistently very high (Postua, S1 File), whereas in other cases the AR and LR methods were at odds (Turin, S1 File), or returned quite different probabilities (Val di Susa, S1 File).

PCA plots (S4 Fig.) reflect the same situation. The signal for Val di Susa (S4C Fig.) is indeed quite noisy, and the observed value of Turin falls on the edge of the variation generated by Model 2-discontinuity with a slight overlap with Model 1-continuity (S4D Fig.), again emphasizing the importance of weighting the simulations with respect to their distance from the observed dataset when assigning the models’ posterior probabilities (LR procedure). Even if it is generally assumed that the logistic regression approach is more reliable than the straightforward acceptance rejection algorithm in estimating posterior probabilities [39], to strengthen these estimates we also evaluated the AR posterior probabilities for each model in a point corresponding to twice the standard error associated with the β coefficients of the fitted regression model (in both directions). For the Val di Susa sample the support was confirmed for Model 2-discontinuity, whereas for Turin the confidence in the estimated posterior probabilities increased with the number of simulations retained, and a stable support for Model 2-discontinuity was reached when considering more than 100,000 simulations (data not shown).

These inconsistences in the probabilities estimated at different thresholds are not surprising, considering the ample degree of overlap between the models we were trying to discriminate. It is probably necessary to analyze more data, both in term of loci and/or individuals, to be able to reject a model with good statistical confidence. Overall these results, along with the PCA plots of the simulated and observed data (S4 Fig.), support a model in which people of Lombard times appear to have contributed little, if at all, to the ancestry of contemporary people in three localities considered, and presumably over the whole area. There is, however, one remarkable exception, namely Trino Vercellese, where the genetic evidence suggests instead the existence of genealogical ties across more than 1,500 years. One can speculate that the existence in Trino Vercellese of an association involving essentially all families since medieval times might have contributed to maintaining a comparatively high level of genetic continuity. Much like in previous ancient DNA analyses, notably in Sardinia [8] and Tuscany [9], this study suggests that the modern population is a patchwork of groups with different genealogical histories; instances of clear-cut, long-term genetic continuity, all the way back to the Early Middle Age (as in the case for Trino Vercellese), or even much earlier (as is the case for Sardinia and Tuscany), exist and can be detected, but seem to represent less the rule than the exception.