High performance computation of landscape genomic models integrating local indices of spatial association

1 Introduction

• Early allele–environment correlations were later adapted for parallel logistic regressions, but the paper reports little subsequent progress in correlative approaches until recently.These approaches were used alongside population-genetic outlier methods as cross-validation.Source: §1 · p. 2
• Newer approaches address demographic effects through Bayesian models, latent factors, or spatially explicit random effects.The paper identifies high-performance whole-genome processing and integrated spatial statistics as remaining needs.Source: §1 · p. 2
• Samβada is presented as an open-source multivariate framework, illustrated on Ugandan cattle and compared with other selection-detection methods.The case study uses 813 cattle genotyped for about 40,000 SNPs.Source: §1 · p. 2

2 New Approaches

• Samβada improves on MatSAM with faster processing, multivariate analysis, and measures of spatial autocorrelation.It models genotype presence with logistic regression and fits models by maximum likelihood; biallelic SNPs are recoded as three genotypes.Source: §2 · pp. 2–3

2.1 Multivariate analysis

• In multivariate selection, each model is tested against simpler parent models formed by dropping one environmental parameter.The parent with the highest log-likelihood is the reference, making the G score the smallest among the parents; models are built from univariate upward to a user-defined maximum.Source: §2.1 · p. 3
• Continuous ancestry coefficients can be included with environmental variables to assess genotype associations while accounting for admixture.Including several ancestral-population coefficients can make detection of selection signatures highly conservative.Source: §2.1 · p. 4

2.2 Spatial autocorrelation

• Moran’s I measures global spatial autocorrelation, while LISA measures spatial dependence for individual points using neighboring values.Samβada offers fixed-distance and density-varying kernels, and assesses significance using permutations.Source: §2.2 · p. 4

3.1 Detection of selection signatures

• BayEnv, LFMM, and Arlequin comparisons used 40,019 SNPs from 804 samples, with method-specific significance procedures and population classifications.Samβada, LFMM, and Arlequin used a 1% threshold before Bonferroni correction; BayEnv candidates exceeded the first percentile of its neutral-locus Bayes-factor distribution.Source: §3.1 · p. 5
• 2,499 SNPs (6.2 %) were identified by Samβada, 1977 (4.9 %) by BayEnv, 280 (0.7 %) by LFMM, and none by Arlequin.Thirty-six loci were identified by all three correlative methods; three were among Samβada’s most significant models and lay close together on chromosome five.Source: §3.1 · p. 5
• 84 significant bivariate models corresponded to 29 loci when population structure was included as an environmental variable.For three models, an environmental variable improved estimation beyond a significant population-structure parent (p ≤7.9 · 10^-10); these loci were detected by all correlative approaches.Source: §3.1 · pp. 5–6

3.2 Spatial autocorrelation

• ARS-113 (GG) showed widespread positive spatial autocorrelation, whereas HM-28 (GG) had generally lower values and significant local autocorrelation only in north-west Uganda.The authors suggest ARS-113’s environmental association could reflect population structure, while HM-28’s lower spatial autocorrelation supports a possible adaptive origin of its association with isothermality.Source: §3.2 · p. 6

4 Discussion

• Multivariate models can account for population structure, while spatial-autocorrelation measures assess sample independence and may lower false-positive risk.Local indicators also map areas prone to spatial dependence.Source: §4 · pp. 6, 8
• Samβada detected the most candidate SNPs; its detections overlapped with 55% of LFMM’s and 22% of BayEnv’s, while LFMM discarded many loci with significant local spatial autocorrelation.The authors suggest the differences may reflect false positives in Samβada or false negatives in more conservative approaches.Source: §4 · p. 8
• Three Samβada detections made with population structure as a covariate were also common detections of correlative methods.HM-28 (GG) showed spatial autocorrelation in northwestern Uganda, overlapping an area with higher tsetse-fly loads; the mutations may be involved in parasite resistance.Source: §4 · pp. 8–9
• Spatially explicit models such as GWR are a possible next step, but tuning their weighting scheme restrains their use on very large datasets.The cited current R implementation of Guillot et al.’s model does not enable whole-genome analysis.Source: §4 · p. 9
• Bonferroni correction may discard adaptive loci in whole-genome sequencing analyses, and false-discovery-rate implementation in Samβada was still ongoing.Source: §4 · p. 9
• The cattle case study took 69 minutes for univariate models and 8.5 hours for univariate and bivariate models; single-process Samβada was about 4.5 times faster than LFMM and 30 times faster than BayEnv.The timings used one 3.1 GHz core and 8 GB of RAM.Source: §4 · pp. 9–10

5.1.1 Sampling design; 5.1.2 Molecular data

• Sampling covered Uganda’s eco-geographic regions with 917 animals from 202 farms; filtering left 804 samples and 40,019 autosomal SNPs.The samples were collected across a 51-cell grid, and the SNP data were filtered using 95% call-rate and 1% minimum-allele-frequency thresholds.Source: §5.1.1 · p. 10, §5.1.2 · p. 10

5.1.3 Population structure

• Ancestry analysis selected four populations, but almost 96% of samples fell into two clusters identified as Ankole and zebu.The remaining two clusters comprised 32 animals and possibly represented introgression from allochthonous gene pools.Source: §5.1.3 · pp. 10–11

5.1.4 Environmental data

• Environmental characterization combined 1-km WorldClim data, 90-m SRTM3 terrain measures, and coordinates; VIF-based screening retained 23 univariate and 15 bivariate predictors.An Ankole ancestry-membership coefficient was added for bivariate analysis, with correlation-based variable selection reapplied to limit VIF to 5.Source: §5.1.4 · pp. 11–12

5.2 Samβada implementation; 5.2.1 Desktop and High Performance Computing

• Samβada is a standalone C++ application, while its desktop and CoreSAM versions target smaller and larger datasets, respectively.The desktop version includes multivariate analysis and spatial autocorrelation; CoreSAM focuses on univariate analysis and computes up to seven times faster. The proposed workflow screens loci with univariate models before investigating spatial autocorrelation and multivariate models.Source: §5.2 · p. 12, §5.2.1 · p. 12

5.2.2 Modules

• Additional modules export spatial results, recode PLINK molecular data, and divide and merge workloads across computers.Supervision can split input data for separate processing, including on heterogeneous computers, then merge results into the output of a single full-data run.Source: §5.2.2 · p. 13

5.3 Alternative methods to detect selection; 5.3.1 BayEnv; 5.3.2 Latent Factor Mixed Models; 5.3.3 Outlier approach

• The comparison used BayEnv, LFMM, and Arlequin’s outlier method alongside Samβada, with approaches differing in their treatment of allele counts, genotypes, and population structure.BayEnv compares environmental models against a neutral-locus null model; LFMM accounts for structure through unobserved variables; Arlequin’s method uses FST outliers under an island-model assumption.Source: §5.3 · p. 13, §5.3.1 · p. 13, §5.3.2 · pp. 13–14

5.4.1 Software availability; 5.4.2 Data availability

• Samβada is open-source software distributed under GNU GPL 3, with compiled versions for Windows, Linux, and MacOS X.Ugandan cattle SNP data are identified as available in PLINK format through the NextGen data repository, subject to its data-use policy.Source: §5.4.1 · p. 14, §5.4.2 · p. 14

Supplementary material

• The supplementary-material passage identifies the paper as arXiv:1405.7658v2, dated 20 November 2014.Source: Supplementary material · p. 18

1 Spatial autocorrelation

• Moran’s I measures global spatial autocorrelation using deviations of point values from their mean and spatial weights between points.The passage defines n as the number of sampling points, wij as the neighbourhood weight, and zi and zj as deviations from the mean.Source: §1 · p. 18
• The weight matrix is usually normalized per line, with S0 equal to the sum of all weights.The section notes a case in which S0 equals n and the global equation can be rewritten.Source: §1 · p. 18
• LISA measures local association between a point’s value and neighbouring points, using normalized neighbourhood weights.Samβada computes a local Moran’s I for each point.Source: §1 · p. 18
• Summed LISA values are proportional to the global spatial-autocorrelation measure, Moran’s I.Source: §1 · p. 18

• Table S1 covers environmental variables used to detect selection signatures with correlative approaches.Source: Table S1 · p. 19
• Sambada is used for both univariate and bivariate analyses.Source: Table S1 · p. 19
• BayEnv and LFMM are listed for univariate analyses.Source: Table S1 · p. 19

• Individuals are grouped by population, with population labels arranged horizontally.Source: Figure S1 · p. 20
• Each individual’s population assignment is based on its highest membership coefficient, Qmax.Source: Figure S1 · p. 20
• Within each population, individuals are ranked by increasing or decreasing Qmax.Source: Figure S1 · p. 20

• Table S2 reports computation durations in hours and provides no runtime for Arlequin.Source: Table S2 · p. 20
• The article dataset includes chromosome X and uses SNP and individual call-rate filters of 5% and a MAF filter of 1%.Source: Table S2 · p. 20
• The second dataset contains 102 Ugandan cattle samples selected from 917 and uses BovineHD SNPs with 5% call-rate and MAF filters.Source: Table S2 · p. 20

• Table S3 lists Samβada loci with the highest G scores, identified by name, chromosome, and position in million base pairs.Source: Table S3 · p. 21
• The table indicates which methods detected each locus and counts the number of detections.Source: Table S3 · p. 21
• Bold loci are shared discoveries of Samβada, LFMM, and BayEnv; spatial autocorrelation indices were computed for lines 1 and 7.Source: Table S3 · p. 21