Extended Data Figure 3: Temporal classification of samples, assemblage variation in time and data robustness.

a–d, K-means clustering of permafrost plant assemblages. a, Cluster identity of samples derived from pre-LGM, LGM and post-LGM periods for values of K between 2 and 10. Each bar represents a separate sample; different colours reflect different cluster identities. b, The Calinski–Harabasz criterion for different levels of K. Higher values indicate stronger support for a level of partitioning. c, d, Heat maps showing the proportional occurrence of samples from pre-LGM, LGM and post-LGM periods in different clusters, for K = 2 (c) and K = 3 (d). Colours vary from red (low values) to white (high values). e–g, Assemblage variation in time and space. e, Nonmetric multidimensional scaling (NMDS) ordination revealed significant variation (PERMANOVA, P < 0.01) in fossil/ancient plant assemblage composition during the three palaeoclimatic periods. f, The effect of spatial distance on similarity when assemblages from different palaeoclimatic periods were compared. The vertical axis represents similarity in floristic composition measured as 1-Bray–Curtis similarity, the horizontal axis depicts ln of distance between sampled communities in kilometres. The greater the spatial distance between pairs of assemblages, the more dissimilar they were. However, the rate of the decay differed depending on which two climatic periods were compared (full model P < 0.001). The weakest distance decay in similarity was observed in the case of comparisons between pre-LGM and post-LGM assemblages. Even if pre-LGM and post-LGM samples came from the same geographic area, their floristic compositions were dissimilar. g, Results of randomization tests. Mean proportional composition of different growth form types in LGM and post-LGM samples. The bars around sample means indicate 95% quantiles derived from 999 bootstrap replicates (where bootstrap N was set to the number of samples in the post-LGM data set; see methods for details). h, Counts of MOTUs exhibiting different growth forms binned over 5-kyr time intervals. The analysis included 218 of the 242 sediment samples, as described in Fig. 4. Numbers immediately below the columns indicate sample sizes. Median (central dot), quartile (box), maximum and minimum (whiskers) counts are shown.