Exploring biotic signals in vegetation assembly from the LGM to the Anthropocene

BIODYNAMICS PROJECT

talking to <ORACLE>:

Ondrej Mottl

CSPE 2026 || 9 September 2026 || Prague, Czechia

NARRATIVE INTERFACE: ORACLE

…/…/…

ANALYTICAL OUTPUTS: REAL

… booting

System online

Greetings Dr. Mottl

I am Observational Runtime for Analysis of Community-Level Ecology

<ORACLE> for short

I am here to assist you in analyzing vegetation patterns on Earth

Awaiting ecological query.

Proceed? [Y]/[N]?

Is there a scale dependence in the amount of unexplained variation (potentially due to biotic interactions) structuring vegetation since LGM?

Query accepted

Plan to partition observed plant co-occurrence into:

spatial structure

climate response

species-species association (residual)

DATA: <VegVault>

USED SAMPLE COVERAGE

<VegVault> database: publicly available, open-source database

Community records, climate predictors, site coordinates, and functional traits are loaded as separate streams

Due to data availability, I will focus on the Northern Hemisphere of the planet since the LGM

Preparing community and climate streams

Data extracted, now preparing …

Community stream normalised

proportions filtered

Taxa are classified automatically against GBIF and filtered

Age uncertainty propagated to estimate counts

Climate stream screened

Redundant predictors are removed before model fitting

Now we need to process the data. I am only mentioning a small subset of the processing pipeline, and it is much more complicated in reality. For the fossil pollen community data, we need to transform pollen counts to proportions. Then we interpolate to an even temporal spacing of 500 years, propagating age uncertainty from the age-depth models. We will harmonise the taxonomy to align with the GBIF backbone. This will allow us to select the taxonomic resolution, like genus or family, used to aggregate the data. For the climate data, I used CHELSA bioclimatic variables, such as temperature, precipitation, and their variability. We remove all constant predictors, then check the correlation among them to create a non-collinear set. Finally, we interpolate to 500 years. All data are then filtered to make sure we have enough samples and cores for the modelling.

Model core: environment, space, association

\(Y_{ij} \sim \operatorname{Bernoulli}\!\left\{\Phi\!\left(\eta_{ij}\right)\right\}\)

\(\eta_{ij} = \alpha_j \;+\;\)

\(\sum_{k=1}^{K}\,\beta_{jk}\,x_{ik} \;+\; \sum_{k=1}^{K}\,\gamma_{jk}\,x_{ik}\,a_i\) \(\;+\;\)

\(\sum_{m=1}^{M}\,\delta_{jm}\,\operatorname{MEM}_{im}\) \(\;+\;\)

\(u_{ij}\)

\(\mathbf{u}_i = (u_{i1}, \ldots, u_{iJ}) \sim \mathcal{N}(\mathbf{0}, \Sigma)\)

Model assembled.

{sjSDM} as the modeling framework

Abiotic predictors explain shared response

Moran Eigenvector Maps (MEMs) absorb spatio-temporal autocorrelation

Residual covariance carries species-species association signal

Env: ~ (x1 + ... + xK) * age - age

Space: ~ 0 + (MEM1 + ... + MEMM)

Association: off-diagonal Sigma from sjSDM::bioticStruct()

Variance decomposition

Decomposition ready

Focus: residual association component

Report what remains after climate and spatial structure have made their claims

Caution: co-occurrence is not proof of interaction.

Three analysis axes

Spatial

Taxonomic

Temporal

Three routes selected:

Spatial-resolution runs : change with spatial scale

Taxonomic aggregation levels : change through classification

Temporal slice tests : change through time

▄█████ ▄▄▄▄   ▄▄▄ ▄▄▄▄▄▄ ▄▄  ▄▄▄  ▄▄       ▄▄▄▄  ▄▄▄▄  ▄▄▄  ▄▄    ▄▄▄▄▄
▀▀▀▄▄▄ ██▄█▀ ██▀██  ██   ██ ██▀██ ██      ███▄▄ ██▀▀▀ ██▀██ ██    ██▄▄
█████▀ ██    ██▀██  ██   ██ ██▀██ ██▄▄▄   ▄▄██▀ ▀████ ██▀██ ██▄▄▄ ██▄▄▄

Spatial results: local to continental

ORACLE, show us how the association signal changes going from continental to local. On this map we can see an animation of how that looks. For the continental scale, we use all data from the whole continent. For the regional scale, we use only a subset, and for the local scale we use an even smaller subset. Then we fit a separate model for each of them. Let’s take a look at the results. The X-axis is the spatial scale, from local to continental. The Y-axis is the amount of variance explained by the species-specific associations. Each point shows one model. We have only three for the continental scale - North America, Europe, and Asia. We have many more for the regional and local scales. The color of the point represents the relative contribution of each component based on the color triangle. Surprisingly, we do not see the species-specific signal getting stronger at the local scale. We can see that there is large variability, but the color is mainly yellowish, suggesting an overall strong impact of climate.

██████ ▄▄▄  ▄▄ ▄▄  ▄▄▄  ▄▄  ▄▄  ▄▄▄  ▄▄   ▄▄ ▄▄  ▄▄▄▄ 
  ██  ██▀██ ▀█▄█▀ ██▀██ ███▄██ ██▀██ ██▀▄▀██ ██ ██▀▀▀ 
  ██  ██▀██ ██ ██ ▀███▀ ██ ▀██ ▀███▀ ██   ██ ██ ▀████ 
                                                      
           ▄█████  ▄▄▄▄  ▄▄▄  ▄▄    ▄▄▄▄▄                        
           ▀▀▀▄▄▄ ██▀▀▀ ██▀██ ██    ██▄▄                         
           █████▀ ▀████ ██▀██ ██▄▄▄ ██▄▄▄

How does the spatial pattern change when we add taxonomic resolution?

Query accepted…

Adding taxonomic axis

Plotting the results

How does the spatial pattern change when we add taxonomic resolution? We assume that increasing the taxonomic resolution will reverse the effect of changing the spatial scale. This is exactly the same figure as before, but we have now added two more panels. The first shows what happens if we rerun the models but aggregate at the family level, and the last one is even more abstract: we replace the taxa with functional types, assigned using the functional traits of each taxon. Interestingly, changing the taxonomic resolution affects the patterns. The family level makes the spatial patterns even stronger, but I think this makes sense, as there is less importance of family-family association at the local scale. The functional types show that there is little interaction between functional types, which is also expected, as their patterns can be mainly explained by abiotic factors. In other words, their climatic niches are different.

██████ ▄▄▄▄▄ ▄▄   ▄▄ ▄▄▄▄   ▄▄▄  ▄▄▄▄   ▄▄▄  ▄▄    
  ██   ██▄▄  ██▀▄▀██ ██▄█▀ ██▀██ ██▄█▄ ██▀██ ██    
  ██   ██▄▄▄ ██   ██ ██    ▀███▀ ██ ██ ██▀██ ██▄▄▄ 
                                                   
         ▄█████  ▄▄▄▄  ▄▄▄  ▄▄    ▄▄▄▄▄                     
         ▀▀▀▄▄▄ ██▀▀▀ ██▀██ ██    ██▄▄                      
         █████▀ ▀████ ██▀██ ██▄▄▄ ██▄▄▄

Is the association signal stable through time?

Query accepted

Temporal mode selected: Slicing the data into 500-year windows

Network diagnostics loaded. Co-occurrence structure can change even when variance components look similar

Each slice receives an independent analysis and diagnostic workflow

Plotting the data distribution

Proceed? [Y]/[N]?

Temporal trajectories

Plotting temporal trajectories for each continent

Paleo predictions

Current SDM models can reconstruct past-to-present biodiversity patterns and prepare the same machinery for future projection experiments

OK, what now? We still have this powerful Species Distribution Model, so we can do what many modelers do, and that is prediction. We will start by predicting into the past. We have the climate data, the spatio-temporal structure, and the species-species association signal; now we will predict the distribution of a specific taxon - in our case, Picea. The blue color is the predicted probability of the taxon, and each tick is a different time slice. Yellow crosses show the actual detected presence of the taxon. We can see the northward migration of the taxon after the Last Glacial Maximum. We have this for all taxa, so we can also look at changes in richness. For me, the logical next step is to use this machinery to predict the future: use the past to better predict future changes in biodiversity.

Take-home messages

Consulting deeper reasoning matrices

Summarising RESULTS together: spatial patterns, taxonomic resolution, temporal dynamics

I have used 1,130 cores with 29,904 spatio-temporal communities, 256 taxa, and 8,977,115 trait values translated into 24 functional types to fit 263 models.

Palaeoecological data CAN be used to reconstruct past biodiversity patterns and provide insights into ecological processes.

SPACE: The association signal is NOT scale-dependent and NOT stronger at local than at continental scales

TAXONOMY: Increasing taxonomic resolution DOES have an impact on the spatial pattern of the association signal, but NOT in the expected direction

TIME: Co-occurrence structure IS responding to major environmental changes, but the association signal IS surprisingly stable through time.

In addition to contemporary single-species diversity models, palaeoecological data can be used to predict future biodiversity patterns and support conservation efforts.

… termination

Execution complete.

Turning off non-essential systems

Thank you for your attention

…

Presenter

Hi! I am Ondřej Mottl

Assistant Professor at 🏛️Charles University, Prague, 🇨🇿

Head of the 🧑‍💻Laboratory of Quantitative Ecology

Interested in macroecology, palaeoecology, biodiversity, and data science

Presentation availability

This presentation is publicly available on the BIODYNAMICS project website

bit.ly/BIODYNAMICS_project

PDF backup contains static figures only; use the HTML deck for animations

MIT LICENSE | Supported by Czech Science Foundation grant GN23-06386I

BIODYNAMICS PROJECT

talking to <ORACLE>:

Exploring biotic signals in vegetation assembly from the LGM to the Anthropocene

Ondrej Mottl

CSPE 2026 || 9 September 2026 || Prague, Czechia

NARRATIVE INTERFACE: ORACLE

…/…/…

ANALYTICAL OUTPUTS: REAL

                                 
██████ ▄▄ ▄▄ ▄▄▄▄▄▄ ▄▄▄▄   ▄▄▄   
██▄▄   ▀█▄█▀   ██   ██▄█▄ ██▀██  
██▄▄▄▄ ██ ██   ██   ██ ██ ██▀██  
                                 
 ▄▄▄▄ ▄▄    ▄▄ ▄▄▄▄  ▄▄▄▄▄  ▄▄▄▄ 
███▄▄ ██    ██ ██▀██ ██▄▄  ███▄▄ 
▄▄██▀ ██▄▄▄ ██ ████▀ ██▄▄▄ ▄▄██▀

BIODYNAMICS PROJECT

… booting

Is there a scale dependence in the amount of unexplained variation (potentially due to biotic interactions) structuring vegetation since LGM?

DATA: <VegVault>

Preparing community and climate streams

Model core: environment, space, association

Variance decomposition

Three analysis axes

Spatial results: local to continental

How does the spatial pattern change when we add taxonomic resolution?

Is the association signal stable through time?

Temporal trajectories

Paleo predictions

Take-home messages

… termination

Presenter

Presentation availability

BIODYNAMICS PROJECT

Results of contemporary data

Is it a issue of model quality?

Is there relationship between species-species association and network modularity?

Is there relationship between network modularity and network size?

How good predictor is the climate?

How are the Functional Types generated ?