Charting the evolutionary consequences of interspecific interactions

Charting the evolutionary consequences of interspecific interactions

From theory to inference

Bob Week  ·  KiTE Fellow  ·  Kiel University

Title. The faint chart behind the words is the map of slide 3.

The question

Opening schematic

0:45. How do interactions alter evolution, and how can we recognise their consequences in variation? The two halves of the question are why the talk has both theory and inference.

A landscape of interaction systems

Opening schematic

2:00. Sixty to ninety seconds, not a taxonomy lecture. Across: partners at the same scale, or one living in the other. Up: more partners. Down: a tighter, inherited partnership. The talk moves from the pair on the left to the systems on the right.

How I approach these questions

Opening schematic

3:45. Start from events among individuals (the close-up on the left); take the limit of many individuals to get a stochastic process for the populations: abundances and trait variation together. The snapshots are a toy predator and prey on a plane whose habitat keeps shifting; abundances cycle, and with trait matching the two mean traits chase each other (Red Queen). From the process, two things in parallel: analytical approximations that give foundational insights, and descriptions of measurable patterns. And already the two ways back: both point at the process, because what we infer is the evolutionary process, not individual outcomes.

Where we are going

Opening schematic

4:30. The route. Start with pairwise coevolution and the interface that decides what happens when two individuals meet. Then the community such interactions are embedded in. Then the spatial patterns that result. A pit stop to put those pieces together for inference. Then carry the lessons into host–microbiome evolution, and end with prospects for inference.

What is the interface between interacting species?

Part I · the interface published

5:15. The long-proboscid fly and the long-tubed iris. Pauw and colleagues’ data: the fly takes more nectar the longer its proboscis is relative to the tube; the flower receives more pollen the other way round. What each gets depends on the difference.

Different interfaces imply different selection

Part I · the interface published

6:45. Two ways to write that down. Trait differences: bigger is always better. Offset matching: best at a particular lead.

Same arms race, different ecological fate

Part I · the interface published

8:45. Both escalate. X benefits throughout in both cases. Under trait differences Y’s benefit falls through zero, so the interaction turns parasitic; under offset matching Y’s benefit shrinks to a third and holds there, well above zero, so it stays a mutualism. Coevolution is not itself a mechanism; the interface decides.

But the fly and the flower are not alone

Part I · community schematic

10:00. The fly is a keystone pollinator: it visits a guild of at least twenty long-tubed species. From here to the resource axis the nodes never move; each slide adds to or redraws the same picture.

But the fly and the flower are not alone

Part I · community schematic

10:45. It is also a nectar thief: it drains Babiana thunbergii, a sunbird flower, almost never touching anthers or stigma. A partner outside the arms race is exploited.

From pairwise interfaces to network architecture

Part I · community schematic

12:00. A segue, not a result. I have worked at this scale too (Week & Nuismer 2021; Nuismer, Week & Aizen 2018). The pair sits in a web of two guilds.

Now change perspective: interactions within a guild

Part I · community schematic

13:30. Look from one side. Pollinators that share plants compete for them.

Diffuse coevolution through competition

Part I · community schematic

15:00. The plant row becomes a resource axis; each species has a utilisation curve; overlap sets competition. This is the set-up of the many-species model.

What can we learn without tracking every interaction?

Part I · community schematic

17:00. Every link carries a competition coefficient and two selection gradients. Strong competitors impose strong linear selection on each other; which way they push, competition alone does not say. In a rich community these relations can be derived without tracking every pair.

Communities are distributed across landscapes

Part I · space schematic

19:00. The geographic mosaic (Thompson; Nuismer, Thompson & Gomulkiewicz): reciprocal selection is strong in some places and weak in others, who is there differs by place, and gene flow connects the local histories. Note the distance d marked between two places: that is where we go next.

One complication at a time: space, with two species

Part I · space schematic

20:30. Going beyond a pair was one complicating direction. Spatial structure is another, so take them one at a time: one host and one parasite, in many places at once. Their mean traits differ from place to place, more alike nearby. Now imagine reciprocal transplants between two places near each other, further apart, far apart. Would local adaptation look the same each time? Empirically it is often weak nearby and strong far away.

A model of coevolution in continuous space

Part I · space published

21:30. The model, in one slide. The local mean trait of each species changes by abiotic selection toward an optimum, by biotic selection (the host evades, the parasite follows), by dispersal, which is diffusion, and by random genetic drift, which is space-time white noise. The solution is a pair of random surfaces: a bivariate Gaussian random field, described fully by how alike two places are as a function of the distance between them.

Coevolution creates spatial scales of variation

Part I · space schematic

22:30. Each species’ mean trait is a surface with its own characteristic length, lambda: the distance over which mean traits stay alike. It is the dispersal distance over the root of genetic variance times abiotic selection. How much places differ depends on population size (drift), selection and dispersal. Biotic selection ties the two surfaces together.

Local adaptation depends on the scale at which we ask

Part I · space published

24:00. Three statements. Host and parasite local adaptation mirror each other. Which one is ahead is set by the two spatial scales: the species whose traits vary over the shorter distance; equal scales, neither. And with unequal abiotic selection and population sizes as well, which one is ahead can change with the distance at which we ask.

Can we run the map backward?

Inference schematic

26:00. The approach slide again, with the two ways back brought forward. From the patterns: link data to mechanistic models, by likelihood or by simulation. From the insights: biologically motivated approximations that keep that inference tractable. Both aim at the process.

Does a correlation mean coevolution?

Inference published

27:30. The same fly and flower. Each site on Pauw’s map gives one pair of means, and they are correlated. That was long read as evidence of coevolution. But Nuismer, Gomulkiewicz and Ridenhour (2010) showed that coevolution is neither necessary nor sufficient for such a correlation: one-sided selection or correlated abiotic optima produce it too. So the data have to be tied to a mechanistic model.

A mechanistic likelihood for coevolution

Inference published

29:00. Under the hood is the offset-matching model from before. Put that pairwise coevolution at many locations, each also under abiotic selection and drift, and the model predicts the spatial pattern of mean-trait pairs. That prediction, P(data | parameters), read as a function of the parameters is the likelihood (theta is the whole set of parameters): its peak is the estimate, its curvature the uncertainty. Coevolution is supported only if both one-sided models are rejected.

Coevolutionary selection in the fly and the flower

Inference published

30:30. Both strengths of biotic selection differ from zero: reciprocal, though far from balanced. And the consequence: proboscis and tube are well beyond where they would sit without it. Aside: the camellia and its weevil give the same verdict (backup slide).

When the likelihood is unavailable

Inference schematic

32:00. Sixty to ninety seconds. Strong selection, gene flow and varying optima break the likelihood. So simulate: draw parameters from priors, simulate data, keep the draws whose data are close to ours. The kept draws are the posterior; these two are a cartoon on simulated data. Analytical likelihood, then ABC, then modern simulation-based inference: one trajectory.

If we simulate, what should the simulator contain?

Inference schematic

33:00. If we build simulation-based approaches to inference, there is a lot we could include. For interspecific interactions in particular, we can look to theory for the key ingredients. (Do not name them; the next slide does.)

Three things keep appearing

Midpoint schematic

34:00. Interface, community, space. Not three projects: three recurring determinants of coevolved variation.

Host–microbiome evolution combines all three

Midpoint schematic

35:30. The same landscape as before, but now nested: localities; at each, a social network of hosts; in each host, a microbial community; underneath, a patchy environmental pool. The interface is genes with microbes; the partner is a community; and space comes three ways: social contacts, local pools, and hosts moving between localities. For one microbe this is a metapopulation; for many, a metacommunity.

…and adds inheritance and timescales

Midpoint schematic

Two things the pair never had: which microbial ancestry stays coupled to host ancestry, and the gap between microbial turnover and host genetic change.

What is the evolutionary interface now?

Part II · the interface published

37:00. However the trait is built from genes and microbes, what we measure is host genomes, microbiomes and the trait, host by host. Fitting them together partitions the variance: host genes, microbes, and their covariance. The covariance is genes and microbes occurring together, not a gene-by-microbe interaction. But explaining variation now does not tell us which of it is inherited.

Where does microbial ancestry run?

Part II · inheritance published

38:30. Ancestral concordance: three patterns read backwards from a realised history. Lineal: follows this host’s ancestors. Non-lineal: stays among hosts, but not these ancestors. Novel: from outside the host population. These are not transmission modes.

Which microbes can carry a response to selection on hosts?

Part II · inheritance published

39:30. What we would expect. Lineal microbes travel with the selected hosts’ line of descent, so they should respond. Novel microbes cannot, by definition. For non-lineal microbes it is not obvious.

Non-lineal microbes: passed on before selection, or after?

Part II · inheritance published

40:30. Whether non-lineal microbes help the response depends on how they are inherited relative to how selection works. Simplest case: are they passed to the next generation before selection on hosts, or after? Before: their variation carries no signal of selection, no response. After: it does.

From describing variation to explaining it

Part II · dynamics schematic

41:30. So far a statistical framework: it describes host variation in terms of genetic and microbiome variation and generates hypotheses about microbiome-mediated adaptation. To understand how that adaptation works, we need mechanistic models, starting simple enough to give clear first answers.

A simple model of microbially-assisted rescue of host populations

Part II · dynamics in prep.

42:30. One host allele, inherited (haploid); one microbe, caught by contact with carriers and lost again. Loss is clearance plus failed inheritance: at birth an offspring receives the microbe with probability v. Four kinds of host. Fitness in the same picture as the trait: the allele, the microbe’s direct effect, and their interaction.

Two routes to rescue

Part II · dynamics in prep.

43:30. Same total microbial benefit. A direct effect buffers the demographic crash; an interaction speeds the spread of the host allele.

Quasi-microbial equilibrium

Part II · dynamics in prep.

45:00. How a clean result came out of a messy process. In multilocus genetics, recombination is fast and selection slow; quasi-linkage equilibrium uses that to connect population genetics to the genetic variances of quantitative genetics. Here transmission and loss are fast: quasi-microbial equilibrium, and the hope is the same connection for microbiome-mediated quantitative genetics. Not tied to the rescue model: any model where microbes turn over faster than hosts evolve.

Next: mechanistic models that meet the variation we observe

Looking forward planned

46:00. Could QME take mechanistic models to observed host-genetic and microbiome variation? Only if the models also hold what else shapes that variation. Two directions. Where hosts live: prevalence differs among places before any selection on hosts, which alone creates gene–microbe associations. Whom they meet: with Aura Raulo and Guilhem Sommeria-Klein, the social network as a microbial metacommunity, where one contact moves several microbes at once.

Can we read interaction history from spatial molecular data?

Looking forward planned

47:00. Two open questions rather than a method. Are two species shaping one another, and where in their genomes? Is the microbiome changing host evolution, and through what? Both lean on spatial replication. What models and data make these mechanisms distinguishable? An invitation.

Charting interactions

Close schematic

48:00. Two ways to close. (1) The map as a summary: we started with a pair and an interface, added a community, then space; the same three reappeared inside a host with its microbes, joined by inheritance and timescale. Wherever we stood on this map, the question was the same: which structure decides what an interaction leaves behind, and can we read it back from data? (2) The map as an invitation: most of this chart is still coastline. The systems studied here in Lyon sit at particular places on it; I would like to know which of these coordinates matter in yours.

Thanks

Coevolution and space

Scott Nuismer · Gideon Bradburd · Stephen Krone

Microbiome-mediated quantitative genetics

Brendan Bohannan · Peter Ralph · Hannah Tavalire · William Cresko

Microbial rescue and QME

Hildegard Uecker · Hinrich Schulenburg

Social and spatial microbiomes

Aura Raulo · Guilhem Sommeria-Klein · Brendan Bohannan

The Schulenburg group for Evolutionary Ecology and Genetics, Kiel

KiTE · Marie Skłodowska-Curie 101081480 · Gordon and Betty Moore Foundation

Collaborators grouped by the story just told.

Backup: drift reorients G

Backup published

The single-species aside in full (Week 2026). Deterministic theory says drift only shrinks G; along a single path it drives the correlation to an extreme.

Backup: the fly and flower, to scale

Backup schematic

Body 16 mm, proboscis 65 mm, tube 60 mm (Pauw et al. 2009).

Backup: local adaptation, all nine cases

Backup published

Week & Bradburd 2024, Fig. 3, recomputed, beside their Fig. 4.

Backup: both systems

Backup published

Week & Nuismer 2019, Figs 3 and 4, with the camellia and its weevil.

Backup: the response over generations

Backup published

Week et al. 2025, Fig. 3, re-run: the response with more and more kinds of factor, non-lineal microbes taken from hosts after or before selection.

Backup: predicting the response

Backup published

Week et al. 2025, Fig. 5. Predictions match when exactly the transmitted factors are counted.

Backup: microbial geography before host selection

Backup planned

Gain and loss place by place; an association nobody selected for; a first sketch of what that does to the variance components.

Backup: social contact is microbial dispersal

Backup schematic

With Aura Raulo and Guilhem Sommeria-Klein. One contact moves several taxa, so taxa do not disperse independently.

Backup: the systems as a grid

Backup schematic

The earlier typology: scale by number of partners.