Thread by @SeppeKuehnLab

Second paper! Learning the community-function landscape via regression -- punchline: we can *predict* emergent functional properties of microbial consortia using only knowledge of strain presence and absence. biorxiv.org/content/10.110…

work I presented at #apsmarch in collab with @mikhtikh @asanchez_lab @jdiazc9 Abby Skwara, @KarnaVGowda , Arjun Raman, Mahmoud Yousef.

Communities exhibit emergent functions - pathogen suppression, compound degradation or production. Given the massive complexity of consortia - how can we understand/predict these emergent properties of communities?

One path forward is to try and dissect interactions in communities, build detailed dynamical models, try and make predictions. It is possible to do! (best example is work by @oventurelli2) It is hard and time-consuming. Is there another way?

Inspired by the idea of fitness landscapes we've been exploring whether these concepts from protein design might help. Perspective piece here - cell.com/cell-systems/p…. The idea is to construct 'community-function' landscapes where strain presence/absence is analogous to

mutation presence/absence. Here the high-dimensional landscape describes a mapping from community composition to the emergent function of interest. In this study we looked at 6 datasets and asked whether we could learn this landscape simply by regression.

Remarkably, we can! ... and predictions are quite good, here are two examples. Model is *linear regression* to second order (e.g. pairwise terms).

Why does this work? It turns out these community-function landscapes are not that rugged -- epistasis is rare and generally weak relative to additive terms. Here we quantified the contribution to the fit of first, second, higher order terms in the model..red curve:

What this means is that we use *only knowledge of which strains are in the community* to predict function. No dynamics necessary. Oddly, we've been making exactly this crude assumption in genetics for >100 years, but now we see how useful it can be to understand communities.

Does it always work? Of course not! @mikhtikh studied a wide range of models (Lotka-Volterra, consumer resource) and he can "break it" by putting in processes like very long trophic chains where function depends on presence of many species.

So what does it mean? What are the limits? We find that these landscapes are navigable. Means we can rationally "design" synthetic consortia for target functions. Limitations: these designs are likely fragile to environmental perturbations, we need synthetic communities (so

working from isolates). We also need to understand more about *why* this works.

Incredible work with amazing friends and collaborators: especially Abby Skwara at Yale, @KarnaVGowda , @mikhtikh, @asanchez_lab @jdiazc9 who were a pleasure to discuss this with every week.