Research

From so simple a beginning endless forms most beautiful and most wonderful have been, and are being, evolved.

Charles Darwin, On the Origin of Species, 1859

The forms are wonderful. They are not endless.

Thirty binary characters already describe more than a billion combinations, far more than the number of species alive today, and almost none of those combinations has ever been built. Allow the description to include size and shape and the gap stops being countable at all. Against a space of possibility that really is endless, life has taken one narrow and uneven path. What determines that path?

Two things about it are worth noticing.

The first is that the path is not uniform. Realized forms cluster, and some regions stay empty while their neighbours fill. Traits that depend on one another are part of the reason, since the state of one character changes which states of another a lineage can reach. Constraint of that kind does not merely thin the occupied region, it gives the region a shape.

The second is that the phenotypes worth explaining are high-dimensional. A form, a spectrum, or a photograph carries far more of what an organism is than a handful of measurements do. Describing organisms that fully raises a question the older methods never had to answer: in a space of shapes, what does it mean for two organisms to be close together, or for a lineage to have travelled a long way? Questions about closeness and distance are questions about geometry, and they return us to the shape the first observation pointed at.

The three lines below follow from those two observations.

Modelling dependencies

Dependencies are the part of the structure we can write down. Traits are bound to one another, to the environment, and to the genotypes that build them. Analyses that treat characters separately miss this, and they can credit the wrong cause for a change. I build models that state the dependencies explicitly. corHMM fits several discrete characters at once, including characters we cannot observe directly, which reduces the false correlations that appear when traits are fitted one at a time.

Plant life history shows what this buys. Annuals wait out bad seasons as seed, so it is natural to ask whether climatic extremes favour that strategy. But life history also shapes which climates a lineage can occupy, and separating the two requires fitting them together. I wrote hOUwie to do this, joining an Ornstein-Uhlenbeck process to a hidden Markov model. Across 32 angiosperm clades and eight climate variables, the maximum temperature of the warmest month was the most consistent predictor of annual life history. The same reasoning applies to antibiotic resistance, where each genetic background determines which mutations come next; with the Michigan Infectious Disease Genomics Center I used phyloAMR to sample the full range of plausible histories rather than settle on one, and found seven distinct routes to resistance in Klebsiella pneumoniae.

Measuring phenotypes at scale

Claims about which forms exist are empirical claims, and testing them takes measurements that stay consistent across thousands of species. Museums and citizen-science platforms have already produced the images. Measuring them is the slow part, because it is laborious and depends on who is doing it.

I use machine learning to turn those images into standard measurements. With support from OpenAI I built pipelines that read flower colour from tens of millions of iNaturalist photographs, covering roughly 80,000 flowering plant species, a trait that had been studied in only a handful of systems because collecting it by hand was impractical. Language models handle colour and description well, but shape needs something more exact, so I am building landmarking tools with convolutional networks. Those now measure head, mandible, and body length across more than 270,000 ant specimens, and intertegular distance, which tracks foraging ecology, across 550,000 bee images. Two R packages came out of this work: SegmentR for segmenting specimen images, and mvh for assembling virtual herbaria.

Recovering geometry

Colour and length are easy to extract partly because they already suit the models we have, in which a phenotype is a point in a low-dimensional space and evolution moves it around. Shape, sound, and spectral reflectance do not reduce that way. Measuring them does not tell us how to compare two of them, or how far apart they lie. Until that question is settled, a pattern we report may belong to our representation rather than to the organism.

The problem appears in miniature in model selection. As characters accumulate, the number of possible models grows faster than we can search it, so I wrote corHMMDredge, which borrows regularization and parameter sharing from statistical learning to keep the search tractable. Applied to mating system and estrus signalling in primates, the best model confined change to a one-dimensional path through a much larger space. Over the next several years I want to build a geometric framework for comparative biology that describes the surface phenotypic evolution actually moves on, and asks whether estimated rates change once we measure them on that surface. The aim across all three lines is a comparative biology that can tell the structure we impose from the structure evolution leaves behind.