Machine learning has quantified the evolution of fish schooling across 2,839 species — nearly 10% of the 39,000 living fish species — and found the behavior was independently gained some 60 times and lost some 320 times over 350 million years, a Cornell University team reports in the Proceedings of the National Academy of Sciences on Nov. 6.
The finding reverses a core assumption in behavioral evolution. Experts had long assumed lineages rarely abandon collective movement once they acquire it. The data says otherwise.
“If you had asked me or other evolution of behavior experts before this study, we’d tell you that losses of collective behavior are probably very rare, but that’s simply not true,” said Andrew Hein, associate professor of computational biology in Cornell’s College of Agriculture and Life Sciences and senior author of the study.
Text transformers read the field notes
It was data. Comparative studies of collective behavior — fish schools, bird flocks, even stadium waves — each covered one or two species, adding up to a few dozen.
Working under first author Jay Love, a postdoctoral researcher in Hein’s lab, the Cornell group applied a machine learning approach called a text transformer — the same approach that powers large language models — to process studies and species descriptions and translate them into numerical representations the researchers could then use to gauge whether a given species might participate in collective movement.
When it came to images, the researchers applied a set of machine learning methods to pictures and footage, transform them into three-dimensional form, and evaluate which way each fish was swimming, then compare each fish’s alignment with that of others to establish whether they were traveling together.
Roughly 10% of all living fish species now have a quantified phenotype for schooling. The new paper, published Nov. 6 in the Proceedings of the National Academy of Sciences, ranks among the biggest investigations into the evolution of any behavior ever carried out across any group of species, the authors say.
Safety in numbers, tested at scale
The dataset let the team test two long-standing hypotheses that had never faced broad data. First: schooling exists for predator avoidance. The prediction — prey low in the food chain should school; predators should go it alone — held, but only loosely.
“There, in the data, species that are lower in the food chain are more likely to engage in collective behavior,” Hein said. But the correlation is not ironclad, he added: plenty of low-trophic species stay solitary, and plenty of predators school.
Second hypothesis: schooling helps individuals find scarce resources. “It would make sense to follow your neighbors around if they are eating the same thing,” Love said. Here the pattern was sharper — Hein said that predators moving in groups virtually all eat prey that themselves move in groups.
The 320 losses are the real story
The asymmetry between the roughly 60 times fish lineages gained schooling and the roughly 320 times they lost it is what should reframe the field. Gains spread through inheritance to descendant species; losses mean a schooling lineage repeatedly left solitary daughter species. “There were hundreds of examples across the tree where a lineage that engaged in collective movement left a daughter species that was solitary,” Hein said.
Why? The researchers don’t know, but they think clues may lie in drawbacks to collective movement. Fish that swim together are more vulnerable to misinformation and information echo chambers — for instance, when a jumpy fish bolts on a false alarm, setting off a scatter among the whole group.
Competition is the second: if food is limited, moving as a group and sharing may not be an advantage. Disease risk is the third, on longer time scales.
“If you live in groups, you’re around lots of other individuals that are susceptible to similar diseases as you are, and that could be exerting some sort of major evolutionary force,” Hein said.
Schooling lineages speciate faster
Another notable finding from the PNAS paper: lineages in which every species exhibited schooling behavior proved far more diverse, producing new species at a faster pace than lineages that lacked schooling.
“We don’t know if the speciation rate is higher or the extinction rate is lower, but we know that through a combination of those two things, we have more species being generated,” Love said.
Why this matters beyond ichthyology
Text transformers and computer vision pipelines extracting behavior estimates from field literature and crowd-sourced photos turn an unquantifiable trait into a dataset.