Working with UVM computer scientist Lapo Frati and colleagues at the University of Vermont and the University of Cambridge, Petak helped design a computational study that followed thousands of generations of digital organisms evolving under many different changing environments. Their analysis, published December 15 in the Proceedings of the National Academy of Sciences (PNAS), shows that the outcomes vary: in some scenarios, environmental change helps populations reach higher fitness peaks, while in others it prevents them from doing so.

The team notes that traditional experimental evolution often tracks a single population in one environment and then generalizes the findings to entire species. In contrast, this work examined many environments and many populations to test how specific patterns of variability shape evolutionary trajectories, highlighting that two populations of the same species facing different types of environmental cycles can follow very different adaptive paths.

Petak illustrates this point with an example involving fruit flies. One population in the United States might adapt to temperature swings between seasons, while another in Kenya might face alternating dry and wet periods. Temperature fluctuations could support adaptation to both hot and cold conditions, she explains, but repeated switches between drought and rainfall might interfere with adaptation to dry conditions by forcing the population to repeatedly reset its evolutionary response after long rainy intervals.

Senior author Melissa Pespeni, a biology professor at UVM, emphasizes that the study's scale allowed the team to replay evolution hundreds of times across many distinct environments. "What's exciting about this study is that we replayed evolution hundreds of times. This gave us a bird's-eye view of how evolution played out across many different environments - something that would be impossible to test in the lab," said Pespeni. "The biggest takeaway for me is that starting point really matters. A population's history shapes how high it can climb and how hard the path is to get there, which means we can't assume one population represents an entire species."

The authors point to several areas where these results could matter. Biologists want to know whether species can adapt quickly enough to keep pace with global climate change, and public health researchers track how bacteria evolve resistance to antibiotics through repeated exposures. Yet many studies still rely on observations from single populations in single fluctuating environments and then draw broad conclusions about the species as a whole. Petak says computational models like the one used in this study can help generate new hypotheses about how real populations might respond under different environmental regimes.

To build the model, the UVM team created artificial organisms and placed them in digital environments that mimicked alternating conditions in nature, including hot - cold cycles and drought - rainfall patterns. Petak notes that a key advance was the decision to construct 105 different variable environments instead of focusing on just one pattern of change. This design allowed the researchers to systematically compare how populations evolved across many distinct scenarios and to identify cases where environmental variability either promoted or constrained access to higher fitness peaks.

The work also connects to questions in artificial intelligence and machine learning. AI systems often have difficulty learning new tasks without losing performance on tasks learned earlier, a challenge that has driven interest in "online continual learning" approaches. UVM computer scientist and co-author Nick Cheney sees parallels between these AI efforts and the team's findings on evolution in dynamic environments and argues that the way evolutionary systems respond to changing conditions can inform how AI models are trained to learn continuously over time.

For Frati, the project links directly to his research on meta-learning, the ability of systems to "learn to learn." He notes that judging an AI system's learning capacity based on a single subject or task can be misleading, just as assessing evolvability from a single environment can miss important dynamics. In both cases, exploring performance across multiple, diverse yet comparable environments provides a more accurate picture of how well a system can adapt and improve.

The study's central conclusion is that both the starting conditions and the nature of environmental variability strongly shape evolutionary outcomes. Petak notes that different histories and different patterns of change can send populations along very different paths, even when they belong to the same species. "Our results show that the choice of variable environment," she says, "can strongly influence the outcome."

Research Report:The variability of evolvability: Properties of dynamic fitness landscapes determine how phenotypic variability evolves

Related Links
University of Vermont
Lands Beyond Beyond - extra solar planets - news and science
Life Beyond Earth