On the Promise (and Practice) of Deep Comparative Methods: Expressivity, Limitations, and Design

Caleb Charpentier , Arthur Porto, Anuj Karpatne, Josef C. Uyeda

Comparative biologists explain the diversity of organismal form and function by deciding which traits to measure and which evolutionary processes to model. These decisions make inference possible, but they also confine it to traits, relationships, and processes that can be specified a priori. Neural networks can expand this space, learning characters directly from complex observations and supporting inference under models without tractable likelihoods. This expressivity comes with a characteristic cost. Neural networks learn from statistical regularities in their training data rather than from the entities and processes biologists use to formulate explanations. This difference reflects a deeper mismatch in reasoning: neural networks generally extend statistical relationships learned from existing data to new observations (induction), whereas comparative biologists often seek the evolutionary processes that best explain how observed patterns arose (abduction). As a result, learned representations can rely on incidental features, fail under distribution shift, or admit many statistically equivalent but biologically distinct interpretations, which post hoc explainability methods cannot by themselves resolve. We argue that this mismatch is a design problem rather than an intrinsic limitation. Architectures, loss functions, training data, and optimization procedures each offer points at which researchers can impose abductive constraints: biologically motivated restrictions on the invariances, scales, and structure of learned representations that make them commensurable with evolutionary theory. Simulations, benchmarks, and targeted validation then test whether those constraints succeeded. We call neural networks designed and evaluated in this way \textit{Deep Comparative Methods}. Their promise lies not in eliminating the need to define biological meaning, but in expanding the range of candidate measurements and explanations that can be made explicit and empirically tested.