Abstract
Ancestral state reconstruction (ASR) is a foundational tool in comparative biology, offering insights into the evolutionary history of lineages. With each new evolutionary model, our ability to estimate ancestral states has improved alongside the increased biological realism of these models. However, the field has primarily relied on reconstructions that focus on individual nodes, known as marginal reconstructions. This framework is analytically tractable but may not accurately represent what biologists want in inference, as evolution is dependent, and phenotypic transitions deeper in time can lead to consistent changes later. We argue that evolutionary history is better represented by joint reconstructions, which estimate the full sequence of states across nodes. Traditionally, joint reconstruction algorithms only estimated the single most likely sequence, but here we develop novel algorithms to estimate all relevant ancestral histories efficiently and provide tools to quantify the uncertainty of joint ASR. Furthermore, through simulations and an empirical case study, we demonstrate that joint reconstructions have higher accuracy than their marginal counterparts, and that the uncertainty surrounding the best joint reconstruction can be biologically meaningful and summarized using novel clustering algorithms. We apply our methods to epidemic multidrug-resistant Klebsiella pneumoniae and find that the evolution of antibiotic resistance is not a single narrative but a series of competing histories. Each of these histories exhibits distinct phenotype-genotype transitions that traditional approaches would struggle to identify, yet have critical implications for predicting resistance evolution.