ICML 2026 Groundbreaking Research: Nested Birth-Death Processes Challenge Neural Networks in Protein Evolution Modeling

A study accepted by ICML 2026 shows via its open-source implementation that nested birth-death processes in traditional probabilistic models can compete with deep learning in time-series modeling of protein evolution. This research challenges the dominance of neural networks in this field and provides a new perspective for computational biology. Its open-source code is maintained by Annabel Large (GitHub link: https://github.com/AnnabelLarge/protein_evolution_icml_2026) and was released on 2026-05-29. Core advantages include higher computational efficiency and stronger interpretability, among others.

Protein evolution is a complex time-dependent process, and understanding the impact of mutations on structure and function is a core problem in computational biology. Traditionally, probabilistic models were used to describe amino acid substitutions, but neural networks have dominated in recent years. However, neural networks have drawbacks such as requiring large amounts of training data, high computational costs, and lack of interpretability, prompting the search for more concise and efficient alternatives.

The study found that nested birth-death processes (a classic stochastic process model) can effectively compete with neural networks in time-dependent protein evolution modeling. This model simulates changes in amino acid sequences as birth-death events (addition/disappearance) in state space, rather than the traditional substitution matrix approach, providing a new methodology for the field.

The open-source implementation of the study uses Jax and PyTorch in collaboration: core computations are done with Jax (leveraging automatic differentiation and JIT compilation for efficient numerical calculations), and the data loading pipeline is built based on PyTorch (ensuring compatibility with the deep learning ecosystem). This architecture lowers the barrier to reproduction/extension, balancing high-performance computing with experimental reproducibility.

The significance of this study goes beyond the field of protein evolution: it reminds us not to overlook the potential of classical statistical methods, and that understanding the structure of a problem may be more important than model capacity. Impacts on bioinformatics include: higher computational efficiency, strong model interpretability, deep mathematical connections with population genetics theory, and possible emergence of hybrid architectures in the future.

The study has its scope of application: the effectiveness of nested birth-death processes depends on specific data characteristics and evolutionary assumptions, which need to be verified in collaboration with domain experts. Additionally, the study focuses on time-dependent evolution modeling and does not cover tasks such as protein structure prediction or functional annotation (where neural networks like AlphaFold still dominate).

This study provides an opportunity for reflection on computational biology methodologies, proving that classical probabilistic models still have unique value in the deep learning era. The open-source implementation provides resources for the community, and we look forward to more research emerging on the trade-offs between traditional models and neural networks in bioinformatics.