Evolutionary Curriculum Learning Enhances Biological Sequence Modeling
A recent preprint on arXiv (2608.00697) presents a novel approach called Evolutionary Curriculum Learning (ECL), designed for training variational autoencoders (VAEs) in the context of biological sequence modeling. Unlike traditional VAE training, which treats sequences as interchangeable and overlooks evolutionary connections, ECL utilizes evolutionary frameworks by gradually introducing sequences that are progressively more distanced from selected anchors, adhering to a power-law expansion model. When implemented in two VAEs—EVE for predicting protein variant effects and RfamGen for generating RNA family sequences—ECL enhances performance in downstream tasks across five random seed configurations. This method addresses significant challenges in generative models for biological sequences, with implications for disease prediction and RNA design. The research team has made the paper available on arXiv.
Key facts
- Paper arXiv:2608.00697 introduces Evolutionary Curriculum Learning (ECL).
- ECL is a training strategy for variational autoencoders (VAEs) on multiple sequence alignments (MSAs).
- Standard VAE training treats sequences as exchangeable, ignoring evolutionary structure.
- ECL progressively exposes the model to sequences of increasing evolutionary distance from anchors.
- ECL uses a power-law expansion schedule.
- ECL was applied to two VAEs: EVE (protein variant effect prediction) and RfamGen (RNA family sequence generation).
- ECL improved downstream task performance across five random seeds per configuration.
- The paper is available on arXiv.
Entities
Institutions
- arXiv