LatentDE: latent-based directed evolution for protein sequence design Enhancing protein function typically relies on directed evolution, where promising variants emerge by mutating an initial sequence and testing improved outcomes. Although widely used, these experimental screens can be time-consuming and expensive, prompting interest in machine learning approaches that search protein space more efficiently. However, operating directly in high-dimensional sequence space remains challenging, especially when only limited fitness data is available. Tran et al. introduce LatentDE, a method that builds on a variational autoencoder to embed protein sequences in a compact latent space. This representation is guided by the state-of-the-art ESM-2 language model, ensuring that functionally related sequences cluster together. The authors regularize the VAE by incorporating fitness signals directly during training, so that latent representations preserve relevant evolutionary patterns. After training, they apply a two-phase strategy that first uses gradient ascent to locate promising high-fitness regions and then performs a directed evolution-like search, but entirely within the latent space. This allows them to generate multiple variants, decode them back to protein sequences, and evaluate fitness with either wet-lab experiments or large ML oracles. By systematically testing LatentDE across eight protein design tasks, the researchers show that it surpasses various baselines in identifying sequences with enhanced fitness. The results highlight how latent space optimization leverages machine learning to overcome the discrete complexity of protein sequences, improving both accuracy and search speed. This approach underscores the promise of combining modern language models, generative architectures, and evolutionary concepts in pursuit of more efficient protein engineering. Paper: iopscience.iop.org/article/1…