Neuroevolution of Multivalent Phage Libraries: Machine Learning-Driven In Vivo Selection for Tissue-Specific Targeting

Landscape phages (Gillespie et al.) display multivalent "elementary binding units" (EBUs) that evolve tissue specificity in vivo but lack predictive optimization. Ito et al. showed ML can extract functional binders from noisy phage libraries. I propose integrating these: deploy landscape phage libraries in animal models, sequence tissue-bound pools at multiple timepoints, and train a reinforcement learning (RL) model to predict EBU combinations maximizing tissue retention. The RL agent would iteratively design new libraries focusing on underexplored sequence motifs, accelerating discovery. This moves beyond traditional panning by treating in vivo evolution as an optimization problem solvable in silico. Unlike Bryant et al.'s EvoBind (2022), which evolves binders computationally, this closes the loop between real-world selection and AI design. It could generate clinical-grade targeting vectors for drug delivery or diagnostics in months, not years.

References:

1. Evolution of a Landscape Phage Library in a Mouse Xenograft Model of Human Breast Cancer. J. W. Gillespie, Liping Yang, L. D. De Plano, M. Stackhouse, V. Petrenko (2019). Viruses.
2. Selection of target-binding proteins from the information of weakly enriched phage display libraries by deep sequencing and machine learning. Tomoyuki Ito, T. D. Nguyen, Yutaka Saito, Yoichi Kurumida, H. Nakazawa, Saki Kawada, H. Nishi, Koji Tsuda, T. Kameda, M. Umetsu (2023). mAbs.
3. EvoBind: in silico directed evolution of peptide binders with AlphaFold. P. Bryant, A. Elofsson (2022). bioRxiv.