Improving de novo protein binder design with deep learning

Abstract

Recently it has become possible to de novo design high affinity protein binding proteins from target structural information alone. There is, however, considerable room for improvement as the overall design success rate is low. Here, we explore the augmentation of energy-based protein binder design using deep learning. We find that using AlphaFold2 or RoseTTAFold to assess the probability that a designed sequence adopts the designed monomer structure, and the probability that this structure binds the target as designed, increases design success rates nearly 10-fold. We find further that sequence design using ProteinMPNN rather than Rosetta considerably increases computational efficiency.

Fig. 1. Monomer and protein complex structure prediction metrics distinguish previously designed binders from non-binders.

a For binder design to be successful, the designed sequence must fold to the designed binder monomer structure (left), and this structure must form the designed interface with the target protein (right). b, c Design failure modes. b Type-1 Failures. The designed sequence does not fold to the designed monomer structure. c Type-2 Failures. The designed sequence folds to the designed monomer structure but does not form the designed interface. d, e The retrospective experimental success rate (YSD SC₅₀ < 4 μM) for the top 1% of designs selected according to different monomer (d) or protein complex (e) based metrics over 10 targets from Cao et al. Source data are provided as a Source Data file.

Fig. 2. Incorporation of structure prediction metrics increases design success rate on new targets.

a Results of Prospective Campaigns. For each target the SC₅₀ from YSD is shown for all designs which showed binding by YSD (like Kd’s, lower values are better). The number of designs included in each library for each target is indicated by the bars in the top panel. The AF2-predicted structure of the top scoring on-target design is shown as a cartoon. No binders were identified to Site 2 of IL2 receptor-ɑ so this campaign is not included here or in panel C. b The experimental success rate for libraries filtered by DL-based filtering versus Physically based filtering for the four prospective targets. c The computational efficiency (the number of designs with pAE_interaction <10 per CPU-s) for the ProteinMPNN sequence design plus Rosetta relax protocol outperforms that of the original Rosetta sequence design protocol. Source data are provided as a Source Data file.