Rational Prediction of PROTAC-compatible Protein-Protein Interfaces by Molecular Docking

Abstract

Proteolysis targeting chimeras (PROTACS) are ligands that can mediate interaction between a protein target and a E3 ligase, forming a ternary complex bound to the ubiquitination machinery, leading to target degradation. This technology has become an exciting new avenue for therapeutic development, with currently two PROTACS in clinical trials targeting cancer. Nonetheless, several PROTAC drug discovery campaigns still rely on serendipity. Here, we describe a general and efficient protocol which combines restraint-based LightDock, energy-based rescoring and minimal solvent-accessible surface distance filtering to produce PROTAC-compatible PPIs. Benchmarking our protocol using a manually curated dataset of 16 ternary complex crystals, accuracies of 94% and 70% were achieved on the redocking and realistic docking experiments starting from unbound protein structures, respectively, evaluated using the CAPRI standards and DockQ. Our protocol is both accurate and computationally efficient, with potential to accelerate PROTAC drug design campaigns, particularly when the ternary complex or PROTAC are unknown.

AlphaFold-Multimer struggles in predicting PROTAC-mediated protein-protein interfaces

“…While some groups have developed tools for this purpose [32][33][34][35][36] , they are typically not general because they require that the PROTAC molecule be known a priori. Within our group we developed PROTACability 25 , which by-passes this constraint and achieves satisfactory accuracy but fails to produce high quality solutions. Thus, there is significant room for improvement and a general tool that uses minimal information and achieves high prediction quality would significantly impact the field of PROTAC-based drug discovery.…”

“…DockQ: To evaluate the similarity between the experimental structures and the predicted structures from AF-Multimer, the DockQ criteria were employed 28 . In short, DockQ provides a continuous score from 0 to 1 (with 1 being perfect similarity) which takes into account the fraction of conserved native contacts, RMSD of the target protein and the interface RMSD between a reference structure and a predicted structure 25 . This metric can be employed either quantitatively or qualitatively, with possible classifications being inaccurate (0-0.229), acceptable (0.23-0.49), medium (0.50-0.799) or high quality(0.80-1) predictions.…”

“…We have reported in another work 25 that PROTAC-mediated interfaces are typically shallow and small and thus, represent a significant challenge to machine learning-based approaches such as AF2. In this letter, we evaluate the accuracy of AlphaFold-Multimer in predicting the structure of heterodimeric protein-protein complexes.…”

Biophysical and Computational Approaches to Study Ternary Complexes: A ‘Cooperative Relationship’ to Rationalize Targeted Protein Degradation

“…Alternative tools applied in TPD include PatchDock, [96][97] FRDOCK, [98][99] MOE packages, [100][101][102][103] ClusPro, [104] and LightDock. [105] The in silico determination of ternary complexes calls for a thorough understanding of protein-protein conformations. However, resolving ligand (PRO-TAC) geometry is also needed to create accurate models, as discussed in the following section.…”

Paralogue-selective degradation of the lysine acetyltransferase EP300

“…We take care not to overstate this observation, as recent analyses have indicated that current versions of AlphaFold are not useful for predicting the final conformation of E3 ligase/ternary complexes formed by induced proximity reagents. 20 However, the potential for differential molecular recognition, together with slight biochemical selectivity of CPI-1612 for EP300, motivated us to explore this HAT ligand's degrader capabilities.…”