De Novo Design of Protein-Binding Peptides by Quantum Computing

TL;DR

This work advances de novo peptide design by integrating a physics-based, multi-scale framework with quantum annealing to explore the combined chemical and conformational space without relying on training data. A coarse-grained lattice model and Miyazawa-Jernigan–type energetics are encoded into a QUBO and solved on a D-Wave hybrid quantum-classical solver, followed by all-atom docking to refine poses. Validation via structure- and sequence-based analyses shows designed peptides reasonably recapitulate native contacts and correlates with experimental binders, while quantum optimization provides a diverse set of low-energy, high-affinity candidates comparable to state-of-the-art classical solvers. The results demonstrate that current quantum technologies can augment physics-based drug design, with potential for scaling to more complex targets and to small-molecule design, receptor flexibility, and ADMET considerations.

Abstract

In silico de novo design can drastically cut the costs and time of drug development. In particular, a key advantage of bottom-up physics-based approaches is their independence from training datasets, unlike generative models. However, they require the simultaneous exploration of chemical and conformational space. In this study, we address this formidable challenge leveraging quantum annealers. Focusing on peptide de novo design, we introduce a multi-scale framework that integrates classical and quantum computing for atomically resolved predictions. We assess this scheme by designing binders for several protein targets. The D-Wave quantum annealer rapidly generates a chemically diverse set of binders with primary structures and binding poses that correlate well with experiments. These results demonstrate that, even in their current early stages, quantum technologies can already empower physics-based drug design.

Figures (6)

• Figure 1: Schematic depiction of the workflow. I: A coarse-grained peptide connecting regions A and B is generated via (quantum) minimization of the problem Hamiltonian (Eq. (\ref{['eq:Hamiltonian_all_parts']})). II: The configuration is frozen and the sequence is optimized in higher resolution by a second (quantum) minimization step. III: A classical molecular docking simulation predicts the all-atom off-lattice representation of the previously generated sequence.
• Figure 2: The area under the precision-recall curve for docking simulations of the PDB entries (a) 3BFW, (b) 4DS1 and (c) 3BRL. For each simulation, the scores of 30 random peptides (red) were compared to the previously removed peptide (orange), a generated peptide optimized without accounting for the sequence free energy (black), and a generated peptide optimized accounting for it (blue). A higher score indicates better binding to the pocket in the framework of the docking validation (see main text and Section \ref{['SM:docking_validation']} of the SM). The redocking pose and the generated molecule's docking pose for 3BRL are shown in (d).
• Figure 3: Comparison of the top 50 generated binders with 111 experimentally known binders of the LC8 hub protein. The eight histograms are corresponding to the eight positions in the anchor motif of LC8. For each position, the frequency of amino acid types in the generated binders (blue) and in the LC8 hub LC8_binding_proteins_database (orange) are displayed. The amino acids have been clustered into five groups as described in Section \ref{['Section:CG_model']}.
• Figure 4: MEV obtained generating binders for the PDB entry 3BRL using classical and quantum optimization. Orange points: Lowest (circles) and average (triangles) MEV obtained in 1000 Gurobi runs as a function of the runtime of an individual simulation. Horizontal blue lines: lowest (dashed) and average (dotted) MEV obtained in 300 D-Wave runs with $5\,\text{s}$ of hybrid annealing time. For a fair comparison, when running classical minimization with Gurobi, we encoded the conditions imposed by $H_{\text{anc}}$, $H_{\text{occ}}$, and $H_{\text{path}}$ as hard constraints.
• Figure 5: Spectrum of MEV obtained in classical and quantum optimization. Left panel: spectra obtained in 1000 Gurobi runs with different minimization times (brown) and in 300 5-s-long runs of the D-Wave hybrid. Right panel: spectra obtained with Gurobi run, with classical Simulated Annealing (SA) and with D-Wave's hybrid solver.