[Paper Review] Novel algorithms and high-performance cloud computing enable efficient fully quantum mechanical protein-ligand scoring

Narbe Mardirossian, Yuhang Wang+3 more|arXiv (Cornell University)|Apr 18, 2020

Computational Drug Discovery Methods43 references4 citations

TL;DR

This paper presents a scalable, fully quantum mechanical approach to protein-ligand scoring using density functional theory (DFT) accelerated by novel algorithms and high-performance cloud computing. By optimizing DFT calculations for cloud architectures, the method achieves accurate binding affinity predictions for 2500-atom protein-ligand complexes in tens of minutes, outperforming classical free energy perturbation in a benchmark on MCL1, enabling practical quantum-level scoring in drug discovery pipelines.

ABSTRACT

Ranking the binding of small molecules to protein receptors through physics-based computation remains challenging. Though inroads have been made using free energy methods, these fail when the underlying classical mechanical force fields are insufficient. In principle, a more accurate approach is provided by quantum mechanical density functional theory (DFT) scoring, but even with approximations, this has yet to become practical on drug discovery-relevant timescales and resources. Here, we describe how to overcome this barrier using algorithms for DFT calculations that scale on widely available cloud architectures, enabling full density functional theory, without approximations, to be applied to protein-ligand complexes with approximately 2500 atoms in tens of minutes. Applying this to a realistic example of 22 ligands binding to MCL1 reveals that density functional scoring outperforms classical free energy perturbation theory for this system. This raises the possibility of broadly applying fully quantum mechanical scoring to real-world drug discovery pipelines.

Motivation & Objective

• To overcome the computational infeasibility of applying fully quantum mechanical methods to large protein-ligand complexes in drug discovery.
• To develop scalable algorithms that efficiently utilize high-performance cloud architectures for large-scale DFT calculations.
• To enable practical, accurate binding affinity prediction using density functional theory without approximations.
• To benchmark the performance of fully quantum mechanical scoring against classical free energy perturbation methods on a realistic drug target.
• To demonstrate the feasibility of integrating fully quantum mechanical scoring into real-world drug discovery workflows.

Proposed method

• The authors implement a novel DFT algorithm optimized for distributed-memory, high-performance cloud architectures to handle large systems.
• The method uses a divide-and-conquer approach to partition the electron density calculation across multiple compute nodes, reducing computational complexity.
• The algorithm leverages efficient linear scaling techniques and parallel I/O to minimize communication overhead in cloud environments.
• The implementation is built on an existing DFT framework but enhanced with task-level parallelism and dynamic load balancing for cloud workloads.
• The system is deployed on a commercial cloud platform, utilizing thousands of CPU cores to accelerate the calculation of electronic structure.
• The method avoids approximations in the DFT formalism, ensuring high accuracy while maintaining computational efficiency.

Experimental results

Research questions

• RQ1 Can fully quantum mechanical DFT calculations be made practical for protein-ligand complexes of drug discovery relevance?
• RQ2 Can novel algorithms and cloud computing reduce the time-to-solution for large-scale DFT to under an hour for systems with ~2500 atoms?
• RQ3 Does DFT-based scoring outperform classical free energy perturbation in predicting binding affinities for a real-world target like MCL1?
• RQ4 Is it feasible to integrate fully quantum mechanical scoring into high-throughput drug discovery pipelines?
• RQ5 How does the performance of the new algorithm scale with system size and number of compute cores in a cloud environment?

Key findings

• The method achieves fully quantum mechanical DFT calculations for protein-ligand complexes of approximately 2500 atoms in tens of minutes on a cloud platform.
• The algorithm scales efficiently across thousands of CPU cores, enabling high throughput on widely available cloud infrastructure.
• On the MCL1 dataset of 22 ligands, DFT-based scoring outperformed classical free energy perturbation in predicting binding affinities.
• The approach maintains full accuracy without approximations in the DFT formalism, unlike many prior methods that sacrifice accuracy for speed.
• The results demonstrate that high-accuracy quantum mechanical scoring is now computationally feasible for realistic drug discovery applications.
• The study establishes a practical pathway for integrating fully quantum mechanical scoring into industrial and academic drug discovery pipelines.