Predict the unpredictable: Disrupting drug lead optimization using quantum mechanics simulation in the cloud

This post was contributed by Klaas Gunst (VP, Scientific Software, QSimulate) and Toru Shiozaki (CEO, QSimulate)

Physics-based computer simulation plays an important role in drug discovery, allowing scientists to accurately model molecular-level interactions. However, traditional simulations do not apply to unconventional binding modalities, limiting their impact. This is because most of the commercial offerings are based on classical-mechanics-based models. These models ignore certain physics, for example, drugs binding to metal sites, those that form covalent bonds, and binders inducing strong protein polarization.

Quantum mechanics (QM) can accurately describe such interactions. However, the direct application of QM to free energy perturbations (FEP) for protein-drug complexes was previously not feasible because it is far more computationally intensive than classical mechanics simulations. QSimulate has made a series of scientific, algorithmic, and software breakthroughs to address this challenge, as partly reported in an NVIDIA technical blog post.

In this blog post, we will review our recent development of a mixed-precision (FP64/FP32) simulation solution for QM-based FEP simulation for lead optimization of small-molecule drugs. This is an extension to QSimulate’s QM FEP product, QUELO, and has led to an order-of-magnitude reduction in computing cost. The QM-based lead optimization solution is now available on AWS using cost-effective instance series such as the G6e instances.

Quantum Mechanics Pushes the Boundary of Physics-Based Simulation

QSimulate’s QUELO product accelerates QM simulation of protein-drug complexes. Each simulation now takes only a few milliseconds, realizing, for the first time, a throughput of 100-nanosecond dynamics per day on a single GPU card. This is in stark contrast to conventional QM simulation solutions that take seconds, or even minutes. In addition to the software implementation, QUELO has also optimized the hardware and orchestration to enable this performance. This has been greatly facilitated by AWS services, such as Amazon Elastic Compute Cloud (Amazon EC2), Amazon Relational Database Service (Amazon RDS), AWS Lambda, Amazon Simple Storage Service (Amazon S3), and AWS ParallelCluster. See Figure 1 for system architecture diagram.

QUELO leverages AWS Lambda to quickly process user inputs that require quick turnarounds. The user inputs, intermediate information, and simulation results are stored in a database in Amazon RDS, and a graphical user interface (GUI) visualizes the task. Each simulation that users submit to QUELO typically spawns tens or even hundreds of tasks to be computed on EC2 instances. Analyses can be stopped or restarted from GUI, and the workflow state is tracked in the RDS database. Large output files such as molecular dynamics trajectories are stored in Amazon S3 and are made downloadable from GUI for expert users.

Figure 1 – The architecture diagram of the QUELO solution in AWS. QUELO utilizes various AWS solutions including AWS EC2, RDS, Lambda, S3, and Parallel Cluster.

Mixed Precision QM Engine: How it works

The mixed precision algorithm for QM FEP, released at the end of 2024, largely follows the standard strategy for classical mechanics simulation. Most of the energy components are computed in FP32 but accumulated in FP64. We took special care with the numerical precision of the quantities that enter into the iterative QM solution to guarantee that the convergence patterns are not negatively impacted by the use of FP64/FP32 mixed-precision arithmetic.

From the users’ perspective, QM-FEP simulation with FP64/FP32 mixed precision brings significant advantages. QUELO can now perform effectively on GPU cards that do not have hardware support for FP64, improving flexibility in hardware choice, like the Amazon EC2 G Instance family.

The GPUs on G6e instances feature 91.61 TFLOPS for FP32 while providing only 1.4 TFLOPS for FP64 (the ratio of 64:1). This FP32 performance is remarkable because even top-end GPUs on P5 instances do not have the same kind of FLOPS counts for FP64.

The use of FP64/FP32 mixed precision significantly improves the price/performance ratio per simulation on AWS. This is because instances in the G Series are generally more cost-effective compared to those in the P Series for the kind of scientific computing workloads like ours, as shown in Figure 2.

Figure 2 – The typical reduction in computing cost and turnaround time using the mixed precision engine and G6e instance series. We measured the cost and timing using a system with 25,000 atoms in a unit cell with a QM region of about 80 atoms.

By taking advantage of FP64/FP32 mixed precision and the cost-effective G-series instances (e.g., g6e.2xlarge instances), we have observed that the time to solution has been decreased by more than a factor of 2, while the computing cost has been reduced by a factor of 7-8.

Conclusion

With QUELO, QM-based lead optimization simulations are now not only feasible but in fact performant and cost effective.

You can take advantage of these new implementation in QUELO. According to Dr. Michael Bartberger, Senior Vice President of Computational Chemistry at Alterome Therapeutics, who is a pioneer in the field of computer-aided drug discovery, lead optimization with production-level QM FEP in the startup setting was challenging in the past because of budgetary considerations. But now, “a mixed-precision QM-FEP engine is a game changer. We at Alterome have been pleased with the performance of the QUELO suite in our drug discovery campaigns. The ability to incorporate quantum mechanics into dynamics-based relative free energy methods bodes well for increasing the predictive accuracy of RBFE calculations against challenging targets. Further, the ability to do this utilizing commodity GPU hardware now allows its use in a routine way.”

Visit the QUELO product page to learn more. QSimulate offers QUELO Experience through which you can witness the power of quantum for drug lead optimization at no cost.

The content and opinions in this blog are those of the third-party author and AWS is not responsible for the content or accuracy of this blog.