
Biotech and pharmaceutical research increasingly relies on high-performance computing (HPC) to power structure prediction, molecular modeling, and simulation workflows. However, the cost, scalability, and complexity of deploying large-batch workloads of models like Boltz-1 remain key obstacles.
Fovus is an AI-powered, serverless HPC platform that makes it easy to run Boltz-1 simulations with high performance, low cost, and minimal effort. This case study highlights how Fovus delivered biomolecular structure predictions for as little as 10 cents per simulation — with no cloud expertise required.
About Boltz-1
Boltz-1 is an open-source model developed by MIT’s Jameel Clinic for predicting 3D biomolecular structures. It supports proteins, RNA, DNA, modified residues, glycans, and covalent ligands. Boltz-1 allows users to define specific interaction pockets or molecular contacts to guide prediction, enabling a wide range of biological use cases.
Performance benchmarks demonstrate its accuracy:
- 65% LDDT-PLI for protein-ligand predictions
- 83% DockQ success rate for protein-protein complexes¹ ²
These capabilities make Boltz-1 ideal for:
- Drug discovery and virtual screening
- De novo protein design and synthetic biology
- Structural biology research in academic and commercial settings
Boltz-1 is released under the MIT license, allowing unrestricted academic and commercial use³.
Challenges in Running Boltz-1 at Scale
Despite its strong capabilities, deploying Boltz-1 for large-batch simulations introduces significant hurdles:
- Limited scalability when running hundreds or thousands of simulations
- High GPU costs in cloud environments
- Lack of tools to evaluate hardware options or select optimal configurations
- Setting up and managing cloud environments for HPC is distracting and time-consuming
- Costs that often exceed expectations due to inefficient resource use
Without both intelligence and automation, teams are left to manually define HPC strategies and cloud configurations. This often results in low throughput, slower time to insight, and unnecessary overspending.
Fovus Intelligent HPC Platform
Fovus removes the complexity of running high-performance computing workloads in the cloud by combining AI-driven strategy selection with full-stack automation. The platform fully manages end-to-end cloud HPC workflows, including strategy optimization, infrastructure provisioning, workload orchestration, data handling, and software environments, through a unified intelligent system that supports the entire HPC lifecycle. Its serverless architecture eliminates the need for cloud setup and management, allowing users to run and scale workloads efficiently without cloud infrastructure overhead.
Key Capabilities of Fovus:
- Automated Benchmarking: Quantifies how different GPU types and parallelization settings impact runtime and cost, helping choose the most efficient setup before launching large workloads.
- AI-Driven HPC Strategy Optimization: Selects the most efficient GPU instance type and system configuration for your workload, based on real-world benchmarking data and infrastructure dynamics — no manual tuning required.
- Dynamic Multi-Cloud-Region Auto-Scaling: Distributes jobs across multiple cloud regions and zones to improve resource availability, reduce wait times, and enable large-scale simulation batches to run in parallel.
- Intelligent Spot Instance Utilization: Minimizes simulation costs by allocating optimal spot instances based on benchmarking data, and current pricing and availability. Automatically fails over to alternate spot instances upon interruption to ensure the integrity of results.
- Continuous Improvement: Auto-upgrade your HPC strategy by adapting benchmarking to new infrastructure options. Stay ahead as cloud environments evolve, with no need for manual updates or reconfiguration.
- Serverless HPC Model: Eliminates infrastructure setup and cloud management headaches. Users can launch simulations with a single command or a few clicks and only pay for the task compute time.
Benchmark Setup
Fovus benchmarked Boltz-1 on three publicly available biomolecular systems of increasing complexity:
System | Link | Description | Atom Count | Number of Residues |
---|---|---|---|---|
System 1 | 7U8C | Crystal structure of Mesothelin C-terminal peptide-MORAb 15B6 FAB complex | 3633 | 444 |
System 2 | 7BZB | Crystal structure of plant sesterterpene synthase AtTPS18 | 4676 | 554 |
System 3 | 7URD | Human PORCN in complex with LGK974 and WNT3A peptide | 5534 | 970 |
Each simulation was run using the Fovus-hosted Boltz-1 Docker container. Fovus hosts the complete run-time environment for Boltz-1, including preloaded model weights and pre-computed Multiple Sequence Alignments (MSAs), eliminating the request rate limit issue of public MSA servers and ensuring the simplicity and scalability to run Boltz-1 out-of-box at scale. To run the workload, we used the following command:
docker run \
--volume=$PWD:/container_workspace \
--workdir=/container_workspace \
--rm \
--ipc=host \
--gpus all \
fovus/boltz:latest \
boltz predict ./input/input.fasta --use_msa_server --out_dir ./output/ --cache ./cache --sampling_steps 600 --recycling_steps 6
Boltz-1 downloads the required data and model parameters online by default, introducing runtime overhead and variance depending on network speed. Fovus hosts the required databases while also building, testing, and maintaining the Docker images for Boltz-1, enabling seamless deployment in a containerized environment. By managing all software and cloud environments, Fovus allows anyone to launch, scale, and optimize Boltz-1, which runs effortlessly out of the box and requires little to no setup. Scientists can focus on their research without the burden of dealing with software or cloud management hassles.
For each biomolecular system, the study was conducted with a specific objective for the HPC strategy optimization. The chosen objective minimized both the cost and runtime of the simulation. This strategy optimizes cost-efficiency and performance to ensure the best outcome for running simulations on the cloud.
Key performance metrics analyzed included:
- Simulation runtime
- Simulation cost
Results
Below are the performance and cost-efficiency results achieved on Fovus:
Input System | Runtime (minutes) | Cost |
---|---|---|
System 1 | 16 | $0.10 |
System 2 | 18 | $0.12 |
System 3 | 21 | $0.29 |
Even the most complex structure completed in just 21 minutes for under 30 cents. Compared to conventional cloud workflows, this represents a 5 to 10 times cost reduction. Performance scaled proportionally with system size, demonstrating Fovus’s ability to optimize runtime and cost automatically across varying workloads.
Additional Benchmarking Highlights
Fovus delivers strong cost efficiency across a wide range of life sciences workloads:
- Vina GPU Docking: Starting from $0.95 per 10,000 receptor-ligand dockings
- OpenMM Molecular Dynamics: Starting from $6.59 /μs ($ per microsecond) with simulation speeds up to 3,359 ns/day (nanoseconds per day)
- MegaDock protein-protein Docking: Starting from $3.09 per 1,000 antibody-antigen dockings
- GROMACS Molecular Dynamics: Starting from $29 /μs (microsecond) with up to 1,139 ns/day (nanoseconds per day)
- High-Throughput Virtual Screening: Screen 500 million ligands in just a week at only $3,750, using an iterative AI-accelerated molecular docking approach that combines traditional molecular docking, AI model training and refinement, and AI inference to augment screening iteratively
Customer Success Snapshots
Fovus is trusted by forward-thinking organizations and industry leaders tackling large-scale simulation challenges to drive scientific discoveries and engineering breakthroughs:
- A Series B next-generation biotech company accelerated its design-make-test-analyze (DMTA) cycles in AI drug discovery by 96 times while reducing cloud costs to just 20 percent of prior cloud spend
- Komatsu, the world’s second-largest heavy equipment manufacturer, adopted Fovus to migrate all of its computer-aided engineering (CAE) workflows to the cloud, accelerating digital innovation and bringing products to market faster
- ChemSpace, a leading compound marketplace and a contract research organization (CRO), reduced the time-to-insight for large-scale compound screening from weeks to hours, while cutting down 85% of the cloud expenses, operating more efficiently, and delivering drug discovery projects faster for global clients
Conclusion
Fovus delivers enterprise-grade HPC performance through intelligent automation. By combining AI-driven optimization with dynamic cloud orchestration, the platform eliminates infrastructure and cloud management burdens while reducing Boltz-1 simulation costs to just cents per prediction. Boltz-1 runs fast and cost-effectively on Fovus, with no manual tuning required.
For teams in biotech, pharma, and engineering, Fovus makes high-performance computing simple, scalable, accessible, and cost-effective. It helps scientists and engineers accelerate large-scale simulations, expedite discovery cycles, and optimize compute usage by addressing the complexity, urgency, and budget pressures shaping modern R&D.
Explore how AI-optimized HPC can streamline your Boltz-1 workflows.
References
- Wohlwend, J., Corso, G., Passaro, S., Getz, N., Reveiz, M., Leidal, K., Swiderski, W., Atkinson, L., Portnoi, T., Chinn, I., Silterra, J., Jaakkola, T., & Barzilay, R. (2024). Boltz-1: Democratizing Biomolecular Interaction Modeling. bioRxiv, 2024.11.19.624167
- Boltz-1 GitHub Repository: https://github.com/jwohlwend/boltz
- MIT License Overview: https://opensource.org/licenses/MIT