This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Building a PC for BioML/ProteinDesign

Hey everyone,

I’m a recent bioinformatics graduate working with a limited budget, exploring the possibility of building a PC to run ML models developed by research groups like the Baker Group and DeepMind. I’m looking for recommendations on PC components or cloud-based alternatives that provide a similar experience.

What I’m Looking For:

The ideal setup would allow me to run models interactively, especially through a debugger, to get a deeper understanding of their inner workings. While I’ve been able to experiment with smaller models like ProteinMPNN on my current setup, I want the flexibility to tinker with larger models as well.

I’m already aware of ColabFold and ColabDesign—both are great tools, but they don’t fully meet my needs. I find the Google Colab interface is much less useable than a dedicated IDE like PyCharm. This can make working with these models a bit challenging. Although I’ve experimented with Google Workstation and tried connecting PyCharm via SSH, I still found the experience underwhelming.

If anyone knows of cloud solutions and accompanying software that provide a near-native local development experience, please share!

Considering Building a PC:

If building a PC is the best way forward, I’d appreciate suggestions on hardware specs. My budget is flexible but ideally under $3,000 USD. Here’s my current rough parts list or criteria:

  • GPU: High-performance with substantial memory (leaning toward an RTX 4090).
  • CPU: A recent AMD processor with plenty of threads for parallel tasks, like MSA construction.
  • RAM: At least 64GB.
  • Storage: Planning on 4TB SSD and 4TB HDD to cover all my storage needs.
  • Everything else: Still to be determined.

Thanks in advance for your help and advice!

protein-design machine-learning
• 648 views • link
updated by Matic Broz • Matic Broz • written by iloveproteins • iloveproteins

1 answer

For this category of work, I would choose hardware around the largest model you realistically need to run, not around the average Python workflow.

A practical split is:

  • ProteinMPNN / sequence design / analysis scripts: modest GPU or even CPU is often fine.
  • ColabFold / AlphaFold-style prediction: GPU memory matters more than raw CPU. A 24 GB consumer GPU can handle many single-chain and smaller multimer jobs, but large complexes can still fail.
  • RFdiffusion / generative structure design: more GPU memory is useful; 24 GB is workable for many examples, but not unlimited.
  • Boltz/Boltz-style complex prediction: plan similarly to other modern structure-prediction models: CUDA/PyTorch compatibility and VRAM are usually the first bottlenecks.

If buying one workstation, I would prioritize:

  1. NVIDIA GPU, because most of these tools are CUDA-first.
  2. At least 24 GB VRAM if you want local structure/design work to be pleasant.
  3. 64-128 GB system RAM.
  4. Fast NVMe storage; databases and model weights can be large.
  5. Linux as the primary environment, or WSL2 only if you are comfortable debugging CUDA/toolchain issues.

That said, I would not build a very expensive local machine until you know your workload. For occasional AlphaFold/ColabFold/Boltz/RFdiffusion jobs, renting cloud GPUs can be cheaper and avoids a lot of driver/database maintenance. A local workstation makes more sense if you will run jobs repeatedly, need privacy, or want interactive development.

Also separate generation from validation. Protein design outputs still need filtering: structure confidence, interface metrics, sequence diversity, developability/expression risk, and ideally experimental testing. A bigger GPU does not remove that step.

• 0 views • link
written by Matic Broz • Matic Broz

Log in to answer this question.

More posts like this