Foldism

Multi-algorithm protein structure prediction using Modal serverless infrastructure.

Runs Boltz-2, Chai-1, Protenix, AlphaFold 2, OpenFold 3, and ESMFold 2.

Quick Start

Set up uv and modal (mac and linux)

curl -LsSf https://astral.sh/uv/install.sh | sh
uvx modal setup

Run dev server or deploy to Modal

uvx modal serve foldism.py
uvx modal deploy foldism.py

Command Line Interface

Run basic tests on command line

echo -e ">Insulin\nGIVEQCCTSICSLYQLENYCN\n>InsulinB\nFVNQHLCGSHLVEALYLVCGERGFFYTPKT" > input.faa
uvx modal run foldism.py --input-faa input.faa
uvx modal run foldism.py --input-faa input.faa --algorithms chai1,boltz2
uvx modal run foldism.py --input-faa input.faa --no-use-msa  # faster but lower quality

MSA pipeline

• All backends share a single ColabSearch/MMseqs2 MSA fetch (Sokrypton's API), cached on a Modal Volume by sequence-set hash.
• Protenix's built-in MSA server (protenix-server.com) is explicitly disabled — protenix_predict refuses to run with use_msa=True and no msa_result.
• ESMFold 2 can run with MSA (PLM + MSA, more accurate) or without (--no-use-msa, PLM-only, faster). Boltz-2 always uses MSA regardless.

Cost

• Modal currently gives you $30 per month free to experiment with
• A run with all 6 methods can cost $1+ depending on sequence length (GPUs cost $1-4 per hour)
• Results are cached, so re-running the same sequence is almost free
• If you modal deploy, the URL is publicly available (anyone who finds it can use it)
• If you want to require users to authenticate, set up a custom domain. I use cloudflare but there are many ways to do this.

License

MIT