modern-structure-prediction_skill

Overview

This skill predicts protein 3D structures using modern machine-learning models including AlphaFold3, ESMFold, Chai-1, and Boltz-1. It supports single-sequence fast predictions, MSA-based high-accuracy runs, and complex or ligand-aware inference. The skill also provides comparison tools to align and quantify differences between outputs.

How this skill works

The skill can call cloud APIs (ESM Atlas, AlphaFold3 server) or run local open-source models (ESMFold via esm, Chai-1, Boltz-1, ColabFold). It returns PDB/CIF coordinates and per-residue confidence (pLDDT in B-factor). Utilities parse outputs, extract confidence metrics (pLDDT, pTM, ipTM, PAE), and compute pairwise comparisons such as RMSD. Examples show typical command-line and Python usage and recommend memory and runtime settings.

When to use it

• Quick single-chain predictions or high-throughput screening (ESMFold API/local)
• Highest-accuracy single-chain or complex predictions where MSAs/templates improve results (AlphaFold3/ColabFold)
• Protein–protein or heteromeric complex modeling (AlphaFold3, Chai-1, Boltz-1)
• Protein–ligand modeling with ligand-aware methods (Chai-1 or AlphaFold3)
• Comparative analysis across methods to assess model agreement and uncertainty

Best practices

• Choose model by task: ESMFold for speed, AlphaFold3/ColabFold for accuracy, Chai-1/Boltz-1 for complexes and ligands
• Verify installed package versions and adapt API calls to local library signatures before running
• Monitor GPU memory requirements (ESMFold ~16GB, Chai-1/Boltz-1 ~24GB) and adjust batch sizes or use cloud APIs if needed
• Extract and inspect per-residue pLDDT and global metrics (pTM/ipTM/PAE) rather than relying solely on visual inspection
• When comparing models, align equivalent residues and use CA-based RMSD or other robust metrics to avoid artifacts

Example use cases

• Run a rapid single-sequence structure prediction for a novel enzyme with ESMFold API and inspect pLDDT to locate confident regions
• Predict a heteromeric complex for two interacting proteins using Boltz-1 or AlphaFold3 and report ipTM and interface contacts
• Benchmark the same sequence across ESMFold, AlphaFold3, and Chai-1, compute pairwise RMSD, and produce a consensus confidence map
• Include a small-molecule ligand SMILES in Chai-1 input to model protein–ligand binding and extract ligand coordinates for downstream docking
• Integrate ColabFold into a pipeline for medium-throughput screening using MMseqs2 to generate MSAs and templates