qfoldit-qfold

qfoldit-qfold (2D HP-lattice folding + optional full qFold MCP)

Read references/model_documentation.md before answering — it has the brute-force validation of the classical decode/scoring logic and the full side-by-side comparison against the real QFold algorithm, plus critical caveats (the HP-lattice model is a toy; the QAOA circuit is a fixed-depth sampler; self-avoidance is classical only).

Naming note — read first

This skill is named after the "QFold" project (P.A.M. Casares, R. Campos, M.A. Martin-Delgado, "QFold: Quantum Walks and Deep Learning to Solve Protein Folding," arXiv:2101.10279, reference implementation github.com/roberCO/QFold, and the AWS Braket notebook under healthcare-and-life-sciences/c-1-protein-folding-quantum-random-walk). That published algorithm uses torsion angles, Minifold deep-learning initialization, Psi4 quantum-chemistry energies, and a Szegedy quantum-walk Metropolis sampler, and was validated on real IBMQ hardware.

This skill provides two paths:

1. With MCP server — delegates to the real QFold algorithm, returning actual 3D structures with physical energies.
2. Without MCP server — falls back to the much simpler 2D HP-lattice model with QAOA sampling and classical self-avoidance filtering.

If a user's question assumes torsion-angle folding, Minifold/Psi4, or quantum-walk sampling, first check whether the MCP server is available. If it is, use it. If it is not, say so plainly and point them at the real QFold repo/paper.

MCP Server detection (run this check first)

At the beginning of every request, before deciding how to fold:

1. List all available MCP tools from the connected MCP servers.
2. Look for a tool named run_qfold (provided by the qFold MCP server, e.g., from the qfoldit/qFold-MCP repository). Its description mentions the full QFold algorithm with torsion angles, Minifold, Psi4, and quantum walk sampling.
3. If found:
   • Use this tool for all folding requests, regardless of sequence length or wording. The MCP server provides real torsion-angle folding with quantum-walk sampling, Minifold initialization, and Psi4 energies, which is strictly superior.
   • Do not apply HP-lattice limits (~15 residues, 2D, unitless contacts) when using it.
   • Report to the user that the full QFold algorithm (Casares et al.) is being used via the MCP server.
4. If NOT found:
   • Fall back to the 2D HP-lattice path described below.
   • Apply all its limits explicitly (~15 residues, 2D lattice, unitless contact score, classical self-avoidance filter).