Pharmaclaw Alphafold Agent

Compliant AlphaFold Agent for protein structure retrieval, ESMFold prediction, binding site detection, and RDKit ligand docking. Fetches public PDB/AlphaFold...

Security Scan

The skill's code, instructions, and external calls are consistent with a lightweight, proof‑of‑concept AlphaFold/ESMFold + RDKit agent; it does not ask for credentials or hidden endpoints, though several capabilities are mocked and the script has bugs and is not production-grade.

Purpose & Capability

The name/description (structure retrieval, ESMFold prediction, binding site detection, RDKit docking) matches the code and dependencies (requests, biopython, RDKit). However ESMFold prediction is explicitly mocked in the script (predict_esmfold writes a placeholder PDB) rather than invoking HuggingFace/ESMFold; docking and pocket detection are simplified/mocked. This is coherent for a demo but not a full production implementation.

Instruction Scope

SKILL.md and the script instruct the agent to fetch public PDBs (RCSB), optionally read a FASTA file, run a local mock prediction, detect pockets and run an RDKit-based docking/emedding. The script writes files locally and performs HTTP requests only to legitimate public endpoints (RCSB and AlphaFold EBI). It does not read or exfiltrate unrelated system files or request credentials.

Install Mechanism

No install spec is provided (instruction-only skill with one included script). That lowers installation risk; packages listed (rdkit-pypi, biopython, requests) are consistent with functionality. No arbitrary downloads or extract/install steps are present.

Credentials

The skill declares no required environment variables, no credentials, and requests no config paths. The runtime code also does not read environment variables or secrets.

Persistence & Privilege

always is false and the skill does not request persistent presence or modify other skills or system-wide settings. It runs as a one-off script that reads/writes local files.

Assessment

This skill appears to do what it says and does not request credentials or hidden endpoints, but treat it as a proof-of-concept/demo: ESMFold prediction is mocked, binding-site detection and docking are simplistic, and the script contains runtime bugs (missing Draw import, minimal error handling). If you plan to use it in production, review and fix the code, add input validation and HTTP/response checks, replace mocks with real ESMFold/fpocket/Vina components, and run it in an isolated/sandbox environment. As with any third-party script from an unknown source, inspect the code and test on non-sensitive systems before trusting it with important data.

Overview

Protein structure retrieval and ligand docking agent for the PharmaClaw drug discovery pipeline. Fetches experimental structures from RCSB PDB and predicted structures from AlphaFold DB, detects binding sites, and performs conformer-based docking with RDKit.

Quick Start

# Fetch structure and dock a ligand
python scripts/alphafold_agent.py '{"uniprot": "P01116", "smiles": "CC(=O)Nc1ccc(O)cc1"}'

# Structure retrieval only
python scripts/alphafold_agent.py '{"uniprot": "P01116"}'

Capabilities

Feature	Method	Source
Structure Fetch	RCSB Search API + AlphaFold DB	Public PDB files
Fold Prediction	ESMFold via HuggingFace	Sequence → 3D structure
Binding Sites	Pocket detection	Residue-level pockets
Ligand Docking	RDKit conformer generation	SMILES → affinity score

Decision Tree

• UniProt ID provided? → Fetch from RCSB PDB / AlphaFold DB
• FASTA sequence provided? → Predict fold via ESMFold
• SMILES provided? → Dock ligand into detected binding pocket
• No structure found? → Fall back to ESMFold prediction

Input Format

{
  "uniprot": "P01116",
  "smiles": "CC(=O)Nc1ccc(O)cc1",
  "fasta": "path/to/sequence.fasta"
}

Output Format

{
  "pdb": "1abc.pdb",
  "sites": [{"res": "G12", "pocket_vol": 150}],
  "docking": {"affinity": -15.2, "viz": "docked.png"},
  "compliance": "Public AlphaFold 2 DB/ESMFold (commercial OK)"
}

Chain Integration

• Receives from: Chemistry Query (SMILES for docking), Literature (target proteins)
• Feeds into: IP Expansion (novel binding modes), Catalyst Design (structure-guided synthesis)

Dependencies

• rdkit-pypi — Conformer generation and molecular descriptors
• biopython — PDB parsing and FASTA sequence handling
• requests — API calls to RCSB and AlphaFold DB

Compliance

Uses only publicly available protein structures (RCSB PDB, AlphaFold DB) and open-source prediction (ESMFold). All data sources are commercially permissible. No proprietary AlphaFold 3 server calls.

Scripts

• scripts/alphafold_agent.py — Main agent: fetch, predict, detect sites, dock

Limitations

• Docking uses RDKit conformer scoring (not full physics-based docking like Vina)
• ESMFold prediction requires significant compute for large proteins
• Binding site detection is simplified; production use should integrate fpocket or P2Rank