syncraft-design

name

syncraft-design

description

Edit unsynthesizable, computationally-generated drug candidate molecules into synthesizable analogs that retain binding to a target protein pocket. Invoke when the user has SMILES from a 3D-aware generative model (Pocket2Mol, ResGen, TargetDiff, DiffSBDD, LIGAN, AR, or any de novo source) and asks for "synthesizable rescue", "make X synthesizable", "fix synthesis cliff", "edit to synthesizable", "lower SA score", or similar. The skill docks the inputs at the target pocket with AutoDock Vina, calls a reasoning LLM (Gemini-2.5-Pro, Gemini-3.1-Pro, or DeepSeek-V4-Pro) to predict atom-level edit sequences that preserve the PLIP interaction profile while removing synthesis liabilities, re-docks the edited products, and returns a ranked list of (original, edited, scores, similarity, reasoning) tuples. Use ONLY when the inputs are already-generated molecules and a defined receptor pocket exist. Do NOT use for: de novo generation, full retrosynthesis route planning, free-energy calculations, or covalent docking.

SynCraft Skill — molecular synthesizability editing

When to invoke

Trigger this skill when the user input matches all three:

1. A list of SMILES from a 3D-aware generative model, or a path to a .smi / .txt file containing one SMILES per line.
2. A target protein: a 4-letter PDB ID, a .pdb file path, or a known pocket descriptor.
3. An intent expressed as "make synthesizable", "edit to easier synthesis", "fix synth cliff", "lower SA", "drug-like rescue", "drug-like edit", or equivalent.

Counter-indications (do NOT invoke):

• User asks to generate new molecules from a pocket → use a generative-model skill instead.
• User asks for full retrosynthesis route → use a dedicated retrosynthesis-planning tool instead.
• User asks for binding-mode analysis only (no editing) → use a PLIP/docking-only skill.

What it does

INPUT                                OUTPUT
─────────────────────────            ─────────────────────────────────
list[SMILES] (generated)       ─→    list[EditResult]
PDB id or .pdb file           ─→     ├── original_smiles
[pocket residues OR box]       ─→    ├── original_dock_score (Vina kcal/mol)
[choice of rescue_llm]         ─→    ├── edited_smiles
                                     ├── edited_dock_score
                                     ├── similarity_to_original (Tanimoto)
                                     ├── edit_operations (list[op, atom_idx])
                                     ├── edit_reasoning (LLM explanation)
                                     └── success (bool)

Single-call entry point

from syncraft import edit_for_synthesizability

results = edit_for_synthesizability(
    molecules=[                                   # list of SMILES
        "NC(=O)[C@@H]1CCC(C(=O)N[C@H]2OC(=O)[C@@H](O)[C@@H](O)[C@H]2NNC2C=CC=CC=C2Cl)C1",
        # ...
    ],
    protein_pdb="7L12",                           # PDB ID or local .pdb
    pocket_residues=None,                         # optional explicit residue list
    rescue_llm="deepseek-v4-pro",                 # gemini-2.5-pro | gemini-3.1-pro | deepseek-v4-pro
    pass_k=5,                                     # rescue attempts per molecule
    output_dir="./edited_hits",
    seed=42,
)

Returns list[EditResult], one per input molecule. Results also written as JSON to output_dir/results.json plus per-pose PDBs in output_dir/poses/.

Pipeline (4 stages, no retrosynthesis)

Stage	Tool	Time per molecule
1. Dock original molecule at pocket	AutoDock Vina	~5-15 s
2. Extract PLIP interactions from docked pose	PLIP	~1 s
3. LLM rescue: edit sequence preserving PLIP profile	reasoning LLM	~5-15 s per pass
4. Apply edits, validate, re-dock edited products	RDKit + Vina	~10-20 s

For 100 candidates × pass_k=5 with Vina-cpu=2 × 8 workers: ~30-60 min wall.

Supported LLMs

Model	License	Local-able	Note
gemini-2.5-pro	API only	no	paper headline; thinking_mode=high
gemini-3.1-pro	API only	no	newer, better at long reasoning chains
deepseek-v4-pro	MIT (open weights)	yes	reproducible long-term; HF: deepseek-ai/DeepSeek-V4-Pro

Default: deepseek-v4-pro (long-term reproducibility per NMI editor's reproducibility requirement).

Required environment

• Conda env with: rdkit ≥ 2022, vina 1.2.x, meeko 0.5+, plip ≥ 2.3, biopython, numpy, pydantic
• API keys (one of): GEMINI_API_KEY (for gemini-*), local vLLM/SGLang endpoint (for deepseek-v4-pro)
• See setup.sh for one-line bootstrap

Required resources at runtime

• A receptor .pdbqt (auto-prepared from .pdb if not provided)
• A Vina box config: 6 floats (center_x/y/z + size_x/y/z); auto-derived from pocket_residues if given
• An "analysis PDB" (receptor with HETATM trimmed) for PLIP

Outputs

output_dir/
├── results.json          # list[EditResult] serialized
├── results.csv           # tabular view (one row per molecule)
├── summary.md            # human-readable summary with top-10 ranked hits
├── poses/
│   ├── orig_NN__rep0.pdb
│   └── edit_NN_pXX__rep0.pdb
└── plip/
    └── *.json            # per-pose interaction profile

Examples

See examples/:

• examples/01_mpro_targetdiff.ipynb — 47 TargetDiff candidates → SARS-CoV-2 Mpro (7L12) → 5-step demo
• examples/02_kinase_p2m.ipynb — Pocket2Mol candidates → RIPK1 (7YDX) kinase rescue
• examples/quickstart.py — minimal 20-line invocation

Citation

If you use this skill in published work, cite:

@article{li2026syncraft,
  title={SynCraft: Guiding Large Language Models to Predict Edit Sequences for Molecular Synthesizability Optimization},
  author={Li, Junren and ... and Lai, Luhua},
  journal={Nature Machine Intelligence},
  year={2026},
  note={under revision}
}

Skill version

v0.1.0 (2026-05-17). Initial release coupled to NMI paper revision.