Adding a Customize Drug Testin Tool

ESMFold + DeepChem Dock + ADMET-AI Pipeline (Proposal)

Purpose
This document proposes a simple end-to-end workflow where a user can provide
a protein sequence (or structure) and ligand, then obtain docked poses and
ADMET predictions.

This is a proposal only. The current DeepChem repo provides docking utilities
but does not include ESMFold integration or ADMET-AI prediction out of the box.

Pipeline Overview
1. Protein structure prediction with ESMFold
2. Docking with DeepChem (Vina or GNINA)
3. ADMET prediction with ADMET-AI

Inputs
- Protein sequence in FASTA format or an existing protein structure in PDB
- Ligand in SDF or SMILES

Outputs
- Docked poses (PDBQT and/or RDKit Mol objects)
- Docking scores (Vina or GNINA)
- ADMET predictions per ligand (model-specific outputs)

Core DeepChem Components
- Docking API: `deepchem/dock/docking.py`
- Pose generation: `deepchem/dock/pose_generation.py`
- Vina scoring terms: `deepchem/dock/pose_scoring.py`

Proposed Workflow
1. Structure prediction
- Input: protein sequence (FASTA)
- Run ESMFold externally to generate a PDB file
- Output: protein PDB

2. Docking (DeepChem)
- Input: protein PDB and ligand SDF/SMILES
- Use `VinaPoseGenerator` or `GninaPoseGenerator`
- Output: docked poses and scores

3. ADMET prediction (ADMET-AI)
- Input: ligand (SMILES) or docked ligand
- Run ADMET-AI externally
- Output: ADMET prediction table

Example Pseudo-CLI
1. Predict protein structure
- `esmfold --fasta target.fasta --out target.pdb`

2. Dock with DeepChem (Vina or GNINA)
- `python run_docking.py --protein target.pdb --ligand ligand.sdf --engine vina`

3. Predict ADMET
- `admet_ai --input ligand.smi --out admet.csv`

Minimal Python Sketch
```python
import deepchem as dc

# 1) ESMFold runs externally, producing target.pdb
protein_file = "target.pdb"
ligand_file = "ligand.sdf"

# 2) Docking with DeepChem
pose_gen = dc.dock.VinaPoseGenerator(pocket_finder=None)
docker = dc.dock.Docker(pose_gen)

poses_and_scores = list(
docker.dock(
(protein_file, ligand_file),
exhaustiveness=8,
num_modes=9,
use_pose_generator_scores=True,
)
)

# 3) ADMET-AI runs externally on ligand SMILES or docked ligands
```

Notes and Constraints
- GNINA currently runs only on Linux (see `deepchem/dock/pose_generation.py`).
- Vina is not available on Windows per the DeepChem docstring and tests.
- DeepChem does not currently provide ESMFold or ADMET-AI integrations.
- This pipeline would require external tooling and simple glue scripts.

Suggested Next Steps
1. Add a small wrapper script, for example `scripts/esmfold_dock_admet.py`,
that orchestrates ESMFold, DeepChem docking, and ADMET-AI.
2. Define a consistent input/output schema, such as a single JSON file or
folder layout for protein, ligand, docking outputs, and ADMET results.
3. Optionally store docked poses with metadata to support downstream analysis.

Start by reading deepchem/dock/docking.py, deepchem/dock/pose_generation.py, and deepchem/dock/pose_scoring.py to understand the existing docking entry points and platform constraints. The proposal mentions scripts/esmfold_dock_admet.py, but the input/output schema, external-tool integration, and acceptance criteria still need to be defined before implementation can be considered done.

Written by the indexing model from the issue text.