inference_app.py

from __future__ import annotations
from pathlib import Path
import time
from biotite.application.autodock import VinaApp

import gradio as gr

from gradio_molecule3d import Molecule3D
from gradio_molecule2d import molecule2d
import numpy as np
from rdkit import Chem
from rdkit.Chem import AllChem
import pandas as pd
from biotite.structure import centroid, from_template
from biotite.structure.io import load_structure
from biotite.structure.io.mol import MOLFile, SDFile
from biotite.structure.io.pdb import PDBFile

from plinder.eval.docking.write_scores import evaluate


EVAL_METRICS = ["system", "LDDT-PLI", "LDDT-LP", "BISY-RMSD"]


def vina(
    ligand, receptor, pocket_center, output_folder: Path, size=10, max_num_poses=5
):
    app = VinaApp(
        ligand,
        receptor,
        center=pocket_center,
        size=[size, size, size],
    )
    app.set_max_number_of_models(max_num_poses)
    app.start()
    app.join()
    docked_ligand = from_template(ligand, app.get_ligand_coord())
    docked_ligand = docked_ligand[..., ~np.isnan(docked_ligand.coord[0]).any(axis=-1)]
    output_files = []
    for i in range(max_num_poses):
        sdf_file = MOLFile()
        sdf_file.set_structure(docked_ligand[i])
        output_file = output_folder / f"docked_ligand_{i}.sdf"
        sdf_file.write(output_file)
        output_files.append(output_file)
    return output_files


def predict(
    input_sequence: str,
    input_ligand: str,
    input_msa: gr.File | None = None,
    input_protein: gr.File | None = None,
    max_num_poses: int = 1,
):
    """
    Main prediction function that calls ligsite and smina

    Parameters
    ----------
    input_sequence: str
        monomer sequence
    input_ligand: str
        ligand as SMILES string
    input_msa: gradio.File | None
        Gradio file object to MSA a3m file
    input_protein: gradio.File | None
        Gradio file object to monomer protein structure as CIF file
    max_num_poses: int
        Number of poses to generate

    Returns
    -------
    output_structures: tuple
        (output_protein, output_ligand_sdf)
    run_time: float
        run time of the program
    """
    start_time = time.time()

    if input_protein is None:
        raise gr.Error("need input_protein")
    print(input_protein)
    ligand_file = Path(input_protein).parent / "ligand.sdf"
    print(ligand_file)
    conformer = Chem.AddHs(Chem.MolFromSmiles(input_ligand))
    AllChem.EmbedMolecule(conformer)
    AllChem.MMFFOptimizeMolecule(conformer)
    Chem.SDWriter(ligand_file).write(conformer)
    ligand = SDFile.read(ligand_file).record.get_structure()
    receptor = load_structure(input_protein, include_bonds=True)
    docking_poses = vina(
        ligand,
        receptor,
        centroid(receptor),
        Path(input_protein).parent,
        max_num_poses=max_num_poses,
    )
    end_time = time.time()
    run_time = end_time - start_time
    pdb_file = PDBFile()
    pdb_file.set_structure(receptor)
    output_pdb = Path(input_protein).parent / "receptor.pdb"
    pdb_file.write(output_pdb)
    return [str(output_pdb), str(docking_poses[0])], run_time


def get_metrics(
    system_id: str,
    receptor_file: Path,
    ligand_file: Path,
    flexible: bool = True,
    posebusters: bool = True,
) -> tuple[pd.DataFrame, float]:
    start_time = time.time()
    metrics = pd.DataFrame(
        [
            evaluate(
                model_system_id=system_id,
                reference_system_id=system_id,
                receptor_file=receptor_file,
                ligand_file_list=[Path(ligand_file)],
                flexible=flexible,
                posebusters=posebusters,
                posebusters_full=False,
            ).get("LIG_0", {})
        ]
    )
    if posebusters:
        metrics["posebusters"] = metrics[
            [col for col in metrics.columns if col.startswith("posebusters_")]
        ].sum(axis=1)
        metrics["posebusters_valid"] = metrics[
            [col for col in metrics.columns if col.startswith("posebusters_")]
        ].sum(axis=1) == 20
    columns = ["reference", "lddt_pli_ave", "lddt_lp_ave", "bisy_rmsd_ave"]
    if flexible:
        columns.extend(["lddt", "bb_lddt"])
    if posebusters:
        columns.extend([col for col in metrics.columns if col.startswith("posebusters")])

    metrics = metrics[columns].copy()
    mapping = {
            "lddt_pli_ave": "LDDT-PLI",
            "lddt_lp_ave": "LDDT-LP",
            "bisy_rmsd_ave": "BISY-RMSD",
            "reference": "system",
        }
    if flexible:
        mapping["lddt"] = "LDDT"
        mapping["bb_lddt"] = "Backbone LDDT"
    if posebusters:
        mapping["posebusters"] = "PoseBusters #checks"
        mapping["posebusters_valid"] = "PoseBusters valid"
    metrics.rename(
        columns=mapping,
        inplace=True,
    )
    end_time = time.time()
    run_time = end_time - start_time
    return metrics, run_time


with gr.Blocks() as app:
    with gr.Tab("🧬 Vina"):
        gr.Markdown(
            "Example model using Vina to dock the ligand with the pocket center defined by the centroid of the input protein."
        )
        with gr.Row():
            input_sequence = gr.Textbox(lines=3, label="Input Protein sequence (FASTA)")
            input_ligand = gr.Textbox(lines=3, label="Input ligand SMILES")
            input_msa = gr.File(label="Input MSA (a3m)")
            input_protein = gr.File(label="Input protein monomer (CIF)")
        max_num_poses = gr.Slider(
            1, 10, value=1, label="Max number of poses to generate"
        )
        inference_btn = gr.Button("Run Inference")
        gr.Examples(
            [
                [
                    "QECTKFKVSSCRECIESGPGCTWCQKLNFTGPGDPDSIRCDTRPQLLMRGCAADDIMDPTSLAETQEDHNGGQKQLSPQKVTLYLRPGQAAAFNVTFRRAKGYPIDLYYLMDLSYSMLDDLRNVKKLGGDLLRALNEITESGRIGFGSFVDKTVLPFVNTHPDKLRNPCPNKEKECQPPFAFRHVLKLTDNSNQFQTEVGKQLISGNLDAPEGGLDAMMQVAACPEEIGWRKVTRLLVFATDDGFHFAGDGKLGAILTPNDGRCHLEDNLYKRSNEFDYPSVGQLAHKLAENNIQPIFAVTSRMVKTYEKLTEIIPKSAVGELSEDSSNVVQLIKNAYNKLSSRVFLDHNALPDTLKVTYDSFCSNGVTHRNQPRGDCDGVQINVPITFQVKVTATECIQEQSFVIRALGFTDIVTVQVLPQCECRCRDQSRDRSLCHGKGFLECGICRCDTGYIGKNCECQTQGRSSQELEGSCRKDNNSIICSGLGDCVCGQCLCHTSDVPGKLIYGQYCECDTINCERYNGQVCGGPGRGLCFCGKCRCHPGFEGSACQCERTTEGCLNPRRVECSGRGRCRCNVCECHSGYQLPLCQECPGCPSPCGKYISCAECLKFEKGPFGKNCSAACPGLQLSNNPVKGRTCKERDSEGCWVAYTLEQQDGMDRYLIYVDESRECCGGPAALQTLFQG",
                    "CC(=O)N[C@H]1[C@H](O[C@H]2[C@H](O)[C@@H](NC(C)=O)CO[C@@H]2CO)O[C@H](CO)[C@@H](O)[C@@H]1O",
                    "input_protein_test.cif",
                    "input_protein_test.cif",
                ],
            ],
            [input_sequence, input_ligand, input_msa, input_protein],
        )
        reps = [
            {
                "model": 0,
                "style": "cartoon",
                "color": "whiteCarbon",
            },
            {
                "model": 0,
                "resname": "UNK",
                "style": "stick",
                "color": "greenCarbon",
            },
            {
                "model": 0,
                "resname": "LIG",
                "style": "stick",
                "color": "greenCarbon",
            },
            {
                "model": 1,
                "style": "stick",
                "color": "greenCarbon",
            },
        ]
        # smiles = molecule2d(input_ligand)
        out = Molecule3D(reps=reps)
        run_time = gr.Textbox(label="Runtime")

        inference_btn.click(
            predict,
            inputs=[
                input_sequence,
                input_ligand,
                input_msa,
                input_protein,
                max_num_poses,
            ],
            outputs=[out, run_time],
        )
    with gr.Tab("⚖️ PLINDER evaluation template"):
        with gr.Row():
            with gr.Column():
                input_system_id = gr.Textbox(label="PLINDER system ID")
                input_receptor_file = gr.File(label="Receptor file (CIF)")
                input_ligand_file = gr.File(label="Ligand file (SDF)")
                flexible = gr.Checkbox(label="Flexible docking", value=True)
                posebusters = gr.Checkbox(label="PoseBusters", value=True)

        eval_btn = gr.Button("Run Evaluation")
        gr.Examples(
            [
                [
                    "4neh__1__1.B__1.H",
                    "input_protein_test.cif",
                    "input_ligand_test.sdf",
                    True,
                    True,
                ],
            ],
            [input_system_id, input_receptor_file, input_ligand_file, flexible, posebusters],
        )
        reps = [
            {
                "model": 0,
                "style": "cartoon",
                "color": "whiteCarbon",
            },
            {
                "model": 0,
                "resname": "UNK",
                "style": "stick",
                "color": "greenCarbon",
            },
            {
                "model": 0,
                "resname": "LIG",
                "style": "stick",
                "color": "greenCarbon",
            },
            {
                "model": 1,
                "style": "stick",
                "color": "greenCarbon",
            },
        ]

        # pred_native = Molecule3D(reps=reps, config={"backgroundColor": "black"})
        eval_run_time = gr.Textbox(label="Evaluation runtime")
        metric_table = gr.DataFrame(
            pd.DataFrame([], columns=EVAL_METRICS), label="Evaluation metrics"
        )

        eval_btn.click(
            get_metrics,
            inputs=[input_system_id, input_receptor_file, input_ligand_file, flexible, posebusters],
            outputs=[metric_table, eval_run_time],
        )

app.launch()