app.py

import os
import sys
import streamlit as st
import pandas as pd
import joblib
import time
import numpy as np
from rdkit import Chem
from rdkit.Chem import Draw
from rdkit.Chem import AllChem
from rdkit import RDLogger
import uuid
from datasets import load_dataset
import requests
from io import BytesIO
import urllib.request
import warnings
warnings.filterwarnings('ignore')
from model import OnTheFlyModel, HitSelectorByOverlap, CommunityDetector, task_evaluator
from myFunctions import *
from morgan_desc import *
from physchem_desc import *
from fragment_embedder import FragmentEmbedder
import onnxruntime as rt

st.set_page_config(
    page_title="Ligand Discovery 5: On-the-Fly ML Model",
    page_icon=":home:",
    layout="wide", # "centered",
    initial_sidebar_state="expanded"
)

st.markdown("""
  <style>
    .css-13sdm1b.e16nr0p33 {
      margin-top: -75px;
    }
  </style>
""", unsafe_allow_html=True)

hide_streamlit_style = """
            <style>
            #MainMenu {visibility: hidden;}
            footer {visibility: hidden;}
            #header {visibility: hidden;}
            </style>
            """
st.markdown(hide_streamlit_style, unsafe_allow_html=True) 

session_id = get_session_id()

RDLogger.DisableLog("rdApp.*")

root = os.path.dirname(os.path.abspath(__file__))

cache_folder = os.path.join(root, ".", "cache")
if not os.path.exists(cache_folder):
    os.mkdir(cache_folder)
clear_old_cache(cache_folder, hours=24)

old_clusters_cache_file = None

df = None

# Functions and variables

CRF_PATTERN = "CC1(CCC#C)N=N1"
CRF_PATTERN_0 = "C#CC"
CRF_PATTERN_1 = "N=N"

crf_0 = "C#CCC1(N=N1)CCNC(CC)=O"
crf_0 = "ONC(=O)CCC1(CCC#C)N=N1"
crf_1 = "C#CCC1(N=N1)CCC(=O)[N]"

SIMILARITY_PERCENTILES = [95, 90]

hits, fid_prom, pid_prom = load_hits()
pid2name, name2pid, any2pid = pid2name_mapper()
fid2smi = load_fid2smi()

st.sidebar.title("Ligand Discovery 5: On-the-Fly Model")
st.sidebar.write("this app builds a quick ML model for your proteins of interest")

st.sidebar.subheader(":mag: UniProt Accession IDs or Gene Name")

text = st.sidebar.text_area("For example, you can query VDAC2", placeholder = "VDAC2", help="Write one protein per line. UniProt AC format is preferred. Only proteins available in the Ligand Discovery interactome will be considered")

input_tokens = text.split()
input_pids = []
for it in input_tokens:
    if it in any2pid:
        pid = any2pid[it]
        if pid in pid_prom:
            input_pids += [any2pid[it]]

input_data = pids_to_dataframe(input_pids, pid2name, pid_prom)

tfidf = True

if input_data.shape[0] == 0:
    has_input = False
    if len(input_tokens) > 0:
        st.sidebar.warning(
            "None of your input proteins was found in the Ligand Discovery interactome.".format(
                len(input_pids), len(input_tokens)
            )
        )
else:
    has_input = True

if has_input:
    print("Instantiating on the fly model")
    model = OnTheFlyModel()
    is_fitted = False
    st.sidebar.info(
        "{0} out of {1} input proteins were found in the Ligand Discovery interactome, corresponding to all statistically significant fragment-protein pairs.".format(
            len(input_pids), len(input_tokens)
        )
    )

    st.sidebar.dataframe(input_data, hide_index=True)

    uniprot_inputs = list(input_data["UniprotAC"])
    if len(uniprot_inputs) == 1:
        print("Only one protein")
        clusters_of_proteins = [uniprot_inputs]
    else:
        graph = get_protein_graph(uniprot_inputs)
        auroc_cut = 0.7
        graph_key = "-".join(sorted(graph.nodes()))
        clusters_cache_file = os.path.join(
            root, "..", "cache", session_id + "_clusters.joblib"
        )
        clusters_of_proteins = None
        if os.path.exists(clusters_cache_file):
            gk, clu = joblib.load(clusters_cache_file)
            if gk == graph_key:
                clusters_of_proteins = clu
        if clusters_of_proteins is None:
            if old_clusters_cache_file is not None:
                os.path.remove(old_clusters_cache_file)
            community_detector = CommunityDetector(tfidf=tfidf, auroc_cut=auroc_cut)
            clusters_of_proteins = community_detector.cluster(model, graph)
            clusters_of_proteins = clusters_of_proteins["ok"]
            joblib.dump((graph_key, clusters_of_proteins), clusters_cache_file)
            old_clusters_cache_file = clusters_cache_file

    cols = st.columns([0.3, 0.7])
    col = cols[0]

    col.subheader(":robot_face: Quick modeling")

    only_one_option = False
    if len(clusters_of_proteins) == 1:
        if sorted(clusters_of_proteins[0]) == sorted(list(input_data["UniprotAC"])):
            only_one_option = True

    options = []
    if not only_one_option:
        for prot_clust in clusters_of_proteins:
            options += [", ".join(sorted([pid2name[pid] for pid in prot_clust]))]
        options += ["Full set of proteins"]
    else:
        if len(clusters_of_proteins[0]) == 1:
            options = [pid2name[clusters_of_proteins[0][0]]]
        else:
            options = ["Full set of proteins"]

    selected_cluster = col.radio(
        "These are some suggested groups of proteins for modeling",
        options=options,
    )

    if selected_cluster == "Full set of proteins":
        selected_cluster = uniprot_inputs
    else:
        selected_cluster = [name2pid[n] for n in selected_cluster.split(", ")]

    print("Cluster selection done")
    default_max_hit_fragments = 100
    default_max_fragment_prom = 500

    max_hit_fragments = col.slider(
        "Maximum number of positives",
        min_value=10,
        max_value=200,
        step=10,
        value=default_max_hit_fragments,
        help="Fragments will be ranked by specificity, i.e. by ascending value of promiscuity.",
    )

    max_fragment_prom = col.slider(
        "Maximum promiscuity of included fragments",
        min_value=50,
        max_value=500,
        step=10,
        value=default_max_fragment_prom,
        help="Maximum number of proteins for included fragments.",
    )

    uniprot_acs = list(selected_cluster)

    data = HitSelectorByOverlap(uniprot_acs=uniprot_acs, tfidf=tfidf).select(
        max_hit_fragments=max_hit_fragments, max_fragment_promiscuity=max_fragment_prom
    )
    print("Selecting hits by overlap")

    num_positives = len(data[data["y"] == 1])
    num_total = len(data[data["y"] != -1])

    subcols = col.columns(3)
    subcols[0].metric(
        "Positives",
        value=num_positives,
        help="Number of positive fragments (i.e. fragments that interact with at least one of the selected proteins). Fragments are ranked by their sum of TF-IDF scores, meaning that the fragments that interact with more proteins will be ranked higher. Interacting with specific proteins will also uprank fragments.",
    )

    subcols[1].metric(
        "Total",
        value=num_total,
        help="Total number of fragments (positive and negative) used in the model. This value decreases as you decrease the maximum promiscuity of included fragments threshold.",
    )

    subcols[2].metric(
        "Rate",
        value="{0:.1f}%".format(num_positives / num_total * 100),
        help="Ratio of positives to total fragments.",
    )

    if num_positives == 0:
        col.error(
            "No positives available. We cannot build a model with no positive data."
        )
        is_fitted = False

    else:
        task_evaluation = task_evaluator(model, data)
        subcols[0].metric(
            label="Corr. prom",
            value="{0:.3f}".format(task_evaluation["ref_rho"]),
            help="Correlation between model outcomes and fragment promiscuity predictors. If you wish to have models that are less correlated with promiscuity, consider lowering the maximum promiscuity of included fragments threshold.",
        )
        subcols[1].metric(
            label="Frag. promiscuity",
            value="{0:.1f}".format(task_evaluation["prom"]),
            help="Average promiscuity of positive fragments. This helps understand how promiscuous the fragments are that are being used to build the model, with a focus on the positive class.",
        )
        subcols[2].metric(
            label="Interactors ({0})".format(len(uniprot_acs)),
            value="{0:.1f}".format(task_evaluation["hits"]),
            help="Average number of query proteins that interact with positive fragments. If you want this number to be higher, consider decreasing the maximum number of positives threshold in order to focus on the fragments that have the highest protein coverage.",
        )
        print("Task evaluator done")

        expander = col.expander("View positives")
        positives_data = data[data["y"] == 1]
        pos_fids = sorted(positives_data["fid"])
        pos_smis = [fid2smi[fid] for fid in pos_fids]
        expander.dataframe(
            pd.DataFrame({"FragmentID": pos_fids, "SMILES": pos_smis}), hide_index=True
        )
        expander = col.expander("View negatives")
        negatives_data = data[data["y"] == 0]
        neg_fids = sorted(negatives_data["fid"])
        neg_smis = [fid2smi[fid] for fid in neg_fids]
        expander.dataframe(
            pd.DataFrame({"FragmentID": neg_fids, "SMILES": neg_smis}), hide_index=True
        )

        if num_positives < 5:
            col.warning("Not enough data to estimate AUROC.")

        else:
            auroc = task_evaluation["auroc"]
            col.metric(
                "AUROC estimation", value="{0:.3f} ± {1:.3f}".format(auroc[0], auroc[1])
            )
            subcols = col.columns(3)

        is_ready = True

    if is_ready:
        col = cols[1]

        col.subheader(":crystal_ball: Make predictions")

        input_prediction_tokens = col.text_area(
            label="Input your SMILES of interest. Ideally, they should have the diazirine fragment",
            help="Paste molecules in SMILES format, one per line. Try to include the CRF pattern in your input molecules. If no CRF pattern is present, it will be automatically attached.",
        )

        col.markdown("Below you can find a few suggested molecules to try out:")
        col.text("\n".join(pos_smis[:2] + neg_smis[:2]))

        pred_tokens = [t for t in input_prediction_tokens.split("\n") if t != ""]

        smiles_list = []
        original_smiles_list = []
        have_crf_count = 0
        attached_crf_count = 0
        for token in pred_tokens:
            if not is_valid_smiles(token):
                continue
            smi = token
            if not has_crf(Chem.MolFromSmiles(smi), CRF_PATTERN):
                smi = attach_crf(smi)
                if smi is None:
                    continue
                attached_crf_count += 1
            else:
                have_crf_count += 1
            smiles_list += [smi]
            original_smiles_list += [token]

        if len(smiles_list) == 0:
            has_prediction_input = False
            if len(pred_tokens) > 0:
                col.warning(
                    "No valid inputs were found. Please make sure the CRF pattern is present: {0}".format(
                        CRF_PATTERN
                    )
                )
        else:
            has_prediction_input = True

        if has_prediction_input:
            model.fit(data["y"])

            col.info(
                "{0} out of {1} input molecules are valid. Of these, {2} had the CRF already, and for {3} of them it was automatically attached".format(
                    len(smiles_list),
                    len(pred_tokens),
                    have_crf_count,
                    attached_crf_count,
                )
            )
            do_tau = False
            if do_tau:
                y_hat, tau_ref, tau_train = model.predict_proba_and_tau(smiles_list)
                dr = pd.DataFrame(
                    {
                        "SMILES": smiles_list,
                        "OriginalSMILES": original_smiles_list,
                        "Score": y_hat,
                        "Tau": tau_ref,
                        "TauTrain": tau_train,
                    }
                )
                for v in dr.values:
                    expander = col.expander(
                        "Score: `{0:.3f}` | Tau: `{1:.2f}` | Tau Train: `{2:.2f}`| SMILES: `{3}`".format(
                            v[1], v[2], v[3], v[0]
                        )
                    )
                    expander.image(get_fragment_image(v[0]))
            else:
                y_hat = model.predict_proba(smiles_list)[:, 1]
                dr = pd.DataFrame(
                    {
                        "SMILES": smiles_list,
                        "OriginalSMILES": original_smiles_list,
                        "Score": y_hat,
                    }
                )
                for v in dr.values:
                    expander = col.expander(
                        "Score: `{0:.3f}` | SMILES: `{1}`".format(v[2], v[1])
                    )
                    expander.image(get_fragment_image(v[1]))
            dr_ = dr[["OriginalSMILES", "Score"]]
            dr_.rename(columns={"OriginalSMILES": "SMILES"}, inplace=True)
            col.download_button(
                label="Download results as CSV",
                data=dr_.to_csv(),
                file_name="prediction_output.csv",
                mime="text/csv",
            )
            del model