app.py

import streamlit as st
st.set_page_config(layout="wide")

import pandas as pd
import numpy as np
from zipfile import ZipFile

import plotly.express as px
import plotly.graph_objs as go

LLR_FILE = 'UniProtKB_human_VESM_3B_llrs.zip'

df = pd.read_csv('UniProtKB_id_names.csv', index_col=0)
if 'shuffled_df' not in st.session_state:
    st.session_state.shuffled_df = df.sample(frac=1)
df = st.session_state.shuffled_df
clinvar = pd.read_csv('clinvar_0325.csv.gz',index_col=0)

f = np.load("logreg_params.npz")
coef, intercept = f["coef"].item(), f["intercept"].item()

def load_LLR(uniprot_id):
    '''Loads the LLRs for a given uniprot id. Returns a 20xL dataframe. 
    Rows are indexed by AA change, 
    (AAorder=['K','R','H','E','D','N','Q','T','S','C','G','A','V','L','I','M','P','Y','F','W'])
    Columns indexed by WT_AA+position e.g., "G 12".
    Usage: load_LLR('P01116') or load_LLR('P01116-2')
    '''
    with ZipFile(LLR_FILE) as myzip:
        data = myzip.open(myzip.namelist()[0] + 'LLRs/' + uniprot_id + '.csv')
    LLR = pd.read_csv(data, index_col=0)
    if sigmoid:
        p = 1/(1 + np.exp(-(LLR.values.ravel()*coef + intercept)))
        LLR = pd.DataFrame(p.reshape(LLR.shape), index=LLR.index, columns=LLR.columns).round(4)
    return LLR

def meltLLR(LLR, gene_prefix=None, ignore_pos=False):
    vars = LLR.melt(ignore_index=False)
    vars['variant'] = [''.join(i.split(' ')) + j for i, j in zip(vars['variable'], vars.index)]
    vars['score'] = vars['value']
    vars = vars.set_index('variant')
    if not ignore_pos:
        vars['pos'] = [int(i[1:-1]) for i in vars.index]
    del vars['variable'], vars['value']
    if gene_prefix is not None:
        vars.index = gene_prefix + '_' + vars.index
    return vars
    
    
def plot_interactive(uniprot_id, show_clinvar=False):
    primaryLLR = load_LLR(uniprot_id)
    template = 'plotly_white'
    zmax=1.1 if sigmoid else 0
    zmin=0 if sigmoid else -18
    cmap='rdbu_r' if sigmoid else 'Viridis_r'
    color = 'score' if sigmoid else 'LLR' 
    fig = px.imshow(
        primaryLLR.values, 
        x=primaryLLR.columns, 
        y=primaryLLR.index, 
        color_continuous_scale=cmap, 
        zmax=zmax, 
        zmin=zmin,
        labels=dict(y="Amino acid change", x="Protein sequence", color=color),
        template=template,
        title=selection
    )

    fig.update_xaxes(tickangle=-90,range=[0,99], rangeslider=dict(visible=True), dtick=1)
    fig.update_yaxes(dtick=1)
    fig.update_layout(
        plot_bgcolor='rgba(0, 0, 0, 0)',
        paper_bgcolor='rgba(0, 0, 0, 0)',
        font={'family': 'Arial', 'size': 11},
        hoverlabel=dict(font=dict(family='Arial', size=14))
    )

    fig.update_traces(
        hovertemplate="<br>".join(["<b>%{x} %{y}</b> (%{z:.2f})"]) + '<extra></extra>'
    )

    if show_clinvar:
        iso_clinvar = clinvar[clinvar.protein == uniprot_id]
        iso_clinvar = iso_clinvar[iso_clinvar.GoldStars > 1]
        b_mut = set(iso_clinvar[iso_clinvar.clinvar_label == 0.0].variant.values)
        p_mut = set(iso_clinvar[iso_clinvar.clinvar_label == 1.0].variant.values)

        hwt_x, hwt_y, cust = [], [], []
        phwt_x, phwt_y, pcust = [], [], []

        for i in primaryLLR.columns:
            for j in list(primaryLLR.index):
                mut = i[0] + i[2:] + j
                if mut in b_mut:
                    hwt_x.append(i)
                    hwt_y.append(j)
                    cust.append(primaryLLR.loc[j, i])
                elif mut in p_mut:
                    phwt_x.append(i)
                    phwt_y.append(j)
                    pcust.append(primaryLLR.loc[j, i])

        # draw pathogenic
        fig.add_trace(go.Scatter(
            x=phwt_x, y=phwt_y, customdata=pcust,
            mode='markers',
            marker=dict(size=8, color='red'),
            showlegend=False,
            hoverlabel=dict(bgcolor='crimson', font_color='black'),
            hovertemplate="<b>%{x} %{y}</b> (%{customdata:.2f})<extra></extra>"
        ))
        # draw benign
        fig.add_trace(go.Scatter(
            x=hwt_x, y=hwt_y, customdata=cust,
            mode='markers',
            marker=dict(size=8, color='white'),
            showlegend=False,
            hoverlabel=dict(bgcolor='white', font_color='black'),
            hovertemplate="<b>%{x} %{y}</b> (%{customdata:.2f})<extra></extra>"
        ))

        fig.update_layout(
            hovermode='closest',
            hoverdistance=10
        )

    return fig


idx = df.index.get_loc('P32245') if 'P32245' in df.index else 0
selection = st.selectbox("uniprot_id:", df, index=idx)
uid = df[df.txt == selection].index.values[0]

col1, col2 = st.columns(2)
with col1:
    sigmoid = st.checkbox(
        "Calibrated VESM predictions (0: benign, 1: pathogenic)",
        value=False
    )
with col2:
    show_clinvar = st.checkbox(
        "Show ClinVar annotations (red: pathogenic, white: benign)",
        value=False
    )

fig = plot_interactive(uid, show_clinvar=show_clinvar)
fig.update_layout(width=800, height=600, autosize=False)
st.plotly_chart(fig, use_container_width=True)

st.download_button(
    label="📥 Download as CSV",
    data=meltLLR(load_LLR(uid)).to_csv(),
    file_name=f"{selection}.csv",
    mime='text/csv'
)
st.markdown("---")

st.markdown("""  
- Bulk download precomputed scores at [VESM Effect Scores](https://huggingface.co/datasets/ntranoslab/vesm_scores) for all UniProt, hg19, and hg38 variants.
- Use VESM locally: Access the source code and installation instructions on [GitHub](https://github.com/ntranoslab/vesm).
- By default the displayed scores are log-likelihood ratios (LLRs) obtained from our largest sequence-only model (VESM-3B). 
""")