app.py

import streamlit as st
import os
import pandas as pd
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity, manhattan_distances, euclidean_distances

st.set_page_config(
    page_title="PhenoGene",
    page_icon="🧑‍💻",
    layout="wide",
    menu_items={
        'Get Help': 'https://www.ncbi.nlm.nih.gov/research/bionlp/',
        'About': "PhenoGene v1.0"
    })

# Constants
embs = []

# Heading
st.title('PhenoGene Interactive Demo')

with st.expander("About", expanded=True):

    st.write(
        """     
-   PhenoGene is a novel gene prioritization method capable of representing HPO terms into embeddings. Utilizing advanced graph embeddings methods, PhenoGene can learn an effective mapping between genes and HPO terms.
-   Given a list of HPO terms, we compute the similarity with a Gene
-   **Input:** List of HPO terms
-   **Output:** Similarity score to the genes
-   Contact: [NLM/NCBI BioNLP Research Group](https://www.ncbi.nlm.nih.gov/research/bionlp/)
	    """
    )

    st.markdown("")



# Gene File, 128 dim embeddings
gdf = pd.read_csv("data/diff2vec_gene_embd.csv").set_index("gene")
genes = gdf.index.tolist()
gene_emb = gdf.values
gene_emb_data = {}
for x,y in zip(genes, gene_emb):
    gene_emb_data[x] = y.reshape(1, 128)
st.session_state['gene_emb_data'] = gene_emb_data


# All HPO embeddings, 128 dim embeddings
hdf = pd.read_csv("data/diff2vec_hpo_embd.csv").set_index("hpo_id")
hpos = hdf.index.tolist()
hpo_emb = hdf.values
hpo_emb_data = {}
for x,y in zip(hpos, hpo_emb):
    hpo_emb_data[x] = y.reshape(1, 128)
st.session_state['hpo_emb_data'] = hpo_emb_data

@st.cache(allow_output_mutation=True)
def compute_similarity_with_gene(emb_src, genes, distance_metric='cosine'):
    data = {}
    for g in genes:
        if distance_metric == "cosine":
            data[g] = cosine_similarity(emb_src, gene_emb_data[g]).item()
    df = pd.DataFrame(data=data.items(), columns=['gene', 'score']).sort_values(by='score', ascending=False).reset_index(drop=True)
    return df

@st.cache
def convert_df(df):
    # IMPORTANT: Cache the conversion to prevent computation on every rerun
    return df.to_csv().encode('utf-8')

st.subheader("🖮 Enter HPO terms separated by a comma")
hpo_terms_text = st.text_area('Example: HP_0000006, HP_0000006', "HP_0000006, HP_0000006")
hpo_terms = list(map(str.strip, hpo_terms_text.split(',')))

#st.write("HPO Terms entered: ")
#st.write(hpo_terms)

st.subheader("💻 Hit Compute to calculate similarity to gene")
metrics = 'cosine'

no_emb = False


if st.button("Compute"):
    with st.spinner('Computing...'):
        for h in hpo_terms:
            if h not in hpo_emb_data.keys():
                st.error("No Embeddings.")
                no_emb = True
                break
            embs.append(hpo_emb_data[h])
        embs_mean = np.array(embs).mean(axis=0)
        result_df = compute_similarity_with_gene(embs_mean, genes, distance_metric=metrics)
    
    if no_emb:
        st.error("Embedding Error.")
    else:
        csv = convert_df(result_df)
        st.success("Done!")
        st.dataframe(result_df)

        st.download_button(
        label="Download results as CSV",
        data=csv,
        mime='text/csv',
)