app.py

import time

import pandas as pd
import gradio as gr
from gradio.themes.utils import sizes
from gradio_leaderboard import Leaderboard
from dotenv import load_dotenv

load_dotenv()  # Load environment variables from .env file (before imports)

from about import ABOUT_INTRO, ABOUT_TEXT, FAQS, SUBMIT_INSTRUCTIONS, WEBSITE_HEADER
from constants import (
    ASSAY_RENAME,  # noqa: F401
    SECOND_DEADLINE,
    SEQUENCES_FILE_DICT,
    LEADERBOARD_DISPLAY_COLUMNS,
    ABOUT_TAB_NAME,
    FAQ_TAB_NAME,
    TERMS_URL,
    LEADERBOARD_COLUMNS_RENAME,
    LEADERBOARD_COLUMNS_RENAME_LIST,
    SUBMIT_TAB_NAME,
    SLACK_URL,
)
from submit import make_submission
from utils import fetch_hf_results, show_output_box, periodic_data_fetch


def format_leaderboard_table(df_results: pd.DataFrame, assay: str | None = None):
    """
    Format the dataframe for display on the leaderboard. The dataframe comes from utils.fetch_hf_results().
    """
    df = df_results.query("assay.isin(@ASSAY_RENAME.keys())").copy()
    if assay is not None:
        df = df[df["assay"] == assay]
    df = df[LEADERBOARD_DISPLAY_COLUMNS]
    df = df.sort_values(by="spearman", ascending=False)
    # After sorting, just add the reason for excluding heldout test set
    # Note: We can also just say the following as a text box at the bottom of the leaderboard: "Note: Results for the Heldout Test Set are only evaluated at competition close"
    # Convert spearman column to string to avoid dtype incompatibility when assigning text
    df["spearman"] = df["spearman"].apply(lambda x: f"{x:.3f}")
    # Cast submission_time to datetime
    df["submission_time"] = pd.to_datetime(df["submission_time"], errors="coerce")
    # Before the first deadline: Say we're busy evaluating
    for metric in ["spearman", "top_10_recall"]:
        if metric not in df.columns:
            continue
        df.loc[
            (df["dataset"] == "Heldout Test Set")
            & (df[metric] == "nan")
            & (df["submission_time"] <= SECOND_DEADLINE),
            metric,
        ] = "Error: All predictions have a constant value"

    # Finally, rename columns for readability
    df = df.rename(columns=LEADERBOARD_COLUMNS_RENAME)
    return df


def get_leaderboard_object(assay: str | None = None):
    filter_columns = ["dataset"]
    if assay is None:
        filter_columns.append("property")
    # Bug: Can't leave search_columns empty because then it says "Column None not found in headers"
    # Note(Lood): Would be nice to make it clear that the Search Column is searching on model name
    current_dataframe = pd.read_csv("debug-current-results.csv")
    lb = Leaderboard(
        value=format_leaderboard_table(df_results=current_dataframe, assay=assay),
        datatype=["str", "str", "str", "number", "str"],
        select_columns=LEADERBOARD_COLUMNS_RENAME_LIST(
            ["model", "property", "spearman", "dataset", "user"]
        ),
        search_columns=["Model Name"],
        filter_columns=LEADERBOARD_COLUMNS_RENAME_LIST(filter_columns),
        every=15,
        render=True,
        height=500,  # Set a fixed height to make it scrollable
    )
    return lb


def refresh_overall_leaderboard():
    # debug-current-results.csv is updated by the outer thread
    current_dataframe = pd.read_csv("debug-current-results.csv")
    return format_leaderboard_table(df_results=current_dataframe)


# Initialize global dataframe
fetch_hf_results()
time.sleep(2)  # Give the outer thread time to create the file at the start
current_dataframe = pd.read_csv("debug-current-results.csv")

# Make font size bigger using gradio theme
with gr.Blocks(theme=gr.themes.Default(text_size=sizes.text_lg)) as demo:
    timer = gr.Timer(3)  # Run every 3 seconds when page is focused

    ## Header

    with gr.Row():
        with gr.Column(scale=6):  # bigger text area
            gr.Markdown(WEBSITE_HEADER)
        with gr.Column(scale=2):  # smaller side column for logo
            gr.Image(
                value="./assets/competition_logo.jpg",
                show_label=False,
                show_download_button=False,
                show_share_button=False,
                show_fullscreen_button=False,
                width="25vw",  # Take up the width of the column (2/8 = 1/4)
            )

    with gr.Tabs(elem_classes="tab-buttons"):
        with gr.TabItem(ABOUT_TAB_NAME, elem_id="abdev-benchmark-tab-table"):
            gr.Markdown(ABOUT_INTRO)
            gr.Image(
                value="./assets/prediction_explainer_v3.png",
                show_label=False,
                show_download_button=False,
                show_share_button=False,
                show_fullscreen_button=False,
                width="30vw",
            )
            gr.Markdown(ABOUT_TEXT)

            # Sequence download buttons
            gr.Markdown(
                """### 📥 Download Sequences
            The GDPa1 dataset (with assay data and sequences) is available on Hugging Face [here](https://huggingface.co/datasets/ginkgo-datapoints/GDPa1),
            but we provide this and the private test set for convenience."""
            )
            with gr.Row():
                with gr.Column():
                    download_button_cv_about = gr.DownloadButton(
                        label="📥 Download GDPa1 sequences",
                        value=SEQUENCES_FILE_DICT["GDPa1_cross_validation"],
                        variant="secondary",
                    )
                with gr.Column():
                    download_button_test_about = gr.DownloadButton(
                        label="📥 Download Private Test Set sequences",
                        value=SEQUENCES_FILE_DICT["Heldout Test Set"],
                        variant="secondary",
                    )

        with gr.TabItem(
            "🏆 Leaderboard", elem_id="abdev-benchmark-tab-table"
        ) as leaderboard_tab:
            gr.Markdown(
                """
                # Overall Leaderboard (filter below by property)
                Each property has its own prize, and participants can submit models for any combination of properties.

                **Note**: It is *easy to overfit* the public GDPa1 dataset, which results in artificially high Spearman correlations.
                We would suggest training using cross-validation to give a better indication of the model's performance on the eventual private test set.
                """
            )
            lb = get_leaderboard_object()
            timer.tick(fn=refresh_overall_leaderboard, outputs=lb)
            demo.load(fn=refresh_overall_leaderboard, outputs=lb)
        with gr.Tab(FAQ_TAB_NAME):
            gr.Markdown("# Frequently Asked Questions")
            for i, (question, answer) in enumerate(FAQS.items()):
                # Would love to make questions bold but accordion doesn't support it
                question = f"{i+1}. {question}"
                with gr.Accordion(question, open=False):
                    if isinstance(answer, list):
                        # Italicize each line
                        italicized_answer = "  \n".join(f"*{item}*" for item in answer)
                        gr.Markdown(italicized_answer)
                    else:
                        gr.Markdown(f"*{answer}*")  # Italics for answers

    # Footnote
    gr.Markdown(
        f"""
        <div style="text-align: center; font-size: 14px; color: gray; margin-top: 2em;">
        📬 For questions or feedback, contact <a href="mailto:[email protected]">[email protected]</a> or discuss on the <a href="{SLACK_URL}">Slack community</a> co-hosted by Bits in Bio.<br>
        Visit the <a href="https://datapoints.ginkgo.bio/ai-competitions/2025-abdev-competition">Competition Registration page</a> to sign up for updates and to register, and see Terms <a href="{TERMS_URL}">here</a>.
        </div>
        """,
        elem_id="contact-footer",
    )

if __name__ == "__main__":
    demo.launch(
        ssr_mode=False,
        app_kwargs={"lifespan": periodic_data_fetch},
    )