app.py

from __future__ import annotations

import csv
import os
import tempfile
from pathlib import Path
from typing import Any

import gradio as gr
import numpy as np
import pandas as pd

from space_runtime import (
    AssayQuery,
    load_compatibility_model_from_hub,
    molecule_ui_metrics,
    rank_compounds,
    serialize_assay_query,
)

MODEL_REPO_ID = os.getenv("MODEL_REPO_ID", "lighteternal/BioAssayAlign-Qwen3-Embedding-0.6B-Compatibility")
MAX_INPUT_SMILES = int(os.getenv("MAX_INPUT_SMILES", "3000"))
DEFAULT_TOP_K = int(os.getenv("DEFAULT_TOP_K", "50"))
ENABLE_BACKGROUND_WARMUP = os.getenv("ENABLE_BACKGROUND_WARMUP", "0") == "1"

CSS = """
@import url('https://fonts.googleapis.com/css2?family=IBM+Plex+Sans:wght@400;500;600;700&family=IBM+Plex+Mono:wght@400;500&family=Fraunces:opsz,wght@9..144,600;9..144,700&display=swap');

:root {
  --paper: #f6f1e6;
  --paper-deep: #ece4d2;
  --ink: #17252d;
  --ink-soft: #5f6d75;
  --accent: #165b55;
  --accent-deep: #0c4641;
  --accent-soft: #dbeee8;
  --accent-warm: #b8643d;
  --accent-bright: #d06a3b;
  --accent-bright-deep: #b25124;
  --accent-warm-soft: #f3e1d5;
  --line: #c6cdbf;
  --warning: #8a4b0f;
  --good: #0e6b48;
  --card: rgba(255, 252, 246, 0.9);
  --card-strong: rgba(255, 255, 255, 0.96);
  --shadow: 0 20px 45px rgba(23, 37, 45, 0.08);
}

.gradio-container {
  --body-text-color: var(--ink);
  --color-text-body: var(--ink);
  --block-title-text-color: var(--ink);
  --input-text-color: var(--ink);
  --input-placeholder-color: #6d7c83;
  font-family: "IBM Plex Sans", sans-serif;
  background:
    radial-gradient(circle at 10% 0%, rgba(22,91,85,0.13), transparent 26rem),
    radial-gradient(circle at 92% 8%, rgba(184,100,61,0.12), transparent 24rem),
    linear-gradient(180deg, #fbf8f1 0%, var(--paper) 100%);
  color: var(--ink);
}

#hero {
  border: 1px solid var(--line);
  background:
    linear-gradient(135deg, rgba(255,255,255,0.98), rgba(241,247,245,0.9)),
    linear-gradient(90deg, rgba(22,91,85,0.06), rgba(184,100,61,0.04));
  border-radius: 30px;
  padding: 1.5rem 1.65rem;
  box-shadow: var(--shadow);
}

.eyebrow {
  font-family: "IBM Plex Mono", monospace;
  font-size: 0.78rem;
  letter-spacing: 0.08em;
  text-transform: uppercase;
  color: var(--accent-warm);
}

.hero-title {
  font-family: "Fraunces", serif;
  font-size: 2.35rem;
  line-height: 1.05;
  margin: 0.2rem 0 0.5rem 0;
}

.hero-copy {
  color: var(--ink-soft);
  max-width: 60rem;
  font-size: 1rem;
}

.hero-grid {
  display: grid;
  grid-template-columns: minmax(0, 1.6fr) minmax(19rem, 0.9fr);
  gap: 1.1rem;
  align-items: start;
}

.hero-side {
  background: rgba(255,255,255,0.75);
  border: 1px solid rgba(198,205,191,0.8);
  border-radius: 20px;
  padding: 1rem 1.05rem;
}

.hero-side-title {
  font-family: "IBM Plex Mono", monospace;
  font-size: 0.74rem;
  letter-spacing: 0.08em;
  text-transform: uppercase;
  color: var(--accent-warm);
  margin-bottom: 0.55rem;
}

.hero-list {
  margin: 0;
  padding-left: 1rem;
  color: var(--ink);
}

.hero-list li + li {
  margin-top: 0.45rem;
}

.metric-strip {
  display: grid;
  grid-template-columns: minmax(0, 1.2fr) minmax(0, 1fr);
  gap: 0.8rem;
}

.metric-card {
  border: 1px solid var(--line);
  background: linear-gradient(180deg, rgba(255,255,255,0.92), rgba(248,244,236,0.9));
  padding: 0.72rem 0.85rem;
  border-radius: 18px;
  min-height: 4.9rem;
  box-shadow: 0 8px 24px rgba(23,37,45,0.04);
}

.metric-card span {
  color: var(--ink-soft);
  display: block;
}

.metric-card strong {
  display: block;
  font-size: 1rem;
  margin-top: 0.15rem;
  color: var(--ink);
}

.metric-card a {
  color: var(--ink);
  text-decoration: none;
}

.metric-card a:hover {
  color: var(--accent-deep);
  text-decoration: underline;
}

.compact-spec {
  margin: 0.7rem 0 1rem 0;
  border: 1px solid var(--line);
  border-radius: 18px;
  background: rgba(255,255,255,0.82);
  padding: 0.78rem 0.9rem;
  color: var(--ink-soft);
  font-size: 0.92rem;
  line-height: 1.45;
}

.workspace {
  gap: 1rem !important;
  align-items: stretch;
}

.workspace > div {
  gap: 1rem !important;
}

.pane {
  background: linear-gradient(180deg, rgba(255,255,255,0.94), rgba(248,244,236,0.92));
  border: 1px solid var(--line);
  border-radius: 24px;
  padding: 0.95rem 1rem 1rem 1rem;
  box-shadow: var(--shadow);
}

.pane-header {
  display: flex;
  align-items: baseline;
  justify-content: space-between;
  gap: 0.75rem;
  margin-bottom: 0.9rem;
}

.pane-title {
  font-family: "Fraunces", serif;
  font-size: 1.35rem;
  line-height: 1.1;
}

.pane-kicker {
  font-family: "IBM Plex Mono", monospace;
  font-size: 0.72rem;
  letter-spacing: 0.08em;
  text-transform: uppercase;
  color: var(--accent-warm);
}

.pane-copy {
  color: var(--ink-soft);
  font-size: 0.95rem;
  margin-bottom: 1rem;
}

.helper-row {
  display: grid;
  grid-template-columns: repeat(3, minmax(0, 1fr));
  gap: 0.7rem;
  margin-bottom: 0.8rem;
}

.helper-chip {
  background: var(--accent-soft);
  border: 1px solid rgba(22,91,85,0.12);
  border-radius: 14px;
  padding: 0.75rem 0.8rem;
}

.helper-chip strong {
  display: block;
  margin-bottom: 0.15rem;
}

.section-note {
  color: var(--ink-soft);
  font-size: 0.88rem;
  margin: 0.1rem 0 0.32rem 0;
}

.action-row {
  display: flex;
  gap: 0.7rem;
  flex-wrap: wrap;
}

.summary-shell,
.results-shell {
  background: linear-gradient(180deg, rgba(255,255,255,0.94), rgba(248,244,236,0.92));
  border: 1px solid var(--line);
  border-radius: 24px;
  padding: 0.9rem 1rem;
  box-shadow: var(--shadow);
}

.results-shell,
.results-shell *,
.results-shell p,
.results-shell li,
.results-shell strong,
.results-shell code {
  color: var(--ink) !important;
}

.results-shell code,
.summary-shell code,
.guide-card code,
.footer-note code,
.section-note code,
.pane code {
  background: rgba(22, 91, 85, 0.08) !important;
  color: var(--ink) !important;
  border-radius: 8px !important;
  padding: 0.08rem 0.35rem !important;
  box-shadow: none !important;
}

.results-callout {
  padding: 0.15rem 0.1rem 0.2rem 0.1rem;
}

.results-callout h3 {
  margin: 0 0 0.55rem 0;
  font-family: "Fraunces", serif;
  font-size: 1.1rem;
}

.results-callout ul {
  margin: 0;
  padding-left: 1rem;
}

.results-callout li + li {
  margin-top: 0.28rem;
}

.results-callout p {
  margin: 0.7rem 0 0 0;
  color: var(--ink-soft) !important;
  line-height: 1.5;
}

.results-metrics {
  display: grid;
  grid-template-columns: repeat(3, minmax(0, 1fr));
  gap: 0.65rem;
  margin-top: 0.55rem;
}

.results-metric {
  border: 1px solid var(--line);
  border-radius: 14px;
  background: rgba(255,255,255,0.9);
  padding: 0.65rem 0.75rem;
}

.results-metric strong {
  display: block;
  margin-bottom: 0.16rem;
}

.examples-note {
  color: var(--ink-soft);
  font-size: 0.92rem;
  margin-top: 0.1rem;
}

.footer-note {
  color: var(--ink-soft);
  font-size: 0.9rem;
  line-height: 1.5;
}

.gradio-container .prose,
.gradio-container .prose *,
.gradio-container label,
.gradio-container .label-wrap,
.gradio-container .label-wrap span,
.gradio-container [data-testid="block-info"],
.gradio-container .gr-form,
.gradio-container .gr-button,
.gradio-container input,
.gradio-container textarea,
.gradio-container select,
.gradio-container .wrap,
.gradio-container .wrap textarea,
.gradio-container .wrap input,
.gradio-container table,
.gradio-container th,
.gradio-container td {
  color: var(--ink) !important;
}

.gradio-container .block,
.gradio-container .block > label,
.gradio-container .block .container,
.gradio-container .block .input-container,
.gradio-container .block .wrap,
.gradio-container .block .wrap-inner,
.gradio-container .block fieldset,
.gradio-container .block .form,
.gradio-container .block .inner,
.gradio-container [data-testid="textbox"],
.gradio-container [data-testid="dropdown"],
.gradio-container [data-testid="number"],
.gradio-container [data-testid="file-upload"],
.gradio-container [data-testid="file"],
.gradio-container [data-testid="accordion"],
.gradio-container [data-testid="dataframe"] {
  color: var(--ink) !important;
}

.gradio-container input,
.gradio-container textarea,
.gradio-container select,
.gradio-container [role="combobox"],
.gradio-container [role="listbox"],
.gradio-container [role="option"],
.gradio-container .choices,
.gradio-container .choices__inner,
.gradio-container .choices__list,
.gradio-container .choices__item {
  background: #fffdf8 !important;
  border: 1px solid #e3e8e0 !important;
  border-radius: 14px !important;
  box-shadow: inset 0 1px 0 rgba(255,255,255,0.7) !important;
  color: var(--ink) !important;
}

.gradio-container [role="listbox"],
.gradio-container .choices__list,
.gradio-container .choices__item {
  background: #fffdf8 !important;
  color: var(--ink) !important;
}

.gradio-container [role="option"][aria-selected="true"],
.gradio-container .choices__item--selectable.is-highlighted {
  background: #e7f0ed !important;
  color: #0d3d38 !important;
}

.gradio-container input:focus,
.gradio-container textarea:focus,
.gradio-container select:focus {
  border-color: var(--accent) !important;
  box-shadow: 0 0 0 3px rgba(22,91,85,0.12) !important;
}

.gradio-container .block > label,
.gradio-container .block .container.show_textbox_border,
.gradio-container .block .input-container,
.gradio-container .block .wrap,
.gradio-container .block .wrap-inner,
.gradio-container .block fieldset,
.gradio-container [data-testid="textbox"],
.gradio-container [data-testid="dropdown"],
.gradio-container [data-testid="number"],
.gradio-container [data-testid="file-upload"] {
  background: #fffdf8 !important;
  border-color: #e3e8e0 !important;
  box-shadow: none !important;
  outline: none !important;
}

.gradio-container .block > label,
.gradio-container .block .container.show_textbox_border,
.gradio-container .block .wrap,
.gradio-container .block .wrap-inner,
.gradio-container .block fieldset {
  border: 1px solid #e3e8e0 !important;
  border-radius: 14px !important;
}

.gradio-container .block > label:focus-within,
.gradio-container .block .container.show_textbox_border:focus-within,
.gradio-container .block .wrap:focus-within,
.gradio-container .block .wrap-inner:focus-within,
.gradio-container .block fieldset:focus-within {
  border-color: rgba(22, 91, 85, 0.4) !important;
  box-shadow: 0 0 0 3px rgba(22, 91, 85, 0.08) !important;
}

.gradio-container .form,
.gradio-container .form.svelte-d5xbca,
.gradio-container .form > .block,
.gradio-container .form > .block.padded,
.gradio-container .form > .block.auto-margin {
  background: transparent !important;
  border: none !important;
  box-shadow: none !important;
}

.gradio-container .block:has(textarea[data-testid="textbox"]),
.gradio-container .block:has(input),
.gradio-container .block:has(select),
.gradio-container .block:has([data-testid="file-upload"]),
.gradio-container .block:has(.svelte-1xfsv4t.container) {
  border-color: transparent !important;
  background: transparent !important;
  box-shadow: none !important;
}

.gradio-container button {
  border-radius: 14px !important;
  font-weight: 600 !important;
  transition: transform 0.18s ease, box-shadow 0.18s ease, background 0.18s ease !important;
}

.gradio-container button:hover {
  transform: translateY(-1px);
}

.gradio-container button.primary {
  background: linear-gradient(180deg, var(--accent-bright), var(--accent-bright-deep)) !important;
  color: #f7fbfa !important;
  border: none !important;
  box-shadow: 0 14px 26px rgba(184,100,61,0.24) !important;
}

.gradio-container button.secondary {
  background: linear-gradient(180deg, #f6e7da, #f0ddcd) !important;
  color: var(--ink) !important;
  border: 1px solid #dbc4b4 !important;
}

.gradio-container .tab-nav button {
  color: var(--ink) !important;
  font-weight: 600 !important;
  border-radius: 999px !important;
  padding: 0.55rem 1rem !important;
  background: rgba(255,255,255,0.62) !important;
  border: 1px solid rgba(198,205,191,0.7) !important;
}

.gradio-container .tab-nav button.selected,
.gradio-container button[role="tab"][aria-selected="true"],
.gradio-container [role="tab"][aria-selected="true"] {
  background: linear-gradient(180deg, #d9ece8, #cde4de) !important;
  color: #0d3d38 !important;
  box-shadow: 0 6px 16px rgba(22,91,85,0.12);
  border-color: rgba(22,91,85,0.18) !important;
}

.gradio-container .gradio-dataframe table,
.gradio-container .gradio-dataframe th,
.gradio-container .gradio-dataframe td {
  background: #fffdf7 !important;
}

.gradio-container .gradio-dataframe,
.gradio-container .gradio-dataframe .wrap,
.gradio-container .gradio-dataframe .table-wrap,
.gradio-container .gradio-dataframe .scrollable,
.gradio-container .gradio-dataframe .table-container,
.gradio-container .gradio-dataframe .cell-wrap,
.gradio-container .gradio-dataframe .cell-input,
.gradio-container .gradio-dataframe .cell-output,
.gradio-container [data-testid="dataframe"],
.gradio-container [data-testid="dataframe"] .wrap,
.gradio-container [data-testid="dataframe"] .table-wrap,
.gradio-container [data-testid="dataframe"] .scrollable,
.gradio-container [data-testid="dataframe"] .cell-wrap,
.gradio-container [data-testid="dataframe"] .cell-input,
.gradio-container [data-testid="dataframe"] .cell-output,
.result-frame,
.result-frame *,
.result-frame .wrap,
.result-frame .table-wrap,
.result-frame .table-container,
.result-frame .scrollable {
  background: #fffdf7 !important;
  border-color: var(--line) !important;
  color: var(--ink) !important;
}

.gradio-container .gradio-dataframe th {
  background: #eaf2ef !important;
}

.gradio-container .gradio-dataframe tr:nth-child(even) td,
.gradio-container [data-testid="dataframe"] tr:nth-child(even) td {
  background: #fcfaf4 !important;
}

.gradio-container .gr-accordion,
.gradio-container .gr-accordion *,
.gradio-container [data-testid="accordion"],
.gradio-container [data-testid="accordion"] * {
  color: var(--ink) !important;
}

.result-accordion,
.result-accordion *,
.result-accordion [data-testid="accordion"],
.result-accordion [data-testid="accordion"] * {
  color: var(--ink) !important;
  background: #fffdf7 !important;
}

.result-file,
.result-file *,
.result-file .wrap,
.result-file .file-preview,
.result-file .file-preview-holder,
.result-file [data-testid="file"] {
  background: #fffdf7 !important;
  color: var(--ink) !important;
  border-color: var(--line) !important;
}

.result-file button {
  background: linear-gradient(180deg, #f6e7da, #f0ddcd) !important;
  color: var(--ink) !important;
  border: 1px solid #dbc4b4 !important;
}

.gradio-container .choices__list--dropdown,
.gradio-container .choices__list[aria-expanded],
.gradio-container .choices__list--single,
.gradio-container .choices__list--multiple {
  background: #fffdf8 !important;
  color: var(--ink) !important;
}

.gradio-container .gr-box,
.gradio-container .gr-group,
.gradio-container .gr-accordion {
  border-color: var(--line) !important;
}

.gradio-container [data-testid="file"] button,
.gradio-container [data-testid="download-button"],
.gradio-container .download-button {
  background: linear-gradient(180deg, #f6e7da, #f0ddcd) !important;
  color: var(--ink) !important;
  border: 1px solid #dbc4b4 !important;
}

@media (max-width: 980px) {
  .hero-grid,
  .helper-row,
  .results-metrics,
  .metric-strip {
    grid-template-columns: 1fr;
  }

  .hero-title {
    font-size: 1.95rem;
  }
}
"""

THEME = gr.themes.Soft(
    primary_hue="emerald",
    secondary_hue="stone",
    neutral_hue="slate",
    font=["IBM Plex Sans", "ui-sans-serif", "system-ui", "sans-serif"],
    font_mono=["IBM Plex Mono", "ui-monospace", "monospace"],
)

EXAMPLES = {
    "JAK2 cell assay": {
        "title": "JAK2 inhibition assay",
        "description": "Cell-based luminescence assay measuring JAK2 inhibition in HEK293 cells.",
        "organism": "Homo sapiens",
        "readout": "luminescence",
        "assay_format": "cell-based",
        "assay_type": "inhibition",
        "target_uniprot": "O60674",
        "smiles": "\n".join(
            [
                "CC1=CC(=O)N(C)C(=O)N1",
                "CC(=O)Nc1ncc(C#N)c(Nc2ccc(F)c(Cl)c2)n1",
                "CCOc1ccc2nc(N3CCN(C)CC3)n(C)c(=O)c2c1",
                "CCO",
            ]
        ),
    },
    "ALDH1A1 fluorescence": {
        "title": "ALDH1A1 inhibition assay",
        "description": "Cell-based fluorescence assay measuring ALDH1A1 inhibition in human cells.",
        "organism": "Homo sapiens",
        "readout": "fluorescence",
        "assay_format": "cell-based",
        "assay_type": "inhibition",
        "target_uniprot": "P00352",
        "smiles": "\n".join(
            [
                "CCOC1=CC=CC=C1",
                "CC1=CC(=O)N(C)C(=O)N1",
                "CCN(CC)CCOC1=CC=CC=C1",
                "CCO",
            ]
        ),
    },
    "BTK binding quick check": {
        "title": "BTK kinase inhibitor binding assay",
        "description": "In vitro kinase-domain binding assay for Bruton's tyrosine kinase inhibitor ranking.",
        "organism": "Homo sapiens",
        "readout": "binding",
        "assay_format": "biochemical",
        "assay_type": "binding",
        "target_uniprot": "Q06187",
        "smiles": "\n".join(
            [
                "CC1=NC(=O)N(C)C(=O)N1",
                "c1ccccc1",
                "CCO",
            ]
        ),
    },
}

DEFAULT_EXAMPLE_NAME = "JAK2 cell assay"
DEFAULT_EXAMPLE = EXAMPLES[DEFAULT_EXAMPLE_NAME]


def _parse_smiles_text(value: str | None) -> list[str]:
    if not value:
        return []
    lines = [line.strip() for line in value.replace(",", "\n").splitlines()]
    return [line for line in lines if line]


def _read_uploaded_smiles(file_obj: Any) -> list[str]:
    if file_obj is None:
        return []
    path = Path(file_obj.name if hasattr(file_obj, "name") else str(file_obj))
    suffix = path.suffix.lower()
    if suffix in {".txt", ".smi", ".smiles"}:
        return [line.strip() for line in path.read_text().splitlines() if line.strip()]
    if suffix == ".csv":
        frame = pd.read_csv(path)
        for column in ("smiles", "canonical_smiles", "SMILES"):
            if column in frame.columns:
                return [str(item).strip() for item in frame[column].tolist() if str(item).strip()]
        first = frame.columns[0]
        return [str(item).strip() for item in frame[first].tolist() if str(item).strip()]
    raise gr.Error("Upload a .csv, .txt, .smi, or .smiles file.")


def _collect_smiles(smiles_text: str, upload_file: Any) -> tuple[list[str], str | None]:
    items = _parse_smiles_text(smiles_text) + _read_uploaded_smiles(upload_file)
    deduped: list[str] = []
    seen: set[str] = set()
    for item in items:
        if item not in seen:
            deduped.append(item)
            seen.add(item)
    warning = None
    if len(deduped) > MAX_INPUT_SMILES:
        warning = f"Input truncated to the first {MAX_INPUT_SMILES} unique SMILES for interactive use."
        deduped = deduped[:MAX_INPUT_SMILES]
    return deduped, warning


def _load_model():
    return load_compatibility_model_from_hub(MODEL_REPO_ID)


def _warm_model_background() -> None:
    try:
        _load_model()
    except Exception:
        # Keep the app usable even if warmup fails; the request path will raise the real error.
        return


def _priority_band(relative_score: float, rank: int, total: int) -> str:
    if total <= 3:
        return "Screen first" if rank == 1 else ("Worth a look" if rank == 2 else "Low priority")
    if relative_score >= 85:
        return "Screen first"
    if relative_score >= 60:
        return "Worth a look"
    if relative_score >= 35:
        return "Middle pack"
    return "Low priority"


def _decorate_valid_rows(valid_rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
    if not valid_rows:
        return []
    scores = np.array([float(row["score"]) for row in valid_rows], dtype=np.float32)
    minimum = float(scores.min())
    maximum = float(scores.max())
    spread = maximum - minimum
    decorated: list[dict[str, Any]] = []
    for idx, row in enumerate(valid_rows):
        score = float(row["score"])
        relative_score = 100.0 if spread <= 1e-8 and idx == 0 else (50.0 if spread <= 1e-8 else 100.0 * (score - minimum) / spread)
        metrics = molecule_ui_metrics(row["canonical_smiles"])
        decorated.append(
            {
                **row,
                "relative_score": round(relative_score, 1),
                "priority_band": _priority_band(relative_score, idx + 1, len(valid_rows)),
                "mol_wt": round(float(metrics["mol_wt"]), 1),
                "logp": round(float(metrics["logp"]), 2),
                "tpsa": round(float(metrics["tpsa"]), 1),
                "heavy_atoms": int(metrics["heavy_atoms"]),
            }
        )
    return decorated


def _build_summary(query_text: str, valid_rows: list[dict[str, Any]], invalid_rows: list[dict[str, Any]], warning: str | None) -> str:
    best = valid_rows[0] if valid_rows else None
    bullets = [f"<li><strong>Ranked:</strong> {len(valid_rows)}</li>"]
    if invalid_rows:
        bullets.append(f"<li><strong>Rejected:</strong> {len(invalid_rows)}</li>")
    if best is not None:
        bullets.append(
            f"<li><strong>Top candidate:</strong> <code>{best['canonical_smiles']}</code> · {best['priority_band']} · relative score {best['relative_score']:.1f}/100</li>"
        )
    if warning:
        bullets.append(f"<li><strong>Warning:</strong> {warning}</li>")
    return (
        "<div class='results-callout'>"
        "<h3>Ranking summary</h3>"
        f"<ul>{''.join(bullets)}</ul>"
        "<div class='results-metrics'>"
        "<div class='results-metric'><strong>Priority</strong>Shortlist cue derived from the within-list ranking.</div>"
        "<div class='results-metric'><strong>Relative score</strong>0–100 rescaling inside this submitted list only.</div>"
        "<div class='results-metric'><strong>Model score</strong>Internal logit retained for debugging and reproducibility.</div>"
        "</div>"
        "</div>"
    )


def _results_to_csv(valid_rows: list[dict[str, Any]], invalid_rows: list[dict[str, Any]]) -> str | None:
    rows = valid_rows + invalid_rows
    if not rows:
        return None
    handle = tempfile.NamedTemporaryFile("w", suffix=".csv", delete=False, newline="")
    writer = csv.DictWriter(
        handle,
        fieldnames=[
            "rank",
            "priority_band",
            "relative_score_100",
            "input_smiles",
            "canonical_smiles",
            "smiles_hash",
            "mol_wt",
            "logp",
            "tpsa",
            "heavy_atoms",
            "model_score",
            "valid",
            "error",
        ],
    )
    writer.writeheader()
    rank = 1
    for row in valid_rows:
        writer.writerow(
            {
                "rank": rank,
                "priority_band": row["priority_band"],
                "relative_score_100": row["relative_score"],
                "input_smiles": row["input_smiles"],
                "canonical_smiles": row["canonical_smiles"],
                "smiles_hash": row["smiles_hash"],
                "mol_wt": row["mol_wt"],
                "logp": row["logp"],
                "tpsa": row["tpsa"],
                "heavy_atoms": row["heavy_atoms"],
                "model_score": row["score"],
                "valid": True,
                "error": "",
            }
        )
        rank += 1
    for row in invalid_rows:
        writer.writerow(
            {
                "rank": "",
                "input_smiles": row["input_smiles"],
                "canonical_smiles": "",
                "smiles_hash": "",
                "mol_wt": "",
                "logp": "",
                "tpsa": "",
                "heavy_atoms": "",
                "model_score": "",
                "valid": False,
                "error": row.get("error", "invalid_smiles"),
            }
        )
    handle.close()
    return handle.name


def run_ranking(
    title: str,
    description: str,
    organism: str,
    readout: str,
    assay_format: str,
    assay_type: str,
    target_uniprot: str,
    smiles_text: str,
    upload_file: Any,
    top_k: int,
):
    smiles_values, warning = _collect_smiles(smiles_text, upload_file)
    if not smiles_values:
        raise gr.Error("Provide at least one SMILES entry by paste or file upload.")
    query = AssayQuery(
        title=title or "",
        description=description or "",
        organism=organism or "",
        readout=readout or "",
        assay_format=assay_format or "",
        assay_type=assay_type or "",
        target_uniprot=[token.strip() for token in target_uniprot.split(",") if token.strip()],
    )
    assay_text = serialize_assay_query(query)
    model = _load_model()
    ranked = rank_compounds(model, assay_text=assay_text, smiles_list=smiles_values, top_k=top_k or None)
    valid_rows = [row for row in ranked if row["valid"]]
    invalid_rows = [row for row in ranked if not row["valid"]]
    valid_rows = _decorate_valid_rows(valid_rows)

    display_rows = [
        {
            "rank": idx + 1,
            "priority": row["priority_band"],
            "relative_score_100": row["relative_score"],
            "canonical_smiles": row["canonical_smiles"],
            "mol_wt": row["mol_wt"],
            "logp": row["logp"],
            "tpsa": row["tpsa"],
            "heavy_atoms": row["heavy_atoms"],
            "model_score": round(float(row["score"]), 4),
        }
        for idx, row in enumerate(valid_rows)
    ]
    invalid_display = [
        {"input_smiles": row["input_smiles"], "error": row.get("error", "invalid_smiles")}
        for row in invalid_rows
    ]
    summary = _build_summary(assay_text, valid_rows, invalid_rows, warning)
    csv_path = _results_to_csv(valid_rows, invalid_rows)
    invalid_frame = pd.DataFrame(invalid_display) if invalid_display else pd.DataFrame(columns=["input_smiles", "error"])
    invalid_title = gr.update(visible=bool(invalid_display))
    invalid_table = gr.update(value=invalid_frame, visible=bool(invalid_display))
    return summary, assay_text, pd.DataFrame(display_rows), invalid_title, invalid_table, csv_path


def load_example(example_name: str):
    example = EXAMPLES[example_name]
    return (
        example["title"],
        example["description"],
        example["organism"],
        example["readout"],
        example["assay_format"],
        example["assay_type"],
        example["target_uniprot"],
        example["smiles"],
    )


with gr.Blocks(title="BioAssayAlign Compatibility Explorer", analytics_enabled=False, css=CSS, theme=THEME) as demo:
    gr.HTML(
        """
<div id="hero">
  <div class="hero-grid">
    <div>
      <div class="eyebrow">BioAssayAlign · assay-conditioned compound ranking</div>
      <div class="hero-title">Rank a compound list against one assay definition</div>
      <div class="hero-copy">
        Define one assay, submit a candidate molecule list, and obtain an assay-conditioned ranking for that specific list.
      </div>
    </div>
    <div class="hero-side">
      <div class="hero-side-title">Operational scope</div>
      <ul class="hero-list">
        <li>One assay at a time</li>
        <li>Relative ranking within the submitted candidate set</li>
        <li>Immediate rejection of malformed SMILES</li>
      </ul>
    </div>
  </div>
</div>
"""
    )

    with gr.Row():
        with gr.Column(scale=9):
            gr.HTML(
                f"""
<div class="metric-strip">
  <div class="metric-card"><span>Model</span><strong><a href="https://huggingface.co/{MODEL_REPO_ID}" target="_blank" rel="noopener">{MODEL_REPO_ID}</a></strong></div>
  <div class="metric-card"><span>Input</span><strong>One assay definition and up to {MAX_INPUT_SMILES} candidate SMILES. Scores are only meaningful within the submitted list.</strong></div>
</div>
"""
            )
    gr.HTML("<div class='compact-spec'>Write your own assay directly or start from a live example. Include target, readout, organism, and format when known. Provide candidates as one SMILES per line or as a CSV with a <code>smiles</code> column.</div>")

    with gr.Tab("Rank Compounds"):
        with gr.Row(elem_classes="workspace"):
            with gr.Column(scale=6, elem_classes="pane"):
                gr.HTML(
                    """
<div class="pane-header">
  <div>
    <div class="pane-kicker">Assay definition</div>
    <div class="pane-title">Define the assay context</div>
  </div>
</div>
<div class="pane-copy">Edit the fields directly. The example selector only pre-populates a realistic starting point.</div>
<div class="helper-row">
  <div class="helper-chip"><strong>Protocol signal</strong>State readout, assay system, target biology, and measurement context.</div>
  <div class="helper-chip"><strong>Target identifiers</strong>UniProt IDs usually improve separation between plausible and implausible candidates.</div>
  <div class="helper-chip"><strong>Ranking scope</strong>Scores are only comparable within the candidate list submitted here.</div>
</div>
"""
                )
                example_name = gr.Dropdown(choices=list(EXAMPLES.keys()), value=DEFAULT_EXAMPLE_NAME, label="Example assay")
                gr.HTML("<div class='examples-note'>Choose a live example or overwrite the fields directly with your own assay and candidate list.</div>")
                assay_title = gr.Textbox(label="Assay title", value=DEFAULT_EXAMPLE["title"])
                description = gr.Textbox(label="Description", value=DEFAULT_EXAMPLE["description"], lines=6, placeholder="Describe the assay in practical lab language.")
                with gr.Row():
                    organism = gr.Textbox(label="Organism", value=DEFAULT_EXAMPLE["organism"], placeholder="Homo sapiens")
                    readout = gr.Textbox(label="Readout", value=DEFAULT_EXAMPLE["readout"], placeholder="binding / fluorescence / luminescence")
                with gr.Row():
                    assay_format = gr.Textbox(label="Assay format", value=DEFAULT_EXAMPLE["assay_format"], placeholder="biochemical / cell-based")
                    assay_type = gr.Textbox(label="Assay type", value=DEFAULT_EXAMPLE["assay_type"], placeholder="binding / inhibition / activation")
                target_uniprot = gr.Textbox(label="Target UniProt IDs", value=DEFAULT_EXAMPLE["target_uniprot"], placeholder="Q06187, P52333")

            with gr.Column(scale=5, elem_classes="pane"):
                gr.HTML(
                    """
<div class="pane-header">
  <div>
    <div class="pane-kicker">Candidate set</div>
    <div class="pane-title">Submit the candidate molecules</div>
  </div>
</div>
<div class="pane-copy">The model ranks molecules relative to the exact candidate set submitted in this run.</div>
"""
                )
                smiles_text = gr.Textbox(
                    label="Candidate SMILES",
                    value=DEFAULT_EXAMPLE["smiles"],
                    lines=14,
                    placeholder="Paste one candidate molecule per line. Example: CCO",
                )
                upload_file = gr.File(label="Upload CSV / TXT / SMI", file_count="single", file_types=[".csv", ".txt", ".smi", ".smiles"])
                top_k = gr.Slider(label="Top-K rows to display", minimum=5, maximum=200, step=5, value=DEFAULT_TOP_K)
                gr.HTML("<div class='section-note'>Use one SMILES per row. Invalid or non-standardizable structures are flagged separately and excluded from ranking.</div>")
                with gr.Row(elem_classes="action-row"):
                    run_btn = gr.Button("Run assay-conditioned ranking", variant="primary")
                    clear_btn = gr.ClearButton(value="Clear inputs", components=[assay_title, description, organism, readout, assay_format, assay_type, target_uniprot, smiles_text, upload_file])

        summary = gr.HTML(elem_classes="results-shell")
        with gr.Accordion("Serialized assay text used by the model", open=False, elem_classes="result-accordion"):
            assay_preview = gr.Textbox(lines=12, label="Model-facing assay text")
        gr.HTML("<div class='section-note'><strong>Ranked candidates</strong></div>")
        ranked_df = gr.Dataframe(label=None, show_label=False, interactive=False, wrap=True, elem_classes="result-frame")
        invalid_title = gr.HTML("<div class='section-note'><strong>Rejected inputs</strong></div>", visible=False)
        invalid_df = gr.Dataframe(label=None, show_label=False, interactive=False, wrap=True, elem_classes="result-frame", visible=False)
        download_file = gr.File(label="Export CSV", elem_classes="result-file")

        example_name.change(
            load_example,
            inputs=[example_name],
            outputs=[assay_title, description, organism, readout, assay_format, assay_type, target_uniprot, smiles_text],
        )
        run_btn.click(
            run_ranking,
            inputs=[assay_title, description, organism, readout, assay_format, assay_type, target_uniprot, smiles_text, upload_file, top_k],
            outputs=[summary, assay_preview, ranked_df, invalid_title, invalid_df, download_file],
        )

if __name__ == "__main__":
    if ENABLE_BACKGROUND_WARMUP:
        import threading
        threading.Thread(target=_warm_model_background, daemon=True).start()
    demo.queue(default_concurrency_limit=4).launch(
        show_error=True,
        quiet=True,
        footer_links=["gradio"],
    )