Datasets:

integMET
/

edge-ML-metabolomics-graph

"""Interactive explorer for the species-coloured UMAP of the Node2Vec embeddings.
    uv run streamlit run umap_app.py
Reads the cached UMAP coordinates, the node properties from DuckDB and the
trained embeddings, all through paths.py — the same files the static figure uses.
The loaders and `build_figure` are importable without starting Streamlit, so the
chart can be rendered and inspected outside the app (see verify_app.py).
"""

import numpy as np
import pandas as pd
import plotly.graph_objects as go
import streamlit as st

from paths import DB_PATH, EMB_PATH, UMAP_NPY_PATH

# Categorical slots 1-3 of the dataviz reference palette, validated all-pairs in
# both modes. A scatter shows every pair of colours at once, so three is the cap;
# further categories fold into a muted "Other" or are highlighted one at a time.
THEMES = {
    "light": dict(series=["#2a78d6", "#eb6834", "#1baf7a"], other="#c3c2b7",
                  surface="#fcfcfb", ink="#0b0b0b", muted="#52514e", grid="#e1e0d9"),
    "dark": dict(series=["#3987e5", "#d95926", "#199e70"], other="#55554f",
                 surface="#1a1a19", ink="#ffffff", muted="#c3c2b7", grid="#2c2c2a"),
}
GROUPINGS = ["species", "Study", "color_group", "tissue1", "Analytics"]
TOP3 = "— top 3 —"
HOVER_COLS = ["idx", "node_id", "species", "Study", "tissue1"]
HOVER = ("<b>%{customdata[1]}</b><br>"
         "species: %{customdata[2]}<br>study: %{customdata[3]}<br>"
         "tissue: %{customdata[4]}<extra></extra>")


@st.cache_data(show_spinner="Loading nodes…")
def load_nodes() -> pd.DataFrame:
    import duckdb

    con = duckdb.connect(DB_PATH, read_only=True)
    df = con.execute("""
        SELECT g.idx, g.id AS node_id, n.species, n.Study, n.Analytics,
               n.color_group, n.factor,
               split_part(trim(n.factor1, '()'), '|', 2) AS tissue1,
               split_part(trim(n.factor2, '()'), '|', 2) AS tissue2
        FROM graph_nodes g JOIN nodes_ML n ON n.id = g.id
        ORDER BY g.idx
    """).df()
    xy = np.load(UMAP_NPY_PATH)
    if len(xy) != len(df):
        raise ValueError(f"{len(xy)} coordinates for {len(df)} nodes")
    df["x"], df["y"] = xy[:, 0], xy[:, 1]
    return df


@st.cache_data(show_spinner="Loading degrees…")
def load_degrees() -> pd.Series:
    import duckdb

    con = duckdb.connect(DB_PATH, read_only=True)
    return con.execute("""
        SELECT id, count(*) AS degree FROM (
            SELECT start_id AS id FROM edge_ML_expected_ge5
            UNION ALL SELECT end_id FROM edge_ML_expected_ge5)
        GROUP BY id
    """).df().set_index("id")["degree"]


@st.cache_resource(show_spinner="Loading embeddings…")
def load_embeddings() -> np.ndarray:
    import torch

    z = torch.load(EMB_PATH, weights_only=False)["embedding"]
    z = torch.nn.functional.normalize(z, dim=1)  # cosine via dot product
    return z.numpy()


@st.cache_data(show_spinner=False)
def graph_neighbours(node_id: str) -> set:
    import duckdb

    con = duckdb.connect(DB_PATH, read_only=True)
    rows = con.execute("""
        SELECT end_id AS nb FROM edge_ML_expected_ge5 WHERE start_id = ?
        UNION
        SELECT start_id FROM edge_ML_expected_ge5 WHERE end_id = ?
    """, [node_id, node_id]).fetchall()
    return {r[0] for r in rows}


def build_figure(df, group_by="species", highlight=TOP3, shown=(), size=3.0,
                 opacity=0.7, show_backdrop=True, mode="light"):
    """The scatter, plus the subset of rows it draws in colour."""
    theme = THEMES[mode]
    # Rank within the restriction, so restricting to two categories colours both
    # rather than leaving one grey because it is not in the global top three.
    ranked = df[df[group_by].isin(list(shown))] if len(shown) else df
    counts = ranked[group_by].value_counts()

    if highlight != TOP3:
        groups = np.where(df[group_by] == highlight, highlight, "Other")
        order = [highlight]
        palette = {highlight: theme["series"][0]}
    else:
        top3 = list(counts.index[:3])
        groups = np.where(df[group_by].isin(top3), df[group_by], "Other")
        order = top3
        palette = dict(zip(top3, theme["series"]))
    palette["Other"] = theme["other"]

    df = df.assign(_group=groups)
    in_view = df[group_by].isin(list(shown)) if len(shown) else pd.Series(
        True, index=df.index)
    visible = df[in_view]

    fig = go.Figure()
    if show_backdrop and len(shown):
        out = df[~in_view]
        fig.add_scattergl(
            x=out["x"], y=out["y"], mode="markers", name="not shown",
            marker=dict(size=size * 0.7, color=theme["grid"]), opacity=0.5,
            customdata=out[HOVER_COLS].to_numpy(), hovertemplate=HOVER,
            showlegend=False)

    n_other = int((visible["_group"] == "Other").sum())
    for name in ["Other"] + order:  # Other first so it sits underneath
        part = visible[visible["_group"] == name]
        if part.empty:
            continue
        label = f"Other ({n_other:,})" if name == "Other" else f"{name} ({len(part):,})"
        fig.add_scattergl(
            x=part["x"], y=part["y"], mode="markers", name=label,
            marker=dict(size=size, color=palette[name]), opacity=opacity,
            customdata=part[HOVER_COLS].to_numpy(), hovertemplate=HOVER)

    fig.update_layout(
        height=680, margin=dict(l=8, r=8, t=8, b=8),
        paper_bgcolor=theme["surface"], plot_bgcolor=theme["surface"],
        font=dict(color=theme["muted"], size=12),
        legend=dict(bgcolor=theme["surface"], bordercolor=theme["grid"],
                    borderwidth=1, itemsizing="constant", y=0.01, x=0.99,
                    xanchor="right"),
        dragmode="pan", hovermode="closest",
    )
    # UMAP axes carry no units; distances between clusters are not meaningful
    fig.update_xaxes(visible=False, scaleanchor="y", scaleratio=1)
    fig.update_yaxes(visible=False)
    return fig, visible


def selection_panel(picked, df, group_by):
    if not picked:
        st.subheader("No selection")
        st.write("Click a point, or box-select a region, to inspect nodes here.")
        return

    if len(picked) == 1:
        idx, node_id, species, study, tissue = picked[0]
        st.subheader(node_id)
        degree = int(load_degrees().get(node_id, 0))
        st.write(f"*{species}* · study `{study}` · tissue {tissue} · "
                 f"degree {degree:,}")

        z = load_embeddings()
        sims = z @ z[int(idx)]
        sims[int(idx)] = -2.0
        nb = np.argsort(-sims)[:10]
        gnb = graph_neighbours(node_id)
        near = df.iloc[nb][["node_id", "species", "Study", "tissue1"]].copy()
        near["cosine"] = sims[nb].round(3)
        near["graph edge"] = [n in gnb for n in near["node_id"]]
        near["same species"] = near["species"] == species
        st.write("**10 nearest embeddings**")
        st.dataframe(near.set_index("node_id"), width="stretch")
        st.caption(f"{int(near['same species'].sum())}/10 share the species · "
                   f"{int(near['graph edge'].sum())}/10 are true graph neighbours")
        return

    st.subheader(f"{len(picked):,} nodes selected")
    sel = pd.DataFrame(picked, columns=HOVER_COLS)
    field = group_by if group_by in sel.columns else "species"
    st.dataframe(sel[field].value_counts().rename("nodes").to_frame(),
                 width="stretch", height=300)


def main():
    st.set_page_config(page_title="UMAP explorer", layout="wide")
    df = load_nodes()
    mode = st.get_option("theme.base") or "light"

    with st.sidebar:
        st.subheader("View")
        group_by = st.selectbox("Colour by", GROUPINGS, index=0)
        categories = list(df[group_by].value_counts().index)
        highlight = st.selectbox(
            f"Highlight one {group_by}", [TOP3] + categories,
            help="Any single category can be highlighted; only three hues stay "
                 "colourblind-safe when every pair is on screen at once.")
        shown = st.multiselect(
            f"Restrict to {group_by}", categories, default=[],
            help="Leave empty for all. Excluded nodes stay as a faint backdrop.")
        st.divider()
        st.subheader("Marks")
        size = st.slider("Point size", 1.0, 8.0, 3.0, 0.5)
        opacity = st.slider("Opacity", 0.1, 1.0, 0.7, 0.05)
        backdrop = st.checkbox("Keep excluded nodes as backdrop", value=True)

    fig, visible = build_figure(df, group_by, highlight, shown, size, opacity,
                                backdrop, mode)

    st.title("Node2Vec embeddings by species")
    st.caption(
        f"UMAP of 128-d embeddings for {len(df):,} MetaboLights comparison nodes · "
        "cosine metric · axes are arbitrary · drag to pan, scroll to zoom, "
        "click or box-select points to inspect them")

    event = st.plotly_chart(fig, width="stretch", on_select="rerun",
                            selection_mode=("points", "box", "lasso"), key="umap")

    left, right = st.columns([2, 3], gap="large")
    with left:
        st.subheader(f"{group_by} in view")
        table = (visible[group_by].value_counts().rename("nodes").to_frame()
                 .assign(share=lambda t: (t["nodes"] / len(visible))
                         .map("{:.1%}".format)))
        st.dataframe(table, width="stretch", height=260)
    with right:
        points = event.selection.get("points", []) if event and event.selection else []
        selection_panel([p["customdata"] for p in points], df, group_by)


if __name__ == "__main__":
    main()

	"""Interactive explorer for the species-coloured UMAP of the Node2Vec embeddings.

	uv run streamlit run umap_app.py

	Reads the cached UMAP coordinates, the node properties from DuckDB and the
	trained embeddings, all through paths.py — the same files the static figure uses.

	The loaders and `build_figure` are importable without starting Streamlit, so the
	chart can be rendered and inspected outside the app (see verify_app.py).
	"""

	import numpy as np
	import pandas as pd
	import plotly.graph_objects as go
	import streamlit as st

	from paths import DB_PATH, EMB_PATH, UMAP_NPY_PATH

	# Categorical slots 1-3 of the dataviz reference palette, validated all-pairs in
	# both modes. A scatter shows every pair of colours at once, so three is the cap;
	# further categories fold into a muted "Other" or are highlighted one at a time.
	THEMES = {
	"light": dict(series=["#2a78d6", "#eb6834", "#1baf7a"], other="#c3c2b7",
	surface="#fcfcfb", ink="#0b0b0b", muted="#52514e", grid="#e1e0d9"),
	"dark": dict(series=["#3987e5", "#d95926", "#199e70"], other="#55554f",
	surface="#1a1a19", ink="#ffffff", muted="#c3c2b7", grid="#2c2c2a"),
	}
	GROUPINGS = ["species", "Study", "color_group", "tissue1", "Analytics"]
	TOP3 = "— top 3 —"
	HOVER_COLS = ["idx", "node_id", "species", "Study", "tissue1"]
	HOVER = ("<b>%{customdata[1]}</b><br>"
	"species: %{customdata[2]}<br>study: %{customdata[3]}<br>"
	"tissue: %{customdata[4]}<extra></extra>")


	@st.cache_data(show_spinner="Loading nodes…")
	def load_nodes() -> pd.DataFrame:
	import duckdb

	con = duckdb.connect(DB_PATH, read_only=True)
	df = con.execute("""
	SELECT g.idx, g.id AS node_id, n.species, n.Study, n.Analytics,
	n.color_group, n.factor,
	split_part(trim(n.factor1, '()'), '\|', 2) AS tissue1,
	split_part(trim(n.factor2, '()'), '\|', 2) AS tissue2
	FROM graph_nodes g JOIN nodes_ML n ON n.id = g.id
	ORDER BY g.idx
	""").df()
	xy = np.load(UMAP_NPY_PATH)
	if len(xy) != len(df):
	raise ValueError(f"{len(xy)} coordinates for {len(df)} nodes")
	df["x"], df["y"] = xy[:, 0], xy[:, 1]
	return df


	@st.cache_data(show_spinner="Loading degrees…")
	def load_degrees() -> pd.Series:
	import duckdb

	con = duckdb.connect(DB_PATH, read_only=True)
	return con.execute("""
	SELECT id, count(*) AS degree FROM (
	SELECT start_id AS id FROM edge_ML_expected_ge5
	UNION ALL SELECT end_id FROM edge_ML_expected_ge5)
	GROUP BY id
	""").df().set_index("id")["degree"]


	@st.cache_resource(show_spinner="Loading embeddings…")
	def load_embeddings() -> np.ndarray:
	import torch

	z = torch.load(EMB_PATH, weights_only=False)["embedding"]
	z = torch.nn.functional.normalize(z, dim=1) # cosine via dot product
	return z.numpy()


	@st.cache_data(show_spinner=False)
	def graph_neighbours(node_id: str) -> set:
	import duckdb

	con = duckdb.connect(DB_PATH, read_only=True)
	rows = con.execute("""
	SELECT end_id AS nb FROM edge_ML_expected_ge5 WHERE start_id = ?
	UNION
	SELECT start_id FROM edge_ML_expected_ge5 WHERE end_id = ?
	""", [node_id, node_id]).fetchall()
	return {r[0] for r in rows}


	def build_figure(df, group_by="species", highlight=TOP3, shown=(), size=3.0,
	opacity=0.7, show_backdrop=True, mode="light"):
	"""The scatter, plus the subset of rows it draws in colour."""
	theme = THEMES[mode]
	# Rank within the restriction, so restricting to two categories colours both
	# rather than leaving one grey because it is not in the global top three.
	ranked = df[df[group_by].isin(list(shown))] if len(shown) else df
	counts = ranked[group_by].value_counts()

	if highlight != TOP3:
	groups = np.where(df[group_by] == highlight, highlight, "Other")
	order = [highlight]
	palette = {highlight: theme["series"][0]}
	else:
	top3 = list(counts.index[:3])
	groups = np.where(df[group_by].isin(top3), df[group_by], "Other")
	order = top3
	palette = dict(zip(top3, theme["series"]))
	palette["Other"] = theme["other"]

	df = df.assign(_group=groups)
	in_view = df[group_by].isin(list(shown)) if len(shown) else pd.Series(
	True, index=df.index)
	visible = df[in_view]

	fig = go.Figure()
	if show_backdrop and len(shown):
	out = df[~in_view]
	fig.add_scattergl(
	x=out["x"], y=out["y"], mode="markers", name="not shown",
	marker=dict(size=size * 0.7, color=theme["grid"]), opacity=0.5,
	customdata=out[HOVER_COLS].to_numpy(), hovertemplate=HOVER,
	showlegend=False)

	n_other = int((visible["_group"] == "Other").sum())
	for name in ["Other"] + order: # Other first so it sits underneath
	part = visible[visible["_group"] == name]
	if part.empty:
	continue
	label = f"Other ({n_other:,})" if name == "Other" else f"{name} ({len(part):,})"
	fig.add_scattergl(
	x=part["x"], y=part["y"], mode="markers", name=label,
	marker=dict(size=size, color=palette[name]), opacity=opacity,
	customdata=part[HOVER_COLS].to_numpy(), hovertemplate=HOVER)

	fig.update_layout(
	height=680, margin=dict(l=8, r=8, t=8, b=8),
	paper_bgcolor=theme["surface"], plot_bgcolor=theme["surface"],
	font=dict(color=theme["muted"], size=12),
	legend=dict(bgcolor=theme["surface"], bordercolor=theme["grid"],
	borderwidth=1, itemsizing="constant", y=0.01, x=0.99,
	xanchor="right"),
	dragmode="pan", hovermode="closest",
	)
	# UMAP axes carry no units; distances between clusters are not meaningful
	fig.update_xaxes(visible=False, scaleanchor="y", scaleratio=1)
	fig.update_yaxes(visible=False)
	return fig, visible


	def selection_panel(picked, df, group_by):
	if not picked:
	st.subheader("No selection")
	st.write("Click a point, or box-select a region, to inspect nodes here.")
	return

	if len(picked) == 1:
	idx, node_id, species, study, tissue = picked[0]
	st.subheader(node_id)
	degree = int(load_degrees().get(node_id, 0))
	st.write(f"{species} · study `{study}` · tissue {tissue} · "
	f"degree {degree:,}")

	z = load_embeddings()
	sims = z @ z[int(idx)]
	sims[int(idx)] = -2.0
	nb = np.argsort(-sims)[:10]
	gnb = graph_neighbours(node_id)
	near = df.iloc[nb][["node_id", "species", "Study", "tissue1"]].copy()
	near["cosine"] = sims[nb].round(3)
	near["graph edge"] = [n in gnb for n in near["node_id"]]
	near["same species"] = near["species"] == species
	st.write("10 nearest embeddings")
	st.dataframe(near.set_index("node_id"), width="stretch")
	st.caption(f"{int(near['same species'].sum())}/10 share the species · "
	f"{int(near['graph edge'].sum())}/10 are true graph neighbours")
	return

	st.subheader(f"{len(picked):,} nodes selected")
	sel = pd.DataFrame(picked, columns=HOVER_COLS)
	field = group_by if group_by in sel.columns else "species"
	st.dataframe(sel[field].value_counts().rename("nodes").to_frame(),
	width="stretch", height=300)


	def main():
	st.set_page_config(page_title="UMAP explorer", layout="wide")
	df = load_nodes()
	mode = st.get_option("theme.base") or "light"

	with st.sidebar:
	st.subheader("View")
	group_by = st.selectbox("Colour by", GROUPINGS, index=0)
	categories = list(df[group_by].value_counts().index)
	highlight = st.selectbox(
	f"Highlight one {group_by}", [TOP3] + categories,
	help="Any single category can be highlighted; only three hues stay "
	"colourblind-safe when every pair is on screen at once.")
	shown = st.multiselect(
	f"Restrict to {group_by}", categories, default=[],
	help="Leave empty for all. Excluded nodes stay as a faint backdrop.")
	st.divider()
	st.subheader("Marks")
	size = st.slider("Point size", 1.0, 8.0, 3.0, 0.5)
	opacity = st.slider("Opacity", 0.1, 1.0, 0.7, 0.05)
	backdrop = st.checkbox("Keep excluded nodes as backdrop", value=True)

	fig, visible = build_figure(df, group_by, highlight, shown, size, opacity,
	backdrop, mode)

	st.title("Node2Vec embeddings by species")
	st.caption(
	f"UMAP of 128-d embeddings for {len(df):,} MetaboLights comparison nodes · "
	"cosine metric · axes are arbitrary · drag to pan, scroll to zoom, "
	"click or box-select points to inspect them")

	event = st.plotly_chart(fig, width="stretch", on_select="rerun",
	selection_mode=("points", "box", "lasso"), key="umap")

	left, right = st.columns([2, 3], gap="large")
	with left:
	st.subheader(f"{group_by} in view")
	table = (visible[group_by].value_counts().rename("nodes").to_frame()
	.assign(share=lambda t: (t["nodes"] / len(visible))
	.map("{:.1%}".format)))
	st.dataframe(table, width="stretch", height=260)
	with right:
	points = event.selection.get("points", []) if event and event.selection else []
	selection_panel([p["customdata"] for p in points], df, group_by)


	if __name__ == "__main__":
	main()