A scatter plot displaying high-dimensional data projected into 2D space using non-linear dimensionality reduction techniques such as t-SNE or UMAP. Points are colored by cluster or class label, revealing groupings and latent structure in the data. This is a standard visualization in machine learning for exploring embeddings, single-cell RNA-seq data, and NLP document clustering, helping practitioners verify that learned representations capture meaningful distinctions.
""" anyplot.ai
scatter-embedding: t-SNE and UMAP Embedding Visualization
Library: altair 6.2.2 | Python 3.13.14
Quality: 93/100 | Updated: 2026-08-11
"""
import os
import sys
sys.path = [p for p in sys.path if not p.endswith("implementations/python")]
import altair as alt
import numpy as np
import pandas as pd
from PIL import Image
from sklearn.datasets import make_blobs
from sklearn.manifold import TSNE
THEME = os.getenv("ANYPLOT_THEME", "light")
PAGE_BG = "#FAF8F1" if THEME == "light" else "#1A1A17"
ELEVATED_BG = "#FFFDF6" if THEME == "light" else "#242420"
INK = "#1A1A17" if THEME == "light" else "#F0EFE8"
INK_SOFT = "#4A4A44" if THEME == "light" else "#B8B7B0"
# Imprint palette (positions 1-7, canonical order)
IMPRINT = ["#009E73", "#C475FD", "#4467A3", "#BD8233", "#AE3030", "#2ABCCD", "#954477"]
# Single-cell RNA-seq scenario with domain-specific cell type labels
CELL_TYPES = ["T cells", "B cells", "NK cells", "Monocytes", "Dendritic cells", "Macrophages", "Plasma cells"]
CELL_ABBR = {
"T cells": "T",
"B cells": "B",
"NK cells": "NK",
"Monocytes": "Mono",
"Dendritic cells": "DC",
"Macrophages": "Mac",
"Plasma cells": "PC",
}
# Data — varying cluster_std for realistic compactness differences across cell types
np.random.seed(42)
n_clusters = 7
cluster_stds = [1.2, 2.0, 1.5, 1.8, 1.0, 2.2, 1.6]
X_high, y_labels = make_blobs(
n_samples=700, n_features=20, centers=n_clusters, cluster_std=cluster_stds, random_state=42
)
X_2d = TSNE(n_components=2, perplexity=30, random_state=42).fit_transform(X_high)
df = pd.DataFrame({"tsne_1": X_2d[:, 0], "tsne_2": X_2d[:, 1], "cluster": [CELL_TYPES[idx] for idx in y_labels]})
centroids = df.groupby("cluster")[["tsne_1", "tsne_2"]].mean()
centroids = centroids.loc[CELL_TYPES].reset_index()
centroids["abbr"] = [CELL_ABBR[c] for c in centroids["cluster"]]
# Tighten the scale domain to the data extent (+6% pad) instead of Altair's
# default "nice" auto-domain — sparse t-SNE clusters otherwise leave large
# unused margins on the canvas.
x_min, x_max = df["tsne_1"].min(), df["tsne_1"].max()
y_min, y_max = df["tsne_2"].min(), df["tsne_2"].max()
x_pad, y_pad = (x_max - x_min) * 0.06, (y_max - y_min) * 0.06
x_domain = [x_min - x_pad, x_max + x_pad]
y_domain = [y_min - y_pad, y_max + y_pad]
# Interactive selection bound to legend — clicking a cell type highlights its cluster
selection = alt.selection_point(fields=["cluster"], bind="legend")
# Marker size/opacity tuned for 700 overlapping points (high-density heuristic)
scatter = (
alt.Chart(df)
.mark_circle(size=45, strokeWidth=0.5)
.encode(
x=alt.X(
"tsne_1:Q",
scale=alt.Scale(domain=x_domain, nice=False),
axis=alt.Axis(labels=False, ticks=False, domain=False, grid=False, title="t-SNE Dimension 1"),
),
y=alt.Y(
"tsne_2:Q",
scale=alt.Scale(domain=y_domain, nice=False),
axis=alt.Axis(labels=False, ticks=False, domain=False, grid=False, title="t-SNE Dimension 2"),
),
color=alt.Color(
"cluster:N", scale=alt.Scale(domain=CELL_TYPES, range=IMPRINT), legend=alt.Legend(title="Cell Type")
),
opacity=alt.condition(selection, alt.value(0.55), alt.value(0.12)),
stroke=alt.value(PAGE_BG),
tooltip=[
"cluster:N",
alt.Tooltip("tsne_1:Q", title="t-SNE 1", format=".2f"),
alt.Tooltip("tsne_2:Q", title="t-SNE 2", format=".2f"),
],
)
.add_params(selection)
)
# Direct abbreviated-name labels on centroids (no legend cross-referencing needed).
# A background-colored halo stroke behind the text keeps labels legible against
# every cluster hue in both themes, independent of the underlying data color.
centroid_marks = (
alt.Chart(centroids)
.mark_text(fontSize=13, fontWeight="bold", dy=-9, stroke=PAGE_BG, strokeWidth=0.75)
.encode(
x=alt.X("tsne_1:Q", scale=alt.Scale(domain=x_domain, nice=False)),
y=alt.Y("tsne_2:Q", scale=alt.Scale(domain=y_domain, nice=False)),
text="abbr:N",
color=alt.value(INK),
)
)
title_text = "scatter-embedding · python · altair · anyplot.ai"
title_params = alt.TitleParams(
text=title_text,
subtitle="t-SNE (perplexity=30) · 20-dimensional synthetic scRNA-seq · 7 cell types, 700 cells",
fontSize=round(18 * min(1.0, 67 / len(title_text))),
subtitleFontSize=12,
color=INK,
subtitleColor=INK_SOFT,
anchor="start",
)
chart = (
alt.layer(scatter, centroid_marks)
.properties(
width=620,
height=320,
padding={"left": 0, "right": 0, "top": 0, "bottom": 0},
title=title_params,
background=PAGE_BG,
)
.interactive()
.configure_view(continuousWidth=620, continuousHeight=320, fill=PAGE_BG, stroke="transparent")
.configure_axis(titleColor=INK, titleFontSize=12)
.configure_legend(
fillColor=ELEVATED_BG,
strokeColor=INK_SOFT,
labelColor=INK_SOFT,
titleColor=INK,
labelFontSize=10,
titleFontSize=10,
cornerRadius=4,
padding=8,
)
)
# Save
chart.save(f"plot-{THEME}.png", scale_factor=4.0)
# Pad-only to the canonical canvas — vl-convert's title/legend padding makes the
# saved PNG larger than width*scale_factor. Never crop: cropping clips title/axis
# labels and trips the AR-09 edge-clipping auto-reject.
TW, TH = 3200, 1800
_img = Image.open(f"plot-{THEME}.png").convert("RGB")
_w, _h = _img.size
if _w > TW or _h > TH:
raise SystemExit(
f"altair vl-convert produced {_w}x{_h}, exceeds target {TW}x{TH}. "
f"Shrink chart .properties(width=, height=) values and re-render."
)
if _w < TW or _h < TH:
_canvas = Image.new("RGB", (TW, TH), PAGE_BG)
_canvas.paste(_img, ((TW - _w) // 2, (TH - _h) // 2))
_canvas.save(f"plot-{THEME}.png")
chart.save(f"plot-{THEME}.html")
Part of t-SNE and UMAP Embedding Visualization on anyplot.ai.