|
|
|
|
|
""" |
|
|
cluster_split.py |
|
|
|
|
|
Author: natelgrw |
|
|
Last Edited: 11/07/2025 |
|
|
|
|
|
Performs spatial KMeans cluster splitting for the AMAX dataset directly on |
|
|
2D UMAP coordinates. This ensures visual consistency between the UMAP |
|
|
visualization and the actual fold assignments, and creates realistic chemical |
|
|
neighborhoods for evaluation. |
|
|
""" |
|
|
|
|
|
import os |
|
|
import random |
|
|
import numpy as np |
|
|
import pandas as pd |
|
|
from collections import defaultdict |
|
|
from rdkit import Chem |
|
|
from rdkit.Chem import AllChem |
|
|
from sklearn.cluster import KMeans |
|
|
import matplotlib.pyplot as plt |
|
|
import seaborn as sns |
|
|
import umap |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
INPUT_CSV = "../amax_dataset.csv" |
|
|
OUTPUT_DIR = "../cluster_split" |
|
|
N_REGIONS = 5 |
|
|
RANDOM_SEED = 42 |
|
|
|
|
|
random.seed(RANDOM_SEED) |
|
|
np.random.seed(RANDOM_SEED) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def compute_fingerprints(df, smiles_column="compound"): |
|
|
""" |
|
|
Compute Morgan fingerprints for all SMILES. |
|
|
Returns array of fingerprints and list of valid indices. |
|
|
""" |
|
|
fps, valid_idx = [], [] |
|
|
print("Computing Morgan fingerprints (radius=2, nBits=2048)...") |
|
|
for i, smi in enumerate(df[smiles_column]): |
|
|
mol = Chem.MolFromSmiles(smi) |
|
|
if mol is None: |
|
|
continue |
|
|
fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=2048) |
|
|
fps.append(fp) |
|
|
valid_idx.append(i) |
|
|
fps_array = np.array([list(fp) for fp in fps], dtype=np.float32) |
|
|
print(f"Valid compounds: {len(fps_array):,}") |
|
|
return fps_array, valid_idx |
|
|
|
|
|
|
|
|
def compute_umap_embedding(fps_array): |
|
|
""" |
|
|
Compute 2D UMAP embedding of fingerprints using Jaccard metric. |
|
|
Returns 2D coordinates for spatial clustering. |
|
|
""" |
|
|
print("\nComputing 2D UMAP embedding (Jaccard/Tanimoto metric)...") |
|
|
fps_bin = (fps_array > 0).astype(float) |
|
|
reducer = umap.UMAP( |
|
|
n_neighbors=25, |
|
|
min_dist=0.1, |
|
|
metric="jaccard", |
|
|
random_state=RANDOM_SEED, |
|
|
) |
|
|
emb = reducer.fit_transform(fps_bin) |
|
|
print(f" UMAP embedding computed: {emb.shape}") |
|
|
return emb |
|
|
|
|
|
|
|
|
def spatial_cluster_split(df, umap_coords, valid_indices, n_regions): |
|
|
""" |
|
|
Performs spatial KMeans clustering directly on UMAP 2D coordinates. |
|
|
Each KMeans cluster becomes one fold - simple 1:1 mapping. |
|
|
Creates visually consistent and chemically meaningful regions. |
|
|
""" |
|
|
print("=" * 65) |
|
|
print("Performing Spatial KMeans Clustering on 2D UMAP Coordinates") |
|
|
print("=" * 65) |
|
|
|
|
|
print(f"\nRunning KMeans with k={n_regions} on 2D UMAP coordinates...") |
|
|
km = KMeans(n_clusters=n_regions, random_state=RANDOM_SEED, n_init=20) |
|
|
labels = km.fit_predict(umap_coords) |
|
|
|
|
|
print(f"\nCluster centroids in 2D UMAP space:") |
|
|
for i, centroid in enumerate(km.cluster_centers_): |
|
|
print(f" Region {i+1}: ({centroid[0]:.2f}, {centroid[1]:.2f})") |
|
|
|
|
|
folds = defaultdict(list) |
|
|
for local_idx, global_idx in enumerate(valid_indices): |
|
|
region = labels[local_idx] |
|
|
folds[region].append(global_idx) |
|
|
|
|
|
total = len(df) |
|
|
print("\nRegion Summary:") |
|
|
for r in sorted(folds.keys()): |
|
|
n = len(folds[r]) |
|
|
p = 100 * n / total |
|
|
print(f"Region {r+1}: {n:,} compounds ({p:.2f}%)") |
|
|
|
|
|
return folds, labels |
|
|
|
|
|
|
|
|
def save_cluster_assignments(df, folds, cluster_labels, umap_coords, valid_idx, output_dir): |
|
|
""" |
|
|
Saves cluster assignments with UMAP coordinates to CSV. |
|
|
Format: compound,cluster,fold,umap_x,umap_y |
|
|
""" |
|
|
print("\nSaving cluster assignments...") |
|
|
fig_dir = os.path.join(output_dir, "figures") |
|
|
os.makedirs(fig_dir, exist_ok=True) |
|
|
|
|
|
assignments = [] |
|
|
for fold_id, indices in folds.items(): |
|
|
for global_idx in indices: |
|
|
if global_idx in valid_idx: |
|
|
local_idx = valid_idx.index(global_idx) |
|
|
compound_smiles = df.loc[global_idx, 'compound'] |
|
|
cluster = int(cluster_labels[local_idx]) |
|
|
fold = fold_id + 1 |
|
|
umap_x = umap_coords[local_idx, 0] |
|
|
umap_y = umap_coords[local_idx, 1] |
|
|
|
|
|
assignments.append({ |
|
|
'compound': compound_smiles, |
|
|
'cluster': cluster, |
|
|
'fold': fold, |
|
|
'umap_x': umap_x, |
|
|
'umap_y': umap_y |
|
|
}) |
|
|
|
|
|
assignments_df = pd.DataFrame(assignments) |
|
|
output_file = os.path.join(fig_dir, "cluster_assignments.csv") |
|
|
assignments_df.to_csv(output_file, index=False) |
|
|
print(f"Saved figures/cluster_assignments.csv: {len(assignments_df):,} entries") |
|
|
|
|
|
|
|
|
def visualize_lambda_max(df, folds, output_dir): |
|
|
""" |
|
|
Plots λmax distributions across regions. |
|
|
""" |
|
|
if "lambda_max" not in df.columns: |
|
|
return |
|
|
print("\nGenerating λmax distribution plot...") |
|
|
os.makedirs(output_dir, exist_ok=True) |
|
|
|
|
|
fig_dir = os.path.join(output_dir, "figures") |
|
|
os.makedirs(fig_dir, exist_ok=True) |
|
|
|
|
|
plt.figure(figsize=(12, 6)) |
|
|
colors = sns.color_palette("husl", len(folds)) |
|
|
|
|
|
for i, (r, idxs) in enumerate(sorted(folds.items())): |
|
|
region_df = df.loc[idxs] |
|
|
sns.kdeplot( |
|
|
data=region_df, |
|
|
x="lambda_max", |
|
|
label=f"Cluster {r+1} (n={len(region_df):,})", |
|
|
linewidth=2.2, |
|
|
color=colors[i], |
|
|
) |
|
|
|
|
|
sns.kdeplot( |
|
|
data=df, |
|
|
x="lambda_max", |
|
|
label=f"Overall (n={len(df):,})", |
|
|
linewidth=2.0, |
|
|
linestyle="--", |
|
|
color="black", |
|
|
alpha=0.7, |
|
|
) |
|
|
|
|
|
plt.title("Lambda Max Distribution Across Cluster Splits", fontsize=14, fontweight="bold") |
|
|
plt.xlabel("λmax (nm)", fontsize=12, fontweight="bold") |
|
|
plt.ylabel("Density", fontsize=12, fontweight="bold") |
|
|
plt.legend(frameon=True, shadow=True) |
|
|
plt.tight_layout() |
|
|
plt.savefig(os.path.join(fig_dir, "cluster_lmax.png"), dpi=300) |
|
|
plt.close() |
|
|
print("Saved figures/cluster_lmax.png") |
|
|
|
|
|
|
|
|
def visualize_umap(umap_coords, valid_idx, folds, output_dir): |
|
|
""" |
|
|
Generates 2D UMAP of chemical space colored by region. |
|
|
""" |
|
|
print("\nGenerating UMAP visualization...") |
|
|
os.makedirs(output_dir, exist_ok=True) |
|
|
|
|
|
colors = sns.color_palette("husl", len(folds)) |
|
|
plt.figure(figsize=(12, 10)) |
|
|
for i, (r, idxs) in enumerate(sorted(folds.items())): |
|
|
local_idx = [j for j, g in enumerate(valid_idx) if g in idxs] |
|
|
label = f"Cluster {r+1} (n={len(local_idx):,})" |
|
|
plt.scatter( |
|
|
umap_coords[local_idx, 0], |
|
|
umap_coords[local_idx, 1], |
|
|
s=10, |
|
|
alpha=0.6, |
|
|
label=label, |
|
|
color=colors[i], |
|
|
) |
|
|
|
|
|
plt.title( |
|
|
"UMAP Projection of Compound Space (Colored by Cluster Split)", |
|
|
fontsize=14, |
|
|
fontweight="bold", |
|
|
) |
|
|
|
|
|
|
|
|
plt.legend(markerscale=2, frameon=True, loc='best') |
|
|
plt.tight_layout() |
|
|
os.makedirs(os.path.join(output_dir, "figures"), exist_ok=True) |
|
|
plt.savefig(os.path.join(output_dir, "figures/cluster_umap.png"), dpi=300) |
|
|
plt.close() |
|
|
print("Saved figures/cluster_umap.png") |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def main(): |
|
|
""" |
|
|
Main function to perform spatial cluster splitting. |
|
|
""" |
|
|
print("=" * 65) |
|
|
print("SPATIAL CLUSTER SPLITTING PIPELINE") |
|
|
print("=" * 65) |
|
|
print(f"Configuration:") |
|
|
print(f"- Number of regions: {N_REGIONS}") |
|
|
print(f"- Clustering dimension: 2D (UMAP coordinates)") |
|
|
print(f"- Random seed: {RANDOM_SEED}") |
|
|
print(f"- Input: {INPUT_CSV}") |
|
|
print(f"- Output: {OUTPUT_DIR}") |
|
|
print() |
|
|
|
|
|
print("Loading dataset...") |
|
|
df = pd.read_csv(INPUT_CSV) |
|
|
print(f"Loaded {len(df):,} rows.") |
|
|
|
|
|
fps_array, valid_idx = compute_fingerprints(df) |
|
|
|
|
|
umap_coords = compute_umap_embedding(fps_array) |
|
|
|
|
|
folds, cluster_labels = spatial_cluster_split( |
|
|
df, umap_coords, valid_idx, N_REGIONS |
|
|
) |
|
|
|
|
|
print("\nSaving cluster CSV files...") |
|
|
os.makedirs(OUTPUT_DIR, exist_ok=True) |
|
|
for r, idxs in folds.items(): |
|
|
output_path = os.path.join(OUTPUT_DIR, f"cluster_{r+1}.csv") |
|
|
df.loc[idxs].to_csv(output_path, index=False) |
|
|
print(f" Saved cluster_{r+1}.csv") |
|
|
|
|
|
save_cluster_assignments(df, folds, cluster_labels, umap_coords, valid_idx, OUTPUT_DIR) |
|
|
|
|
|
visualize_lambda_max(df, folds, OUTPUT_DIR) |
|
|
visualize_umap(umap_coords, valid_idx, folds, OUTPUT_DIR) |
|
|
|
|
|
print("\n" + "=" * 65) |
|
|
print("CLUSTERING COMPLETE!") |
|
|
print("=" * 65) |
|
|
print(f"Output directory: {OUTPUT_DIR}") |
|
|
print(f"- {len(folds)} region CSV files") |
|
|
print(f"- figures/cluster_assignments.csv") |
|
|
print(f"- figures/cluster_lmax.png") |
|
|
print(f"- figures/cluster_umap.png") |
|
|
print() |
|
|
print("Note: Clustering performed in 2D UMAP space for visual consistency") |
|
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
|
main() |