FireProtDB2 / src /06_make_weighted_splits.py
drake463's picture
final pipeline and updated subsets
ae74a72
#!/usr/bin/env python3
"""
Re-assigns cluster-based splits to account for mutation data counts per protein (via greedy partition).
Reads in the previous split parquet file, and then generates a new split column
Ouputs:
- A parquet/CSV with new assigned splits.
Usage:
python 06_make_weighted_splits.py \
--input ../data/fireprotdb_cluster_splits.parquet \
--output ../data/fireprotdb_splits_balanced_ddg.parquet \
--ratios 0.8,0.1,0.1 \
--task ddg
Notes:
- Can balance splits based on the amount of different data types. Useful if the splits based on ddG result in large imbalances for other datatypes. Specified with --task [type]
"""
from __future__ import annotations
import argparse
import pandas as pd
SPLITS = ["train", "validation", "test"]
def assign_weighted_splits(cluster_weights: pd.Series, ratios=(0.8, 0.1, 0.1)) -> pd.DataFrame:
total = float(cluster_weights.sum())
targets = {
"train": total * ratios[0],
"validation": total * ratios[1],
"test": total * ratios[2],
}
current = {s: 0.0 for s in SPLITS}
assignment = {}
# Largest-first
for cid, w in cluster_weights.sort_values(ascending=False).items():
w = float(w)
# Choose split that minimizes relative fill after adding this cluster.
# (current+w)/target; lower is better.
def score(s):
t = targets[s] if targets[s] > 0 else 1.0
return (current[s] + w) / t
chosen = min(SPLITS, key=score)
assignment[cid] = chosen
current[chosen] += w
out = pd.DataFrame({"cluster_id": list(assignment.keys()), "split": list(assignment.values())})
# Print expected vs achieved (cluster-weighted, i.e. row-weighted)
print("Target totals:", {k: round(v) for k, v in targets.items()})
print("Achieved totals:", {k: round(v) for k, v in current.items()})
return out
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--input", required=True)
ap.add_argument("--output", required=True)
ap.add_argument("--ratios", default="0.8,0.1,0.1")
ap.add_argument("--task", choices=["dg", "ddg", "tm", "dtm", "fitness", "binary"], default="ddg")
args = ap.parse_args()
ratios = tuple(float(x) for x in args.ratios.split(","))
df = pd.read_parquet(args.input)
if "cluster_id" not in df.columns:
raise ValueError("Input must contain cluster_id")
# IMPORTANT: remove any existing split so we don't accidentally reuse it
if "split" in df.columns:
df = df.drop(columns=["split"])
has_mut = df["mutation"].notna()
if args.task == "dg":
df_task = df[has_mut & df["dg"].notna()].copy()
elif args.task == "ddg":
df_task = df[has_mut & df["ddg"].notna()].copy()
elif args.task == "tm":
df_task = df[has_mut & df["tm"].notna()].copy()
elif args.task == "dtm":
df_task = df[has_mut & df["dtm"].notna()].copy()
elif args.task == "fitness":
df_task = df[has_mut & df["fitness"].notna()].copy()
else:
df_task = df[has_mut & df["stabilizing"].notna()].copy()
# Ensure cluster_id is a plain string key
df_task["cluster_id"] = df_task["cluster_id"].astype("string").fillna("NA_CLUSTER")
# Cluster weights = number of task rows
w = df_task.groupby("cluster_id").size()
print(f"Task={args.task} rows: {len(df_task):,}")
print(f"Task clusters: {len(w):,}")
print("Top 10 clusters by rows:")
print(w.sort_values(ascending=False).head(10))
assign = assign_weighted_splits(w, ratios=ratios)
# Join back to all rows (clusters without task rows -> train by default)
df["cluster_id"] = df["cluster_id"].astype("string").fillna("NA_CLUSTER")
df = df.merge(assign, on="cluster_id", how="left")
df["split"] = df["split"].fillna("train")
df.to_parquet(args.output, index=False)
print("Wrote:", args.output)
# Quick verify on task rows
df_task_out = df_task.merge(assign, on="cluster_id", how="left")
df_task_out["split"] = df_task_out["split"].fillna("train")
print("Task rows by split:")
print(df_task_out["split"].value_counts())
if __name__ == "__main__":
main()