fix: bug in num_filters_passed counting

[chAwater]

fix bug in num_filters_passed

• fix suspected bug in "num_filters_passed" #126

  boltzgen/src/boltzgen/task/filter/filter.py

  Lines 408 to 409 in f3392fe

  	self.df["num_filters_passed"] += self.df[filter_cols].all(axis=1)
  	self.df["pass_filters"] = self.df[filter_cols].all(axis=1)

• should be

-            self.df["num_filters_passed"] += self.df[filter_cols].all(axis=1) # cumulative all-pass check
+            self.df["num_filters_passed"] += self.df[filter_col].astype(int)  # only cumulative this filter
             self.df["pass_filters"] = self.df[filter_cols].all(axis=1) 

• verify the bug from data (in the final_ranked_designs dir)

import pandas as pd
df = pd.read_csv("./all_designs_metrics.csv")
filter_cols = [c for c in df.columns if c.startswith("pass_") and c.endswith("_filter")]
expected = df[filter_cols].sum(axis=1).astype(int)

# buggy data
print(df["num_filters_passed"].astype(int).value_counts().sort_index())
# expected data
print(expected.value_counts().sort_index())

mismatch = (df["num_filters_passed"].astype(int) != expected).sum()
print(f"Mismatched rows: {mismatch} / {len(df)}")

• Luckily, the budget is typically small so that final_designs_metrics_{n}.csv only contains designs passing all filters. The final selection is usually unaffected (in my experience)

---

remove duplicate ligand_iptm

• fix The dictionary eval_keys_confidence in src/boltzgen/data/const.py contains two entries for ligand_iptm. #135

  boltzgen/src/boltzgen/data/const.py

  Lines 3461 to 3463 in 81b5e1d

  	eval_keys_confidence = [
  	"ligand_iptm",
  	"interaction_pae",

  and

  boltzgen/src/boltzgen/data/const.py

  Lines 3477 to 3483 in 81b5e1d

  	"target_to_design_ipsae",
  	"ligand_iptm",
  	"complex_plddt",
  	"complex_iplddt",
  	"complex_pde",
  	"complex_ipde",
  	]

• just remove the duplicate one