use BulkTanimotoSimilarity instead of pairwise FingerprintSimilarity loops

[seungchan-an]

Summary: Tc is currently computed via pairwise FingerprintSimilarity loops and can be accelerated using BulkTanimotoSimilarity.

While profiling Tc-based nearest-neighbor computations, I noticed that FingerprintSimilarity(fp1, fp2) is used inside nested Python loops.

CLM/src/clm/commands/write_nn_Tc.py

Line 63 in 2cf5e22

if max_tc < (fps := FingerprintSimilarity(target_fps, ref_fp)):

CLM/src/clm/commands/write_structural_prior_CV.py

Line 169 in 2cf5e22

FingerprintSimilarity(row["fp"], target_fp)

CLM/src/clm/commands/create_training_sets.py

Line 136 in 2cf5e22

FingerprintSimilarity(input_fp, target_fp) for input_fp in input_fps

CLM/src/clm/functions.py

Line 408 in 2cf5e22

tcs.append(FingerprintSimilarity(fp1, fp2))

RDKit provides a bulk API (BulkTanimotoSimilarity) that computes the same Tanimoto scores but is significantly faster for this use case. Here is a simple benchmark comparing the following approaches. Using Morgan bit vectors, all methods produced identical outputs, but performance differed substantially:

• pairwise FingerprintSimilarity: ~32 sec
• pairwise TanimotoSimilarity: ~24 sec
• BulkTanimotoSimilarity: ~1.4 sec