[padix-key]

PartnerIndexEncoder._find_residue_partners() built the O(n²) pairwise distance matrix with numpy.ma operations (ma.dot, masked arithmetic, nan_to_num on a masked array). Those dominated to_3di().

Now the matrix is computed on plain arrays: masked residues are set to NaN (which becomes inf, like the masked entries that argmin() filled with inf before), and nan_to_num(nan=inf) is replaced by an equivalent clip() + NaN assignment. The partner search itself becomes ~3.8x faster (0.54 -> 0.14 ms, 213 residues).

to_3di(), model 1, amino acids only:

Input	Longest chain	Before	After	Speedup
4ZXB (5 chains)	800 res	13.0 ms	5.8 ms	2.2x
5H73	363 res	2.13 ms	0.96 ms	2.2x
5UGO	326 res	1.82 ms	0.89 ms	2.1x
1IGY (4 chains)	434 res	7.67 ms	3.72 ms	2.1x
1DIX	208 res	1.03 ms	0.58 ms	1.8x
1AKI	129 res	0.68 ms	0.45 ms	1.5x

The speedup grows with chain length (the quadratic part); for short chains the remaining per-chain overhead dominates.

Output checked identical to main (codes and PartnerIndexEncoder.encode() results) for all structures in tests/structure/data/pdb, including variants with randomly removed atoms to produce masked residues (also at the first position).

Tests: tests/structure/alphabet

Authored by Claude

[codspeed]

Merging this PR will improve performance by 91.22%

1 warning

⚠️ 17 benchmarks with a significant change ran on different hardware

Their base and head results were measured on machines with a different CPU, core count or memory, so the changes might not come from the code. Run base and head on the same runner type.

Learn more about Macro Runners

See affected benchmarks

⚡ 17 improved benchmarks
✅ 92 untouched benchmarks
🆕 2 new benchmarks
⏩ 14 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	benchmark_find_rotatable_bonds	177.1 ms	26.8 ms	×6.6
⚡	benchmark_superimpose[superimpose]	3,198.9 µs	643.9 µs	×5
⚡	benchmark_hbond	29.7 ms	10.1 ms	×2.9
⚡	benchmark_similar_kmers	483 µs	197.4 µs	×2.4
⚡	benchmark_superimpose_multi_model[superimpose]	14.4 ms	6.4 ms	×2.3
⚡	benchmark_get_assembly	1,594.7 ms	835.6 ms	+90.85%
⚡	benchmark_superimpose[superimpose_homologs]	60.6 ms	33.5 ms	+80.57%
⚡	benchmark_score_threshold_rule[11*11*1*1***111]	3.3 s	1.9 s	+70.9%
⚡	benchmark_score_threshold_rule[None]	3.3 s	1.9 s	+70.78%
⚡	benchmark_superimpose_multi_model[superimpose_homologs]	74.2 ms	44.8 ms	+65.7%
⚡	benchmark_superimpose[superimpose_without_outliers]	31.9 ms	20.8 ms	+52.91%
⚡	benchmark_structural_alphabet_methods[to_3di]	21.1 ms	14.1 ms	+49.96%
⚡	benchmark_superimpose[superimpose_structural_homologs]	53.4 ms	35.8 ms	+49.32%
⚡	benchmark_match_kmer_selection[KmerTable-None]	642.6 µs	527.6 µs	+21.8%
⚡	benchmark_match[KmerTable-None]	739.6 µs	621.8 µs	+18.94%
⚡	benchmark_set_structure[bcif-True]	39 ms	32.8 ms	+18.79%
⚡	benchmark_set_structure[bcif-False]	17.7 ms	15.1 ms	+16.84%
🆕	benchmark_stack[10]	N/A	6.7 ms	N/A
🆕	benchmark_stack[2]	N/A	2 ms	N/A

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing padix-key:perf/ana-3di-partners (588a1b9) with main (81df912)²}

Footnotes

1. 14 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on main (b0d32ca) during the generation of this report, so 81df912 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩