Add interface contact probability scores

[DimaMolod]

Resolves #21.

Summary

• Add contact-probability support to Confidence / Interface and write four new CSV columns:
  • interface_contact_prob_source
  • interface_contact_prob_max
  • interface_contact_prob_top10_mean
  • interface_expected_contacts
• AF3: read native contact_probs from confidences.json, align token matrices to residue matrices where possible, symmetrize, and expose as af3_contact_probs.
• AF2: derive contact probabilities from full result-pickle distogram logits using softmax mass for bins with upper bound < 8 A, matching the AlphaPulldown diagnostics fallback convention, and expose as af2_distogram_lt_8A.
• Keep these scores out of interface_meta_score for now; the benchmark comparison should decide how useful/calibrated they are.
• Add scripts/evaluate_contact_probability_scores.py for labelled benchmark comparison without mutating prediction directories. It can score all chain pairs, including pairs without coordinate-detected contacts, which is useful for interaction benchmarks.

Validation

• python -m compileall -q src/alphajudge scripts/evaluate_contact_probability_scores.py
• pytest -q test/test_parsers_and_runner.py::test_af2_distogram_contact_probs_softmax_cutoff_is_deterministic test/test_parsers_and_runner.py::test_contact_probability_scores_math_is_deterministic test/test_parsers_and_runner.py::test_af3_parser_accepts_alphapulldown_layout test/test_parsers_and_runner.py::test_af3_contact_probability_scores_match_raw_contact_probs -> 4 passed in 21.38s
• Runner smoke on copied AF2/AF3 fixtures with PNG/biophysical extras disabled:
  • AF2 fixture without result pickle: contact source blank, contact scores nan
  • AF3 fixture: af3_contact_probs, max 0.16, top10 mean 0.146, expected contacts 140.22
• I also started a broader five-test run including the existing full schema runner test; the four contact-focused tests passed, and I interrupted the older broad runner path after 3:47 while it was CPU-bound in existing expensive numeric work.

Slurm benchmark jobs

• Smoke job: 58905391, human AF3, one pos/one neg per group, output under /g/transform/kosinski/dima/IntAct_BioGRID_STRING/benchmark_26/contact_probability_benchmark/.
• Full benchmark job: 58905885, dependency afterok:58905391, output CSVs:
  • /g/transform/kosinski/dima/IntAct_BioGRID_STRING/benchmark_26/contact_probability_benchmark/contact_probability_scores.csv
  • /g/transform/kosinski/dima/IntAct_BioGRID_STRING/benchmark_26/contact_probability_benchmark/contact_probability_metrics.csv

[DimaMolod]

Pushed review-response commit 85b129d. Main updates: AF2 distogram scoring now uses the lower-bound <8A convention with logsumexp, compressed AF2 pickles are covered by tests, AF3 contact_probs alignment prefers token_res_ids and ignores extra non-residue tokens, summaries only symmetrize the requested chain-pair block, and the benchmark script now uses this checkout explicitly, complete-case metrics, tie-aware AP, checkpointing, and per-model error handling.

Benchmark status: the first full CSV in contact_probability_benchmark finished, but it imported the dirty local checkout and AF2 contact scores were present for only 12 rows, so I treat it as preliminary AF3-only signal. Corrected rerun is in /g/transform/kosinski/dima/IntAct_BioGRID_STRING/benchmark_26/contact_probability_benchmark_pr22_fix. Smoke job 58939682 completed; full Slurm job 58939710 is running.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6fce50a27d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".