jiosephlee/assay-transfer-record-level-v24-3-bbb-martins-l3-intern

BBB assay transfer V24.3

This revised V24.3 rejects buckets with a within-parent / between-parent-mean variance ratio strictly greater than 1.0 in all five independent BBB levels. Ratios equal to 1.0 are allowed, as are undefined estimates. Ratios use retained training, validation and test evidence under the existing estimation policy. The calibration metadata records both the cutoff and each bucket's diagnostics. Existing transform-validity checks still reject invalid log10 inputs. This build is stored in bbb_v24_3_ratio1; the original ungated bbb_v24_3 artifacts are preserved.

The release is rebuilt from frozen BBB V10 evidence and exact source-row UID level mapping. Assay bucket and level boundaries, reviewed identity/log10 transforms, same-bucket identical-query binary conflict exclusions, and record support rules are retained. ID requires 12 training records / 6 parents; OOD requires 9-11 non-test records / 4 parents. Official test scaffold units stay outside training and native validation. Existing per-level molecule promotions are retained; no new molecules or scaffolds are promoted. Every selected query molecule is absent from that level's training universe.

Training retains V24.2 minority-first sampling with no hard binary balancing limit. Binary and continuous use separate 4096-pair molecule budgets per role and level, and degree96 per record role. Both aim for half minority pairs before majority filling. Compatible-endpoint solver repair remains enabled. Continuous sigmoid targets use center0.4 SD and temperature0.1. Binary match/mismatch probabilities are refitted from realized continuous training pairs within each level, so changes to the retained continuous evidence propagate into binary targets.

Existing evaluation bucket-selection rules are rerun over the resulting eligible universe. Validation is exhaustive over eligible parents in L1/L2/L3/L5; L4 retains its bucket and per-bucket query caps. Test retains its existing caps. Finally, deterministic equal bucket rounds retain at most500 queries per level, ID/OOD regime, and binary/continuous type, separately for validation and test. Each selected query keeps its full reference pool: top-50 Morgan training records for ID, or all other eligible records in the held-out OOD bucket. Same-molecule OOD references remain allowed.

The three previously excluded L5 binary eval buckets remain excluded; their surviving records can still train. Surviving inherited test-to-validation query copies remain explicitly recorded. Their validation and test results are not independent. No new test-query copies are introduced.

Build with: python -m assay_transfer.record_level.bbb_v24_1.build all --version v24.3 Prior releases and oral datasets are preserved.

Level: L3

Pair rows: {'train': 342830, 'validation_ranking': 8431, 'test_ranking': 4210}