You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
M2: scoring profiles — the triage policy becomes a choice #93
Goal: make the triage score a choice of profile, not a fixed policy. Part of #87.
WEIGHTS in s2f/m2_triage/score.py encodes an antibacterial target-selection policy: human homology is a penalty, essentiality a bonus, membrane proteins deprioritised. That is right for bacteria and wrong for other questions — a viral protein, a single hypothetical, or the breast-cancer test case in #84 (where the target is human).
--profile NAME or --weights FILE; run.json records which profile produced the ranking, so two runs cannot be confused.
With one or a few sequences, skip ranking and report evidence only.
--top (default 50) stays the default selection size.
Definition of done
run.json names the profile and its weights
Changing the profile re-scores recorded proteins.tsv with no re-querying
Any change to the default profile still needs a dated row in the change log in docs/02a-m2-pdb-evidence.md (pitfall M2: triage scoring and selection #12)
Status:#102 implements the profiles (open, ready to merge, last in the #98 → #102 chain).
Done:
--profile bacterial-target (the default, exactly today's WEIGHTS), --profile evidence-only (provisional), and --weights FILE. run.json records the profile and its weights.
python -m s2f.m2_triage.rescore re-ranks recorded proteins.tsv under another profile with no re-querying. Under the default profile it reproduces M2 exactly, and a test pins this.
The default profile is still governed by the change log in docs/02a-m2-pdb-evidence.md.
Left before closing:
Skip ranking for very small inputs. With a handful of sequences (a single query, a small virus, see Decision: viral proteins — scope, and what changes per module #88) a rank is meaningless, so M2 should report evidence without ranking below some size. The threshold is a decision for the team.
Compare profiles on real genomes. The recorded G37 and HS11286 runs cannot be rescored: their proteins.tsv predates surface_bonus, ligandable_homolog and membrane_penalty, and the tool refuses them rather than guessing zeros. A fresh M2 run on each is needed first.
First real comparison of profiles (2026-10-10), from fresh M2 runs rescored with python -m s2f.m2_triage.rescore --profile evidence-only:
Genome
Proteins
Uncharacterized
Top 50 uncharacterized: bacterial-target
evidence-only
Overlap of the two top 50s
G37 (--full-evidence)
542
190
39
7
12 of 50
HS11286 (default flags)
5,338
1,338
33
4
10 of 50
The two profiles select almost disjoint sets. evidence-only favours well-characterised proteins with good PDB hits; bacterial-target favours hypotheticals through annotation_gap (0.30). That is the expected direction, and it shows the choice of profile matters more than any single weight.
These fresh runs also fix the earlier limitation that the recorded runs predated three components.
Activity
Jiahuic commentedon Oct 9, 2026
Status: #102 implements the profiles (open, ready to merge, last in the #98 → #102 chain).
Done:
--profile bacterial-target(the default, exactly today'sWEIGHTS),--profile evidence-only(provisional), and--weights FILE.run.jsonrecords the profile and its weights.python -m s2f.m2_triage.rescorere-ranks recordedproteins.tsvunder another profile with no re-querying. Under the default profile it reproduces M2 exactly, and a test pins this.docs/02a-m2-pdb-evidence.md.Left before closing:
evidence-only. Its weights (pdb_evidence0.40,ligandable_homolog0.10) are a placeholder; see M2: calibrate the triage weights on a genome that can test them #61.proteins.tsvpredatessurface_bonus,ligandable_homologandmembrane_penalty, and the tool refuses them rather than guessing zeros. A fresh M2 run on each is needed first.Jiahuic commentedon Oct 10, 2026
First real comparison of profiles (2026-10-10), from fresh M2 runs rescored with
python -m s2f.m2_triage.rescore --profile evidence-only:bacterial-targetevidence-only--full-evidence)The two profiles select almost disjoint sets.
evidence-onlyfavours well-characterised proteins with good PDB hits;bacterial-targetfavours hypotheticals throughannotation_gap(0.30). That is the expected direction, and it shows the choice of profile matters more than any single weight.These fresh runs also fix the earlier limitation that the recorded runs predated three components.
Merge pull request #102 from NIAID-BRC-Codeathons/feat/93-scoring-pro…