Repository navigation
docs: post-workshop plan and the three directions for the restart - #84
Merged
Merged
Conversation
The team agreed to continue from mid-October. This writes down where the project actually stands so the restart argues about something concrete. What it records: - M1-M3 run end to end (5.5 min on G37, 418 tests, validating report.json); M4 and M5 were never built, and M4 is the binding evidence the project is named for. Three PRs have been open since 18 September. - The pipeline has never run on a genome whose relatives rather than itself are in the PDB: 47 of 50 selected proteins rest on a same-genus hit, the weights are uncalibrated, and two circularity defects surfaced late. - About 2,500 lines have no BV-BRC coupling at all -- the structure search, AlphaFold lookup, quality gates, contract and renderers. Coupling is concentrated in M1 and three M2 files. So the project is a generic evidence factory plus an antibacterial selection policy, and generalizing means making the policy pluggable rather than rewriting the core. Three directions rather than two: finish what exists (sparse-time work), build M4 (needs sustained attention, and Ping's script is the starting point), and write the generic/domain seam down before refactoring toward it. Breast cancer is proposed as the test case for that seam because it inverts the policy -- human homology stops being a penalty, DepMap replaces flux balance, and Open Targets becomes native rather than awkward. Food-science peptides are flagged as a weaker fit: short peptides defeat the evidence modules specifically, so what transfers is the architecture, not the modules. CEPI is placed in the same bucket: a teammate's services are another instance of "can a third party's evidence plug in without touching the core", and the fit needs establishing since their public work is viral and ours is bacterial. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Carla Mann, Alexander Partin, Ping Hu, Ashita Jawali and Jiahui Chen are continuing; Moises Gualapuro, MD Shakhaowat Hossain, Heng Ma and Archit Vasan are not. There is no longer a co-lead, so sequencing decisions land in one place, and five people working evenings is the constraint the proposals are measured against. Two things change substantively: M4 is the best-staffed gap on the board, which was not visible until the team shrank. Ping's prototype already produces prepared receptors, pocket boxes and Boltz-2 inputs; Alex holds Boltz-2 co-folding and built the structure collection M4 consumes; Ashita holds ligand candidate collection. Three of the five are already adjacent to it, and the work is less "build M4" than "bring Ping's prototype inside the contract" -- it currently has its own ranking and output directories and reads none of report.json. #62 becomes a different decision now its author has left: merging ~7,900 lines means owning them. Its most important finding is filed separately as #85 -- M1's --seq-cap removes 58% of hypothetical proteins before M2 sees them, because m1_priority scores a hypothetical 0 while annotation_gap pays 0.30 for exactly that property -- so the finding survives whichever way the PR goes. Refs #85. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This was referenced Oct 6, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The team agreed to continue from mid-October. This writes down where the project stands so the restart argues about something concrete.
docs/post-workshop-plan.md, linked from the docs index as the starting point for the restart.What it records
Works: M1 → M2 → M3 end to end — 5½ minutes on G37, 542 proteins in, 50 selected, 50 structures, validating
report.json, 418 tests.Does not: M4 and M5 were never built, and M4 is the binding evidence the project is named for. Three PRs open since 18 September (#62, #76, #77).
Not shown: the pipeline has never run on a genome whose relatives rather than itself are in the PDB. 47 of 50 selected proteins rest on a same-genus hit, weights are uncalibrated (#61), and two circularity defects surfaced late (#12's self-match, #68's essentiality reference).
The structural finding
About 2,500 lines have no BV-BRC coupling at all —
foldseek.py,pdb_evidence.py,afdb.py, all ofm3_fold/, the contract, the renderers. Coupling is concentrated in M1 (24 refs incollect.pyalone) and three M2 files.So the project is a generic protein-evidence factory plus an antibacterial target-selection policy. Generalizing means making the policy pluggable, not rewriting the core — a much cheaper proposition than it looked.
Three directions, not two
Breast cancer is proposed as the test for (3), because it inverts the policy:
human_homolog_penaltystops being a penalty when the target is human, DepMap replaces flux-balance essentiality, and Open Targets — awkward for bacteria because it is human-centric — becomes native. The parts weakest for bacteria are strongest there, which is the best evidence the split is genuine.Food-science peptides are flagged as a weaker fit: 5–20 residue peptides defeat the evidence modules specifically (PDB search returns noise against a 1e-5 gate, AlphaFold is unreliable at that length, Foldseek needs a fold). What transfers is the architecture, not the modules.
CEPI is placed in the same bucket — a teammate's services are another instance of "can a third party's evidence plug in without touching the core." Their public work is viral and vaccine-facing while ours is bacterial, so the plan says establish the fit rather than assume it.
Sequencing
A and C are not in competition: the refactor that makes the policy pluggable is also what lets us run a null-policy control on the bacterial case, which is what calibration needs. What should not happen is building B and C at once on an unvalidated core.
Proposal, not decision — the document ends with five open questions for the restart.
🤖 Generated with Claude Code