| # Task 3 — self-supervised representation learning (RxRx1, RxRx3-core) |
|
|
| Run traces of the 25 % subset comparison (seed 0, October 2026): adapting a DINOv3 ViT-B/16 to Cell Painting crops |
| by self-supervised fine-tuning on a quarter of each plate's training wells, scored on held-out wells. Code and |
| instructions to rerun: [SHAPE-Revision, branch `repro-sweep`](https://github.com/xuekunwu/SHAPE-Revision/tree/repro-sweep/reproduce/sweep_2026_10) |
| (experiment code: branch `luna-sonnet`, commit `aafdce04`). |
|
|
| | Task | Data | Score (held-out wells) | |
| |---|---|---| |
| | HEPG2, RPE | RxRx1 | well mAP of siRNA retrieval, native embeddings | |
| | HUVEC | RxRx3-core | EFAAR known-relationship recall (mean of CORUM, HuMAP, Reactome, SIGNOR, StringDB) | |
|
|
| Arms (every one has the same tool, lever menu, 8-round budget, 64-epoch cap and label-free validation signal): |
|
|
| - `shape_es` — SHAPE-ES, the agent driving the analyzer's fine-tuning levers; |
| - `react` — ReAct at SHAPE-ES's control frequency (a call resuming a finished call inherits its parameters; the |
| learning rates change in at most one call per run); |
| - `policy_fixed` — a fixed-recipe search over the same levers, no LLM; |
| - `policy_random` — a random search over the same levers, no LLM; |
| - `default` — the plain DINOv3 fine-tune at the tool's defaults, no LLM; |
| - frozen pretrained models, no training and so no trace: DINOv3 (the analyzer's starting point), DINO4Cells-S/B, |
| DCM DINOv2-B, OpenPhenom, Phenom-1/2 and CellProfiler (`frozen_models.tar.zst`). |
|
|
| LLM arms ran with gpt-5.6-luna (`_luna`) and claude-sonnet-5-5 (`_sonnet`), reasoning effort high. |
|
|
| ## Scores (seed 0) |
|
|
| | Arm | HEPG2 | RPE | HUVEC | |
| |---|---|---|---| |
| | SHAPE-ES · luna | 0.0184 | 0.0206 | 0.295 | |
| | SHAPE-ES · Sonnet | 0.0215 | 0.0335 | 0.295 | |
| | ReAct · luna | 0.0116 | 0.0040 | 0.220 | |
| | ReAct · Sonnet | 0.0021 | 0.0151 | 0.260 | |
| | policy-fixed | 0.0137 | 0.0331 | 0.300 | |
| | policy-random | 0.0147 | 0.0380 | 0.287 | |
| | default fine-tune | 0.0086 | 0.0198 | 0.262 | |
| | frozen DINOv3 | 0.0145 | 0.0237 | 0.293 | |
| | frozen DINO4Cells-S (Cell Painting) | 0.0150 | 0.0295 | 0.265 | |
| | frozen DINO4Cells-B (Cell Painting) | 0.0126 | 0.0240 | 0.280 | |
| | frozen DCM DINOv2-B (JUMP-CP) | 0.0072 | 0.0168 | 0.268 | |
| | frozen OpenPhenom (MAE-S) | 0.0058 | 0.0090 | 0.267 | |
| | frozen Phenom-1 (MAE-L) | — | — | 0.318 | |
| | frozen Phenom-2 (MAE-G) | — | — | 0.358 | |
| | CellProfiler features | — | — | 0.286 | |
|
|
| The 25 % subset changes only the training wells. Validation and held-out wells are the full prep's, so the frozen |
| models' scores apply unchanged (frozen DINOv3 scores the same to the last digit on both preps). Phenom-1/2 and |
| CellProfiler exist only as Recursion's released RxRx3-core tables, so they have no RxRx1 score. The summary figure |
| (`summary.tar.zst`) shows the arms of the main comparison only, without policy-random and the frozen models other |
| than DINOv3. |
|
|
| ## Archives |
|
|
| One archive per run, `<arm>_<task>_s0[_<llm>].tar.zst`, unpacking to `<name>/`: |
|
|
| - `run/logs/agent_log.jsonl` — the agent's trace (reasoning, plans, verdicts, every step); |
| - `run/logs/reasoner.jsonl` — every LLM request and its reply; |
| - `run/logs/tool_calls.jsonl` — every tool call: the parameters set, the training record (epochs, validation |
| history, timing) and the readouts; |
| - `run/RUN_META.json`, `run/AGENT_RUN.json` — settings, model, wall time, token use; |
| - `run/submission/decision.json` — the delivered model; |
| - `run/cache/cell_state_analysis/ssl_calls/<call>/` — each call's `CALL_RECORD.json` and `loss_curve.png`; |
| - `run/react/` (ReAct only) — the ReAct session trace; |
| - `lane.log` — the run's console log; `score.json` — its held-out score. |
|
|
| The no-LLM archives (`policy_fixed_*`, `policy_random_*`, `default_*`) also hold the full computation record: |
|
|
| - `run/cache/cell_state_analysis/ssl_calls/<call>/CALL_RESULT.json` (policy arms) — each call's complete tool result: |
| training configuration, per-epoch loss and validation history, early-stop decision, batch and resume details; |
| - `control/loss_curve.png`, `control/archiver.log` (default) — the fine-tune's loss curve and per-epoch checkpoint log; |
| - `scoring/score_native.json`, `scoring/score_matched.json` (HEPG2, RPE) — the held-out scoring record: well mAP on |
| held-out, seen and validation wells with the AP of every well (`per_unit`), chance level and the sha256 of the |
| checkpoint, prep and scorer code; `scoring/score_arm.log` — the scorer's console log; |
| - `scoring/final_eval.json`, `scoring/final_eval.log` (HUVEC, and the default's RxRx1 finals) — the final embedding |
| and EFAAR evaluation (recall per source, compound-gene benchmark per concentration) and its log. |
|
|
| The well → siRNA map is removed from every scoring file (per-well APs are kept, keyed by well). |
|
|
| `frozen_models.tar.zst`: the frozen pretrained models' scores (`scores.json`, with the per-plate matched table and |
| per-source recall), their scorer outputs, and the embedding and EFAAR scoring code (see its `README.md`). |
| `default_<task>_s0.tar.zst` hold the control stage's `CONTROL.json` (training record of the default fine-tune and of |
| frozen DINOv3), its console log and score. `summary.tar.zst`: `results.json` (every score, wall-clock hours, training |
| hours, LLM tokens), the summary figure, frozen DINOv3's scores and `VALIDATION.md` (the reproduction code against the |
| original pipeline: all 16 checked values match). |
|
|
| Not included: model checkpoints and embeddings (several GB per run), and the siRNA labels of the data (no file here |
| holds them). Wall-clock times were measured with up to four trainings sharing one H200 GPU. |
|
|
