add thermostability

thermostability/README.md

@@ -0,0 +1,113 @@
+# thermostability
+
+Folding free energy of a monomer, predicted from a frozen encoder's residue
+representations. Regression against cDNA-proteolysis dG.
+
+**This directory holds the protocol only. The data was not copied.** The LMDB is large and
+already staged where training reads it, so moving it would duplicate it for no gain. Paths
+are absolute and given below. This is a deliberate exception to the collection's rule 6.
+
+## Task
+
+```
+input:   one protein chain, as 3-view CHI tokens (seq / bb / fa)
+output:  one scalar, dG in kcal/mol
+metric:  Spearman correlation on the held-out set
+```
+
+A **probe, not a fine-tune**. The encoder is frozen; only the head trains.
+
+## Probe head
+
+Replicated from `k-fold-stability/stability_ft/model_esmc.py`, which is the SaProt recipe,
+so numbers are comparable to that line of work.
+
+```
+masked residue mean-pool over valid positions
+Linear(D, D) -> ReLU -> Linear(D, 1)
+clamp(-20, 20)
+SigmoidScaling(min=-1, max=5)
+```
+
+**No input LayerNorm.** Our encoder output already passes `encoder_norm_after`, and adding a
+second normalisation measurably hurt: that is the ours-0714f finding recorded in the
+k-fold-stability head-design note. Anyone porting this head to an encoder that does not
+normalise its output should revisit that, not copy it blindly.
+
+`SigmoidScaling` is the piece that is easy to get wrong. It is not a plain sigmoid. It maps
+the raw scalar onto the cDNA range piecewise:
+
+```
+x < 0     2 / (1 + exp(-alpha2 * x)) - 1        soft tail down to -1
+0 <= x <= 4   x                                 identity through the bulk of the range
+x > 4     2 / (1 + exp(-alpha1 * (x - 4))) + 3  soft tail up to 5
+```
+
+`alpha1` and `alpha2` are trainable, initialised at 0.8. The exponent argument is clamped at
+30 to stop `exp` overflowing to NaN, and the raw MLP output is clamped to +-20 before the
+scaling for the same reason. Both clamps are load-bearing: without them a diverged head
+produces NaN into the loss rather than a large error.
+
+The pooling is masked, `sum(reps * mask) / sum(mask)`, so padding never enters the average.
+
+## Labels
+
+```
+authority:   cDNA proteolysis dG (Tsuboyama et al. mega-scale dataset)
+provenance:  staged into an LMDB plus eval jsonl files by the k-fold-stability project
+validation:  authority is the source; not recomputed here
+```
+
+## Data
+
+Not in this directory. Absolute paths:
+
+```
+/cache/ssahn_lab/datasets/lmdb/stability_monomer                      records (node-local)
+/mnt/parallel_storage/ssahn_lab/datasets/stability/pairs/val_eval.jsonl    2,357 rows
+/mnt/parallel_storage/ssahn_lab/datasets/stability/pairs/test_eval.jsonl   3,282 rows
+/mnt/parallel_storage/ssahn_lab/datasets/stability/pairs/pairs_train.jsonl
+```
+
+`/cache` is node-local. The LMDB exists only on nodes it was staged to, and a run on a
+fresh node needs it copied there first.
+
+## Split
+
+The eval sets are the jsonl files above; the training pairs are separate. **The grouping
+behind the split was not established here.** The collection asks for the grouping key,
+threshold and the number of groups spanning splits, and none of the three is recorded for
+this benchmark. It is inherited from k-fold-stability.
+
+## Baselines
+
+Reported as `stability_eval/{val,test}_spearman` during training runs. This directory does
+not carry a fixed baseline table, because the numbers on record come from mid-training
+checkpoints of various runs rather than from one controlled sweep over encoders. The
+comparison that is established is against SaProt: our 3B dG fine-tune lands at roughly
+SaProt's level with the same head.
+
+**No trivial baseline is recorded, and one matters here.** Sequence length alone correlates
+with dG in the cDNA dataset, since the assay is length-limited. Measure it before reading
+anything into a Spearman value.
+
+## Known limitations
+
+- **Data is not in this directory** and the LMDB half of it is node-local. A reader on a
+  new node cannot run this benchmark without staging.
+- **Split provenance unknown**, as above.
+- **No trivial baseline**, as above, and length is a real confound for this label.
+- **The head is tuned to the cDNA range.** `SigmoidScaling(-1, 5)` bakes in the assay's
+  dynamic range. Applying this head to dG from a different assay would need the bounds
+  changed, and the numbers would not be comparable.
+- **LR sensitivity.** This head produced NaN at lr 4e-4 in an earlier run; the SaProt recipe
+  uses 1e-4. If a run goes NaN, that is the first thing to check, not the clamps.
+
+## Rebuild
+
+```
+k-fold-multimer-tuning/model/stability_head.py     SigmoidScaling and StabilityHead
+k-fold-multimer-tuning/model/callbacks.py          the eval loop that logs Spearman
+```
+
+Head reproduced from `k-fold-stability/stability_ft/model_esmc.py`. Written 2026-08-13.

thermostability/baselines/baselines.json

@@ -0,0 +1,9 @@
+{
+ "task": "cDNA proteolysis dG of a monomer, regression",
+ "metric": "stability_eval/{val,test}_spearman",
+ "protocol": "frozen encoder, masked mean-pool, Linear(D,D)-ReLU-Linear(D,1), clamp +-20, SigmoidScaling(-1,5)",
+ "encoders": null,
+ "encoders_note": "no controlled sweep exists. The numbers on record come from mid-training checkpoints of different runs. The one established comparison is that our 3B dG fine-tune lands at roughly SaProt level with the same head",
+ "trivial_baseline": null,
+ "trivial_baseline_note": "sequence length correlates with dG in the cDNA assay because the assay is length-limited. Measure it before reading anything into a Spearman value"
+}