Senba

Senba is a source-anchored adaptation of MarS-FM (Valence Labs, ICLR 2026) for mdCATH backbone transitions at 450 K and one 50-frame lag. It is published as a validation-stage candidate together with the selection receipt, every comparison, and the preregistered protocols, so the claim below can be audited.

Release page: https://senba.papercrane.bio/

Identity

Field	Value
Run	mars-fm-e010-a07350-exactlag-128x32-20260825-s1-v1
Checkpoint	senba.ckpt · 136,711,098 bytes · SHA-256 f02dcf6a97848c834e3c0be7d848c1068d1d0eb7976d8d6bcfbc98d4114e9c7c
Base model	valencelabs/mars-fm revision 1e957cd4b9f1a054ceb0c2104f40ba6f3cc0d8d0, mars_fm.ckpt SHA-256 f56794cc5d1a8e5263031944f932384dd785fe945553612fe76fe1e9191fa879
Source code	valence-labs/mars-fm commit 4fc17d86cd7e22dda6353f2f1f6bf4b40c9ce946
Architecture	Unchanged MarS-FM; 34,152,521 parameters, all trainable during adaptation
Adaptation data	mdCATH, bounded development subset of 128 train and 32 validation domains (manifest sha256:163ab0043d86b5a2731e1b1a233af27fb069049897d657fd5c39bd54b1343b3f), 450 K replicas
Objective	Observed same-replica frame pairs separated by exactly 50 frames (exact_lag), cluster-balanced starts, no replica crossing
Anchoring	Source-checkpoint weight interpolation, α = 0.0735 (7.35% of the fine-tuning displacement retained)
Frozen	2026-08-25
Test records opened	0

The checkpoint is a PyTorch Lightning checkpoint. Its hyper_parameters contain an argparse.Namespace, so loading under torch>=2.6 requires torch.serialization.add_safe_globals([argparse.Namespace]) (the example script does this).

How it was made

1. Parent fine-tune. All MarS-FM parameters were fine-tuned from the pinned source checkpoint for 10 epochs (160 optimizer updates, 30,720 sampled exact-lag transitions), batch size 8 domains, learning rate 3e-6, fresh Adam, seed 20260825, on one NVIDIA H100 in 121.80 s (peak 64.7 GB allocated). Validation loss on a fixed transition population with replayed flow randomness moved from the replayed source value 1.724479 to 1.693136. Parent run mars-fm-epoch-sweep-e010-exactlag-128x32-20260825-s1-v1, checkpoint SHA-256 1bd5c20b5f3aa8587dce6440af6580f0c2def98d0b868a39969ef854bd729a25.
2. Anchoring. The parent damaged the source model's equilibrium ensembles and was rejected as a deployable candidate. Senba is source + 0.0735 · (parent − source) in weight space.
3. Selection. α was chosen on the frozen epoch-sweep protocol, then the candidate had to pass the full-panel transition comparison and the equilibrium-retention suite without any change to the checkpoint, inference settings, metrics, or population.

Full eligible validation panel (28 domains, ≤256 residues)

Comparison	Ensemble-mean endpoint C-alpha RMSD improvement	95% CI	Domains improved
vs source MarS-FM	+0.0326 Å	[0.0253, 0.0401]	27 / 28
vs coordinate persistence	+0.436 Å	[0.203, 0.681]	22 / 28
vs previous Senba (5-epoch, α = 0.08)	+0.00216 Å	[0.00086, 0.00345]	22 / 28

Secondary summaries on the same outputs: best-of-k RMSD improvement vs source +0.0372 Å, single-sample RMSD improvement vs source +0.0689 Å, displacement-magnitude MAE improvement +0.0292 Å, direction-cosine improvement +0.0014. Per-domain rows are in receipts/full-transition-comparison.json.

Equilibrium retention (8 domains, official tree sampling, 500 samples)

Degradation is measured in each observable's unfavorable direction against source MarS-FM; negative values are improvements. Limits were frozen before any candidate output existed.

Observable	Mean degradation	95% CI	Mean limit	Worst domain / limit	Pass
DCCM MAE	−0.000660	[−0.001625, 0.000285]	0.005	0.001768 / 0.01	✓
Native-contact fraction W1	+0.000271	[−0.001594, 0.001946]	0.015	0.003055 / 0.03	✓
Pairwise RMSD W1 (Å)	−0.0437	[−0.1260, 0.0493]	0.25	0.2076 / 0.5	✓
PCA subspace RMSIP	+0.00416	[−0.00023, 0.01075]	0.03	0.02536 / 0.06	✓
Clash fraction	−0.0000012	[−0.0000246, 0.0000225]	0.0005	0.0000575 / 0.001	✓
Radius of gyration W1 (Å)	−0.0422	[−0.0775, −0.0080]	0.15	0.0319 / 0.3	✓
RMSF RMSE (Å)	−0.0222	[−0.0776, 0.0396]	0.15	0.1471 / 0.3	✓

Replication (three independently trained parents, 8-domain compact panel)

Parent seed	Gain vs source	95% CI	Domains improved	Matched difference vs previous Senba
1 (20260825)	+0.0301 Å	[0.0184, 0.0431]	8 / 8	+0.00308 Å
2 (20260826)	+0.0229 Å	[0.0133, 0.0343]	8 / 8	−0.00110 Å
3 (20260827)	+0.0290 Å	[0.0183, 0.0386]	8 / 8	+0.00557 Å

Three-seed mean gain over source: +0.0273 Å. Mean matched difference vs the previous Senba: +0.00252 Å. The seed-2 difference is negative and its interval crosses zero, so the correct reading is a modest average improvement over the previous checkpoint, not a uniform win at every seed.

Intended use

Research inspection of single-chain, all-heavy-atom protein structures of at most 256 residues: draw independent samples of where MarS-FM-style flow matching places the backbone after the trained lag, and compare against source MarS-FM.

How to run

git clone https://github.com/valence-labs/mars-fm.git && cd mars-fm
git checkout 4fc17d86cd7e22dda6353f2f1f6bf4b40c9ce946
pip install "torch>=2.5" "pytorch-lightning>=2.4" torchdiffeq "numpy<2" \
            dm-tree fair-esm biopython pandas scipy
hf download pikachuandme1/senba senba.ckpt inference_example.py --local-dir senba

PYTHONPATH=. python senba/inference_example.py \
  --pdb 1ubq.pdb --chain A \
  --checkpoint senba/senba.ckpt \
  --expect-sha256 f02dcf6a97848c834e3c0be7d848c1068d1d0eb7976d8d6bcfbc98d4114e9c7c \
  --samples 8 --ode-steps 10 --seed 20260904 \
  --out senba-1ubq.pdb

The script writes a multi-model PDB: MODEL 1 is the input, MODELS 2 to K+1 are K independent draws. It refuses structures with missing heavy atoms, non-standard residues, or several chains rather than silently zero-filling, and warns when the input exceeds the validated 256 residues. A GPU is recommended; CPU works but is slow.

To load the checkpoint yourself, follow inference_example.py: build the same batch layout as MarS-FM (torsions, torsion_mask, trans, rots, seqres, mask), then call MarSModule.inference(batch, num_steps=10).

Files

Path	Contents
senba.ckpt	The selected checkpoint, byte-identical to the receipt hash
inference_example.py	Sampling script against the pinned MarS-FM source
receipts/selection-v2.json	Selection receipt (dynaview-mars-fm-epoch-sweep-selection-v2)
receipts/interpolation-config.json, receipts/interpolation-result.json	Anchoring receipt for the selected checkpoint
receipts/parent/training-config.json, receipts/parent/training-result.json	Seed-1 parent fine-tune receipt with per-epoch validation losses
receipts/full-transition-comparison.json	28-domain transition comparison vs source and persistence
receipts/head-to-head-vs-previous-senba.json	28-domain comparison vs the previous Senba checkpoint
receipts/equilibrium-suite-comparison.json	Seven-observable retention suite with per-domain rows
receipts/replications/seed2/, receipts/replications/seed3/	Seed-2 and seed-3 parent, anchoring, and compact-panel receipts
protocols/	The four preregistered protocols, frozen before the outputs they govern existed
release-manifest.json	SHA-256 of every published file

Compute

The three ten-epoch parents took 378.21 H100-seconds in aggregate (121.80, 126.07 mean). Observed Modal billing for the whole 2026-08-25 search and validation campaign was $19.45 ($10.60 adaptation, $8.85 validation).

License and attribution

Weights and code: MIT (see LICENSE). MarS-FM source model and code are MIT, Copyright (c) 2026 Valence Labs. The adaptation data, mdCATH, is CC BY 4.0; attribute it when you use these weights.

Citation

If you use Senba, cite MarS-FM and mdCATH:

@inproceedings{marsfm2026,
  title     = {MarS-FM: Generative Modeling of Molecular Dynamics via Markov State Models},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  note      = {arXiv:2509.24779}
}

@article{mdcath2024,
  title   = {mdCATH: A Large-Scale MD Dataset for Data-Driven Computational Biophysics},
  journal = {arXiv preprint arXiv:2407.14794},
  year    = {2024}
}

Senba itself: Paper Crane, "Senba: a source-anchored MarS-FM adaptation for mdCATH transitions," model release, 2026, https://huggingface.co/vkali08/senba.