Skip to content
papercrane
Written by Vedant K. and Harsha P.

Senba is a 34.2 million parameter protein dynamics model we built from Valence Labs' MarS-FM checkpoint. Give it one structure and it generates possible conformations 50 frames later.

MarS-FM had already learned molecular motion from simulation. We gave Senba a narrower job: model the transition from one frame to the frame exactly 50 steps later. Training pairs came from mdCATH trajectories at 450 K. We used 128 protein domains to train and 32 to choose the checkpoint.

How Senba was trained

What we did

We fine-tuned all 34,152,521 parameters for ten epochs at batch size 8 and a learning rate of 3 × 10⁻⁶. That came to 160 optimizer updates over 30,720 sampled transitions.

The direct fine-tune learned the transition but moved too far from MarS-FM's ensemble behavior. We used source-weight interpolation to retain 7.35 percent of its weight displacement.

What it took

The selected training run took 121.8 seconds on one H100 and peaked at 60.2 GiB of CUDA memory. The two repetitions took 126.0 and 130.4 seconds.

We also tried a 160-epoch branch, which learned a bigger update. Most of it had to be removed to preserve the ensemble, and the anchored checkpoint did not beat the ten-epoch version.

The whole search and benchmark campaign cost $19.45 on Modal. That includes the epoch sweep, rejected and interrupted runs, weight blending, the 28-protein comparison, equilibrium checks, and two confirmation seeds.

Results

Senba vs. MarS-FM

Senba and the source checkpoint used the same 28 mdCATH validation proteins, transitions, and sampling protocol. Endpoint RMSD tells us how far each generated endpoint lands from the target structure; lower is better. Senba improved on 27 of the 28 proteins. The mean improvement was 0.0326 Å, with a 95% bootstrap interval from 0.0253 to 0.0401 Å.

Endpoint RMSD improvement by protein.
Senba improvement over the source model for 28 proteinsTwenty-seven proteins improved and one had a small negative result. The mean improvement was 0.0326 angstrom.mean 0.0326 Å1jeyB031e7uA044qncB004pkiG023pt1A021nktA023i01A035k2mA011h4yA003atsA012izwC002k54A003aekA031q16B032ji7A032f2cA022incC002k4qA001d1rA004b96A003mg3A012kjrA013tdgA012dp9A011rquA014nreA013d8lA004deqA02−0.0100.020.040.060.08

Coordinate persistence assumes the structure does not move. Senba beat this baseline on 22 proteins, by an average 0.4364 Å. It also beat the first five-epoch Senba checkpoint by 0.0022 Å across all 28 proteins and won 22 comparisons.

Did it repeat?

On a compact eight-protein panel, three independently trained checkpoints improved on MarS-FM by 0.0301, 0.0229, and 0.0290 Å. Each checkpoint improved all eight proteins, and each confidence interval stayed above zero.

Ten-epoch training across three seeds, with 95% confidence intervals.
Improvement over source across three random seedsAll three seed means are positive and all three 95 percent confidence intervals exclude zero.seed 01+0.0301 Åseed 02+0.0229 Åseed 03+0.0290 Å0.000.020.04

Did the ensemble hold up?

We also checked the generated ensemble: cross-correlation, pairwise distances, principal-component overlap, radius of gyration, per-residue flexibility, steric clashes, and native contacts.

Mean change relative to each allowed degradation gate. Negative values favor Senba; 1.0 is the limit.
Mean degradation for seven equilibrium metrics relative to their gatesEvery metric is below its mean rejection boundary. Five metrics improved. PCA overlap and native contacts changed slightly but stayed inside their limits.Dynamic cross-correlation−13.2%Pairwise ensemble distance−17.5%Principal-component overlap+13.9%Radius of gyration−28.1%Per-residue flexibility−14.8%Predicted steric clashes−0.2%Native-contact distribution+1.8%-25%0%50%limit

Five of the seven aggregate means improved relative to the source. PCA overlap and native contacts moved slightly the other way, but both stayed well inside their limits. Every protein passed the individual limits on all seven measurements.

Senba is still a validation-stage model. Next, we'll test it across more lags and compare it with other runnable dynamics models.

References

  1. Kapuśniak, K., Gabellini, C., Bronstein, M., Tossou, P., and Di Giovanni, F. MarS-FM: Generative Modeling of Molecular Dynamics via Markov State Models. ICLR 2026. Paper, code, and weights.
  2. Mirarchi, A., Giorgino, T., and De Fabritiis, G. mdCATH: A Large-Scale MD Dataset for Data-Driven Computational Biophysics. Scientific Data 11, 1299 (2024). Paper and dataset.