Long compute time for final energy minimisation

[Tenchberry]

Hi there,

I've recently downloaded and installed RhoFold on our compute cluster. Everything runs smoothly, though when trying to generate a predicted structure for a test RNA sequence (no MSA used, just single structure prediction to test the installation and inference times), the calculation reaches the final energy minimisation step and stalls for hours.

Are there any reasons why this could be happening? This is inference.py running on CPU, for reference.

[jiaye01]

Hi, I encountered a similar issue where the inference pipeline stalls during the final energy minimization step, even without MSA.

After some testing, I suspect this might be due to structural distortion during initial prediction, leading to a conformation that fails or freezes during the Amber relaxation step in OpenMM.

Here's an example sequence that consistently reproduces the issue in website and local:
CCUGUCCCUUUUUUUUUUUUUUUUAGGUCCUGGGCA
The prediction proceeds smoothly up to the minimization stage, then stalls indefinitely. I found that the predicted structure contains base misplacements—some nucleotides are severely shifted or rotated, forming geometries that deviate significantly from realistic RNA structures.

This could suggest a potential internal issue with how the model handles certain sequence motifs, resulting in unrealistic geometries that cause OpenMM to hang.

Let me know if others can reproduce this or have found workarounds.

[jiaye01]

[WangJiuming]

Hi @Tenchberry,

Thanks for raising this issue, and apologies for the delayed response.

We have also observed similar problems in practice. As @jiaye01 mentioned, this is primarily due to distorted atoms that cannot be fully corrected by the relaxation program. While relaxation typically resolves minor geometric inconsistencies, more severe structural conflicts can cause the process to stall or take an unusually long time. These cases show the current limitations of the model in handling certain structures that fall outside of our scope (in @jiaye01's case, most likely due to the artificial poly-U fragment).

We’re actively exploring improvements to address this and will share updates as they become available. Meanwhile, please feel free to let us know if there is any further question or suggestion.