Paper’s end-to-end pipeline vs. repo code, and missing training

[xiaoxiaokuye]

Hi, thanks for sharing E3-CryoFold!

In the paper “End-to-end cryo-EM complex structure determination with high accuracy and ultra-fast speed” (Nat. Mach. Intell. 2025), the method is described as end-to-end: density map and sequence are directly fed into the model, fused by cross-attention, and the SE(3)-GNN outputs backbone coordinates.

But in this repo, the workflow seems different — it requires extracting Cα atoms, constructing trajectories, and concatenating sequences, and I don’t see the cross-attention interaction described in the paper.

Could you clarify:

Why is the released code more modular instead of the “direct end-to-end” pipeline shown in the paper?

The paper reports large-scale pretraining and fine-tuning, but no training code is included here. Will the training scripts and data preparation pipeline be released?

Thanks again for your work!

[wangjue1111]

Hi, thanks for highlighting this confusion. The model you mentioned in our paper is the E3-CryoFold V1, which was initially released. However, we received some feedback on our models and codes regarding the instability in Cryo-EM map modelling, especially for maps with irregular shapes. Hence, we updated our codes and models for E3-CryoFold V2, which utilize the predicted Ca atoms to guide structure generation, incorporating improved spatial constraints. This update is also mentioned in our paper for future work: “Furthermore, we believe that enhancing the alignment of E3-CryoFold’s predicted structures with the coordinates of Cα atoms from density maps could yield additional improvements. ”

Furthermore, this update not only improves E3-CryoFold's stability and generation but also expands its applications, including sequence-free modeling, low-resolution map modeling, and others, as detailed in the README.

Regarding the problem with the training script, we consider releasing our training script, but there are some issues with the patent and license. We hope we can release the training script soon.

[Josephine-Young]

Like @xiaoxiaokuye pointed out, I also haven't found code for the cross-attention mechanism in the latest version. Will you please explain where and how cross-attention interaction is implemented?