sheet_fraction and helix_fraction

[zhulinyu0527-eng]

I have a question I’d like to ask,thank you.
When designing binders against a target, among 600 protein backbones, 591 of them have a sheet_fraction of 0, and the remaining 9 protein backbones that do have a non-zero sheet_fraction also show relatively low sheet content. The structures are dominated by helices. Among 300 protein backbones, every single one contains both helices and loops.
I used the parameters inference_sampler.step_scale=3, inference_sampler.gamma_0=0.2, and is_non_loopy: true. Is the fact that 591 out of 600 protein backbones have a sheet_fraction of 0 related to the parameters I set above, or is it related to the specific target and hotspots?

Before setting these parameters, I tested the generation of 10 protein backbones. In all three groups of parameters I tried, is_non_loopy: true was enabled.

• Group 1: inference_sampler.step_scale=3 and inference_sampler.gamma_0=0.2
• Group 2: inference_sampler.step_scale=1.5 and inference_sampler.gamma_0=6 (the model’s default values)
• Group 3: inference_sampler.step_scale=2.25 and inference_sampler.gamma_0=0.4

In all cases the results still showed a high helix content and extremely low sheet content. I would like to know whether this is caused by the way I set the parameters. Thank you.

[rclune]

Thank you for your question.

RFD3 does not have settings that will directly impact which secondary structure motifs are present in the designs it generates. Setting the step_scale to lower values and gamma_0 to higher values should increase the diversity of your designs, but that does not directly translate to designs with higher sheet fractions, it just makes them more likely.

It seems like your adjustments decreased the gamma_0 value compared to the default and increased the step_scale from default. Have you tried increasing the gamma_0 value and decreasing the step_scale value?

You can try changing the noise_scale as well, this will also influence the diversity of your designs.

[zhulinyu0527-eng]

Thank you.
I noticed that inference_sampler.step_scale=3 and inference_sampler.gamma_0=0.2 seem to already represent a low-temperature sampling strategy. Should I further increase the gamma_0 value and decrease the step_scale value based on this setting?What ranges of values are generally recommended for these two parameters?

Do these parameter adjustments affect the designability of the generated structures? How should we balance diversity and designability during parameter optimization? The other two parameter adjustments I tried above were all based on the default values and aimed at improving designability, rather than increasing structural diversity.

Regarding noise_scale — this parameter sets the scaling factor for the noise during inference (default value: 1.003). A smaller value introduces less noise into the system, resulting in lower diversity among the generated outputs. What range of values is generally recommended for this parameter? Thank you.