1 of 16

FrameDiff: “SE(3) diffusion model with application to protein backbone generation” Yim et al. (2023)

​

Molecular ML Reading Group

Meeting 2: 09/27/2023

2 of 16

Meeting Outline

  • Updates for Molecular ML Reading Group (MMLRG)
    • Logistics, schedule, and future topics

​

  • “SE(3) diffusion model with application to protein backbone generation” Yim et al. (2023)
    • Background and context
    • Figures and results (discussion!)
    • Conclusions

2

3 of 16

MMLRG Logistics

  • Meet every 2 weeks
    • Next meeting planned for Wed, Oct 11th at 11 am EDT (UTC-4)

​

  • Google form to select next papers for discussion!

​

  • Potential Methods for Discussion:
    • RFDiffusion, DiffDock-PP, Genie, ProtPardelle, EigenFold, DiffPack, EvoDiff, ProteinGenerator

3

4 of 16

What is diffusion? How does it work?

4

Reverse Diffusion*

 

 

 

Sohl-Dickstein et al. (2015); Yang & Ermon (2019); Ho et al. (2020); Song et al. (2021)

Forward Diffusion

 

 

random prior

 

 

 

 

 

 

 

 

 

 

 

5 of 16

 

5

Ho et al. (2020)

 

Compute predicted noise

Take a step to remove the noise

(inference)

Diffusion on Protein Backbones?

 

 

 

Unphysical bond lengths and angles!

6 of 16

 

6

Yim et al. (2023)

 

 

 

Frame

 

 

Rotation

Translation

 

 

  • Reduced degrees of freedom!�
  • Ideal intra-residue bond lengths and angles!�
  • Unphysical peptide bond geometry

7 of 16

SE(3) Diffusion

7

Yim et al. (2023)

 

 

 

 

Rotation

Translation

 

 

Forward Diffusion Process

 

 

 

Rotation

Translation

 

Priors:

 

8 of 16

SE(3) Diffusion for Backbone Generation

8

(Figure 1)

Yim et al. (2023)

9 of 16

FrameDiff Architecture

9

Yim et al. (2023)

(Figure 2)

10 of 16

Sampling Algorithm of FrameDiff

10

Ho et al. (2020); Yim et al. (2023)

(Algorithm 1)

 

Compute predicted noise

Take a step to remove the noise

(inference)

Sample from prior

Predict true data

Estimate score

Translation denoising step

Rotation denoising step

11 of 16

Designability Test for Sampled Backbones

11

Yim et al. (2023)

(Figure 5)

12 of 16

Generation Results of FrameDiff

12

Yim et al. (2023)

(Figure 3)

13 of 16

Generation Results of FrameDiff

13

Yim et al. (2023)

(Figure 6)

14 of 16

Effect of Noise Scale on Generated Backbones

14

Yim et al. (2023)

(Figure 7)

15 of 16

Conclusions

  •  

15

Yim et al. (2023)

16 of 16

Additional Questions?

16

https://forms.gle/LMDuCBP76Uog2KiM8

Topic Selection!