WEBVTT

NOTE week11-clip2-reverse
NOTE Narration read by an AI-generated voice.
NOTE The clip draws its own caption along the bottom of the frame. These cues
NOTE sit there too, so that no part of the animation is ever covered.

1
00:00:00.150 --> 00:00:04.929
A single network sees the noisy
structure x t and the step number t,

2
00:00:04.929 --> 00:00:08.501
and predicts the noise that was added.

3
00:00:08.501 --> 00:00:13.072
Predicting the noise and predicting
the clean structure are the same job:

4
00:00:13.072 --> 00:00:16.283
either one gives you the other.

5
00:00:16.483 --> 00:00:21.786
One reverse step moves a little way
toward the predicted clean structure,

6
00:00:21.786 --> 00:00:25.833
then adds a little fresh noise.
Repeat from a thousand down to one.

7
00:00:25.833 --> 00:00:30.550
The fresh noise is what
keeps the samples diverse.

8
00:00:30.750 --> 00:00:32.893
Run that step a thousand times

9
00:00:32.893 --> 00:00:37.466
and the Gaussian cloud tightens
into a backbone that could fold.

10
00:00:37.466 --> 00:00:41.467
The drawing is illustrative,
but the trajectory is real:

11
00:00:41.467 --> 00:00:45.350
structure appears
gradually, not all at once.

12
00:00:45.550 --> 00:00:50.838
Start from different noise and you get a
different structure. This is generation,

13
00:00:50.838 --> 00:00:57.083
not retrieval: none of these has to
appear anywhere in the training set.

14
00:00:57.283 --> 00:01:03.570
RFdiffusion fine-tunes the structure
predictor RoseTTAFold into its denoiser.

15
00:01:03.570 --> 00:01:08.000
It works on residue frames, a
C-alpha position plus an orientation,

16
00:01:08.000 --> 00:01:13.350
predicts the clean structure at every
step and feeds it back as a template,

17
00:01:13.350 --> 00:01:19.274
and runs about 200 steps. One
twist: for motif scaffolding,

18
00:01:19.274 --> 00:01:22.482
adding no fresh noise
on the way back worked

19
00:01:22.482 --> 00:01:27.683
better in 17 of the 23 solved problems.

20
00:01:27.883 --> 00:01:31.093
RFdiffusion makes a shape, not a protein.

21
00:01:31.093 --> 00:01:35.676
ProteinMPNN then designs
sequences for that backbone,

22
00:01:35.676 --> 00:01:41.168
typically about eight of them. Those
sequences are what you can actually order,

23
00:01:41.168 --> 00:01:43.750
express, and test.

24
00:01:43.950 --> 00:01:47.898
AlphaFold 3 uses a
diffusion module as well,

25
00:01:47.898 --> 00:01:51.331
but it predicts raw atom
coordinates directly,

26
00:01:51.331 --> 00:01:55.172
with no amino acid frames
and no torsion angles.

27
00:01:55.172 --> 00:02:01.133
Dropping that protein-specific machinery
is what lets one model cover proteins,

28
00:02:01.133 --> 00:02:04.480
nucleic acids, ligands, and ions.