WEBVTT

NOTE week11-clip3-conditioning
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 --> 00:00:05.026
Unconditional sampling gives you a valid
structure, but not the one you need.

2
00:00:05.026 --> 00:00:08.521
A real design project asks
for something specific:

3
00:00:08.521 --> 00:00:13.883
hold this functional motif in
place, or bind this target protein.

4
00:00:14.083 --> 00:00:16.601
Conditioning is the direct answer.

5
00:00:16.601 --> 00:00:21.826
Give the denoiser the motif's coordinates
and hold them fixed at every step.

6
00:00:21.826 --> 00:00:25.213
The motif never moves, and
the scaffold is sampled,

7
00:00:25.213 --> 00:00:29.083
so one motif gives you
many different scaffolds.

8
00:00:29.283 --> 00:00:34.066
Guidance is a different lever,
borrowed from image diffusion.

9
00:00:34.066 --> 00:00:38.261
Mix the conditional and
unconditional predictions:

10
00:00:38.261 --> 00:00:43.547
0.2, plus 3 times 0.5 minus 0.2,

11
00:00:43.547 --> 00:00:49.422
gives 1.1. A larger w follows the
condition harder and cuts diversity.

12
00:00:49.422 --> 00:00:55.417
RFdiffusion does not do this; it
conditions the denoiser directly.

13
00:00:55.617 --> 00:01:01.256
On 25 motif-scaffolding
benchmarks, RFdiffusion solved 23,

14
00:01:01.256 --> 00:01:06.605
against 19 for RFjoint inpainting
and 15 for hallucination,

15
00:01:06.605 --> 00:01:10.316
using one set of settings
for every problem.

16
00:01:10.316 --> 00:01:13.883
Note that this success
is scored in silico.

17
00:01:14.083 --> 00:01:17.793
Designed binders were
tested against five targets,

18
00:01:17.793 --> 00:01:23.707
among them influenza hemagglutinin
and PD-L1. 19% worked experimentally,

19
00:01:23.707 --> 00:01:27.334
from fewer than a hundred
designs per target.

20
00:01:27.334 --> 00:01:30.929
The best flu binder reached 28 nanomolar,

21
00:01:30.929 --> 00:01:36.550
and a cryo-EM map matched
the design to 0.63 angstroms.

22
00:01:36.750 --> 00:01:39.400
Read the in silico numbers with care.

23
00:01:39.400 --> 00:01:43.542
The designs are scored by
AlphaFold family predictors trained

24
00:01:43.542 --> 00:01:48.982
on the same structures, so the designer
and the judge can agree and both be wrong.

25
00:01:48.982 --> 00:01:55.713
The 19% is the number that counts,
because it came from a wet lab.