Nate Bennett
@naterbennett0
Followers
1,059
Following
187
Media
8
Statuses
61
Co-Founder at Xaira Therapeutics. De novo protein binder design. Former Postdoctoral Scholar at @UWproteindesign .
Seattle, WA
Joined June 2022
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Oprah
• 389348 Tweets
JIN AT MILAN FASHION WEEK
• 232655 Tweets
FACE OF GUCCI JIN
• 219299 Tweets
FAYEYOKO X SRICHAND LIVE
• 139927 Tweets
XIAO ZHAN REPRESENTS THE HEART
• 76298 Tweets
TRUEDTAC5G X FB
• 66445 Tweets
GULF BA GUCCI IN MFW
• 56032 Tweets
GUCCI PRINCESS HANNI
• 55197 Tweets
大本丸博
• 36038 Tweets
#lovefighters
• 33399 Tweets
真島吾朗
• 31664 Tweets
アニポケ
• 29522 Tweets
ロバート秋山
• 22551 Tweets
Majima
• 19330 Tweets
#獅子堂あかり3D
• 18066 Tweets
貴景勝引退
• 15728 Tweets
転売ヤー
• 13237 Tweets
レイエス
• 11761 Tweets
オースティン
• 11341 Tweets
ビーズリー
• 10219 Tweets
Pinned Tweet
We’re excited to share our preprint where we show, for the first time, the atomically accurate design of VHH antibodies!
16
163
643
1/7 Improving de novo Protein Binder Design with Deep Learning () We show that AF2 is an effective predictor of whether a de novo designed miniprotein will bind to the intended target or not.
5
74
368
It’s exciting to see our work on improving binder design with deep learning methods published in Nature Communications!
4
57
231
RFdiffusion is now free and open source. We’re excited to see what cool proteins you all will diffuse!
Today we're making RF Diffusion, our guided diffusion model for protein design with potential applications in medicine, vaccines & advanced materials, free to use. The software has proven much faster and more capable than prior protein design tools.
6
325
1K
0
11
82
Turns out RFdiffusion is incredible at designing binders to arbitrary targets. We see a 100x increase in experimental success rate over Rosetta-based binder design methods!
We’re very happy to announce that our RFdiffusion manuscript is now on bioRxiv! A lot can change in a week - we’ve now tested over a thousand designs and there’s so much exciting new data! 🧵
28
273
1K
0
2
55
Diffusion models excel at protein design including binder design! Check out Joe’s thread to see how we made a pM binder straight off the computer!
DALL-E’s amazing images are popping up all over the web. That software uses something called a diffusion model, which is trained to remove noise from static until a clear picture is formed.
Turns out diffusion models can design proteins too!
34
529
2K
0
11
51
Here’s a link to the paper on BioRxiv!
1
7
35
@UWproteindesign
Protein binder design has never been easier and we are excited to see what you design! A big thanks to my co-lead on this paper: Brian Coventry (who is not on Twitter).
0
3
17
This work is a proof-of-concept and needs to be refined before it is on-par with RFdiffusion for mini binder design. I'm incredibly excited, however, about the potential of this type of technology for therapeutic design and I can't wait to see where this leads!
2
2
18
2/7 This paper builds on the work of Cao et al. () that presented a pipeline to computationally design miniprotein binders. The major issue with this pipeline has been the low experimental success rate of designed binders.
1
2
17
@UWproteindesign
This method is what we use for sequence design and in silico filtering of RFdiffusion protein binder backbones! RFdiffusion is already open source and we are making this method free and open source as well:
1
4
14
5/7 The confidence AF2 assigns to interchain contacts between target and miniprotein (a metric we call pAE_interaction) is by-far the best predictor of whether a miniprotein will work experimentally that we have ever seen. A new RoseTTAFold (RF2) is similarly predictive to AF2.
1
2
15
RFdiffusion is capable of designing diverse antibodies, that are truly de novo (no similarity to existing antibodies to those epitopes/in the training set).
@RobertRagotte
led the experimental effort, and characterised numerous VHH antibody binders to four different targets.
1
4
14
4/7 Crucially, initializing AF2’s prev_pos variable with the Rosetta design structure helps AF2 find the correct dock more often (we call this protocol “AF2 with an initial guess”).
1
2
14
We wanted to know precisely how accurate these VHH designs were.
@andrewjborst
along with the IPD EM core solved the structure of one of these VHHs binding to Flu Hemagglutinin, and the structure is essentially exactly as designed, even in the challenging-to-model CDR H3 loop.
1
3
13
6/7 We put this new metric to the test in a prospective experiment against 4 new targets. We found that filtering by AF2 pAE_interaction led to ~10-fold experimental success rate improvements over the previous state-of-the-art.
3
1
9
Last year, we released RFdiffusion, which showed state-of-the-art performance across a broad range of protein design tasks. However, the design of antibodies is a particularly challenging problem because of the flexible loop-mediated interactions they make with target proteins.
1
2
9
We had so many amazing collaborators on this project:
@dejsee
,
@wanderingriti
,
@definitelyphil
,
@SingerBenedikt
,
@SusanaVazTor
, as well as many Twitter-less collaborators!
1
3
9
7/7 We also find that using ProteinMPNN to design miniprotein binders is ~6-fold more computationally efficient than the previous state-of-the-art design method (below). We also verify that binders designed with ProteinMPNN work experimentally.
1
2
8
3/7 Using templating, we configure AF2 to fix the structure of the target protein in place and predict the structure and docked pose of the miniprotein binder from single-sequence. This lets us model challenging target proteins (like COVID Spike, below; AF2: Blue, CryoEM: Green).
1
1
8
@andrewwhite01
We have tried using hallucination for designing binders! Hallucination, however, takes ~10 minutes on a GPU to make a design whereas ProteinMPNN takes ~2 seconds on a CPU. They pass the AF2 filter at similar rates so we prefer to the faster method!
2
0
6
Antibodies are one of the biggest therapeutic modalities, however, so we desperately want to be able to rationally design them.
@_JosephWatson
and I therefore developed RFdiffusion further to extend it to the design of antibodies.
1
2
5
Thanks Stephan for the in-depth article on RFdiffusion! This article summarizes our work really well and describes how it fits within the broader context of de novo protein design. I recommend anyone interested in RFdiffusion to read this article!
The past month we have seen some amazing AI news 🤖. But we should be careful not to miss out on what I believe could be one of the most disruptive proteomics technologies this decade. Protein Diffusion could usher in a new Protein Design Era.
4
49
205
0
0
4
Very cool work showing RFdiffusion can generate very high affinity binders to many different peptides-of-interest!
So happy to have co-lead this project ! And so excited for this new era of protein design ! Is finally out !
@PreethamVi
@definitelyphil
@_JosephWatson
@DaveJuergens
1
37
165
0
1
3
Since we published the preprint of this work last year, this pipeline has been used to create a ton of new protein binders at
@UWproteindesign
.
1
2
1
@andrewwhite01
@_JosephWatson
@DaveJuergens
@jueseph
and
@sid_thesci_kid
have done most of the work on hallucinating binders!
0
0
2
@Eyesgack
We are looking into trying to learn physically what AF2 knows! To express these energies in Rosetta, though, they must be pairwise decomposable which makes it difficult
0
0
1
@jrobsontull
We describe in the paper how we use AF2 to validate the binders in silico. Scripts to do that protocol are also publicly available!
0
0
1
@SassSeabass
pAE_interaction is closely correlated with pLDDT of the binder! So filtering stringently on pAE_interaction is already selecting for designs that AF2 thinks will be stable as a monomer. We are very curious what AF2 is still getting wrong though...
0
0
1
@Eyesgack
Thank you! Generating large datasets of successful and unsuccessful designs to train on is probably the most straightforward way to improve the success rate. A larger dataset should also allow us to better understand, physically, what makes binders work
0
0
1