Varun Shanker
@varunrshanker
Followers
785
Following
377
Media
8
Statuses
33
@StanfordMed Biophysics MD-PhD | Caltech BioE ‘21 @caltechmten
Palo Alto, CA
Joined September 2020
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
FB X MLB OPENING STORE
• 1009216 Tweets
Amorim
• 189246 Tweets
LINGLING THAI TOUR EP2
• 129949 Tweets
Steve Bannon
• 122847 Tweets
Kabir Is God
• 103837 Tweets
Sporting
• 102794 Tweets
#Jin_HBD
• 38044 Tweets
JAEMIN REAL CITY TEASER
• 32501 Tweets
ゼノブレイドクロス
• 31738 Tweets
#YutaJinguji
• 29732 Tweets
Our Dopamine Giselle Day
• 24611 Tweets
#神宮寺勇太誕生祭2024
• 23796 Tweets
ディフィニティブエディション
• 21798 Tweets
ホークス
• 20091 Tweets
#الهلال_الطايي
• 19564 Tweets
少女終末旅行
• 19095 Tweets
Xenoblade
• 18958 Tweets
Tom Lembong
• 16713 Tweets
GALA EN EXA
• 15799 Tweets
Switch 2
• 14542 Tweets
José Dirceu
• 13990 Tweets
横浜優勝
• 12550 Tweets
#松田迅祝22歳_にゃんとも愛でたいBD
• 11559 Tweets
神宮寺くん
• 11387 Tweets
اكرم عفيف
• 11335 Tweets
レ・ダウ
• 10607 Tweets
Pinned Tweet
Excited to share our latest work in
@ScienceMagazine
. We describe how a structure-informed language model can be used for diverse protein engineering tasks, and guide the evolution of therapeutic antibodies for improved potency and resilience against viral escape.
12
69
335
In our latest work, we describe how inverse folding with a multimodal language model enables unsupervised ML-guided protein evolution with the strongest experimental success rates to-date. We evolve antibodies to have ~10-30X better pseudoviral neutralization with improved
5
42
204
Great to see our latest work from the Kim Lab, with some new updates, officially in print!
1
1
25
Excited to share our work out now in
@ScienceTM
describing a new strategy to optimize vaccine sequence design for enhanced immune responses, applied to treat glioblastoma
@TopNeuroDocs
1
3
19
Affinity mature any antibody with just a click! Excited to share our new work led by
@BrianHie
which shows zero-shot protein language models can improve affinity of several Abs without any information about target antigen, structure, or training data!
The evolutionary velocity paper ended on a cliffhanger: protein language models could predict evolution retrospectively, but could they also run evolution forward to prospectively design new proteins? So, I retrained as a protein biochemist to find out...
10
54
317
1
1
17
@CarolynBertozzi
@Stanford_ChEMH
@StanfordMed
@BrianHie
@StanfordEng
Thanks for sharing our work
@CarolynBertozzi
! Here’s a brief tweetorial
In our latest work, we describe how inverse folding with a multimodal language model enables unsupervised ML-guided protein evolution with the strongest experimental success rates to-date. We evolve antibodies to have ~10-30X better pseudoviral neutralization with improved
5
42
204
2
3
15
Protein design is often challenged by the fact that mutations that improve protein function can, in some cases, compromise protein stability and evolvability. By modeling the inverse folding problem—predicting amino acid sequences from backbone structure coordinates—using a
1
0
12
A great piece about the intuition behind our new approach to protein/antibody engineering w/
@BrianHie
!
Therapeutic
#antibodies
, the fastest growing drug class, can treat
#cancer
,
#autoimmune
diseases, and more.
@Stanford
/
#CZBiohubSF
researchers are helping to create those treatments even faster, leveraging the
#AI
technology that brought us
#ChatGPT
.
3
17
56
1
1
13
Seminal work by
@francesarnold
&
@jbloom_lab
showed that functionally beneficial mutations can still decrease protein stability and evolvability. We consider a structure-guided paradigm using inverse folding with ESM-IF1 to focus on mutations that preserve the fold of the
1
1
13
Antibodies are important class of protein therapeutics used across many disease indications. Here, we leveraged this approach to identify affinity and neutralization-enhancing mutations that enabled rapid engineering of former FDA-approved therapeutic bebtelovimab against a viral
1
2
12
We are really excited about these new results and see this approach as an attractive low-N alternative to conventional brute-force experimental techniques. I am especially grateful to
@theobruun
,
@BrianHie
, Peter, and the entire Kim Lab for their invaluable help with this work!
2
1
13
In comparison to other machine-learning-guided directed evolution methods, we find that ours has the strongest experimental success rate to-date, surpassing both sequence-only language models and models with supervision on task-specific training data. (5/n)
1
1
9
I am incredibly grateful to have been able to work alongside
@theobruun
,
@BrianHie
, and Prof. Peter Kim on this project. I would also like to thank the entire Kim Lab for their input, support, and contributions that made this effort possible! (n/n)
0
0
9
Our approach takes minutes to highlight desirable, yet rare, mutations that were previously only identified via exhaustive experimental mutational scanning. We also show that performance is superior to sequence-only language models across diverse proteins. (3/n)
1
1
9
We hope this powerful method will enable the development of new, efficacious therapeutics and, more broadly, contribute to democratizing protein engineering efforts via low-N screening campaigns. (5/n)
1
0
8
We find that this is an effective prior to identify functionally beneficial mutations for directed evolution tasks across protein families and definitions of fitness (3/n).
1
0
7
We find it also generalizes to engineering protein complexes. We evolve two antibodies, including Bebtelovimab (a formerly FDA-approved COVID mAb), to have resilience to escape variants Omicron BQ.1.1 and XBB.1.5, achieving an unprecedented 26-fold improvement in neutralization.
1
2
4
@CyrusMaher
Thanks Cyrus! Yes it is. We include only designs in both rounds that improve neutralization (there's actually more if the fitness definition was affinity, as discussed in the text). We think the IF framework is especially valuable in round 2 since mutations are more likely to be
0
0
1