Chloe Hsu
@chloehsu0
Followers
486
Following
15
Statuses
16
Joined April 2022
It's been a year since we first started writing this primer, and so much has changed already. Always amazed by how fast the field grows🌱 Also check out the rest of this @NatureBiotech issue focused on protein engineering. Many more primers, reviews, and news & views 👇
Our primer "Generative models for protein structures and sequences" is now live @chloehsu0 @seafann (free version:
1
6
43
RT @ricomnl: if you feel appalled by this and want to help advance the status quo of ML for single-cell data, let's talk
0
4
0
The February issue of @NatureBiotech is a focus on protein engineering. There are so many great news & views, primers, and reviews. 1/
0
0
0
RT @alexechu_: Had fun putting together this review! Lots of good surveys of the literature lately so mostly we used it as a chance to look…
0
10
0
Pre-print: Code: Twitter threads from me and @alexrives:
Excited to share our new ESM-IF1 inverse protein folding model. The result of scaling inverse folding with millions of predicted structures. Paper: Model:
0
0
3
This large scale inverse folding project wouldn't have been possible without @adamlerer. Very grateful for the mentorship during the internship and the opportunity to work together. Also would love to see this gets used for designing new proteins.
It was a pleasure to work with @chloehsu0 during her internship last year! Our preprint is out describing our large scale inverse protein folding model, i.e. predicting sequence from 3D structure, trained on millions of sequences using AlphaFold2 predicted structures.
0
0
4
RT @alexrives: Excited to share our new ESM-IF1 inverse protein folding model. The result of scaling inverse folding with millions of pre…
0
72
0
RT @BrianHie: The evolutionary velocity paper ended on a cliffhanger: protein language models could predict evolution retrospectively, but…
0
53
0
Fun internship collaboration with @adamlerer @alexrives @BrianHie @TomSercu and the @MetaAI protein team @robert_verkuil @jason_liu2 @ebetica. Still feels a bit surreal to do all of this together remotely during the pandemic!
0
0
8
The ESM-IF1 model uses GVP-GNN encoder layers to extract geometric features, followed by a generic autoregressive encoder-decoder Transformer. We found that this simple architecture is sufficient to learn inverse folding at scale. Model weights & code:
1
0
9
Existing inverse folding models are limited by the relatively small number of experimentally determined structures. Larger models especially benefit from these 12M new predicted structures. Grateful that such a scale is possible at all today, and curious to see what comes next.
0
0
5
We next show that ESM-IF1 is an effective zero-shot predictor of mutational effects. Examples: mutational effects on the binding affinity of SARS-CoV-2 RBD to human ACE2, AAV packaging (gene delivery), stability of de novo mini proteins, and more.
0
0
5
Beyond existing benchmarks, we also make the sequence design task more challenging along three dimensions: (1) introducing masking on coordinates; (2) generalization to protein complexes; and (3) conditioning on multiple conformations. Our new training data help with all three!
0
0
6
The best model trained with predicted structures improves native sequence recovery by 9.4 percentage points (51.6% vs 42.2%) over the previous state-of-the-art model. Sequence recovery (accuracy) measures how often sampled sequences match the native sequence at each position.
0
0
8