Keyulu Xu

We’re very excited to share BoltzGen — a biomolecular binder design model that matches state-of-the-art folding and turns that capability into generalizable binder design! Unlike prior tools, BoltzGen spans nanobodies, miniproteins, peptides, and cyclic scaffolds and designs binders for proteins, nucleic acids, small molecules, and more. A declarative specification language (binding sites, covalents, structure groups, secondary structure, design masks) lets you steer generation toward precise objectives. With multiple wet-lab partners, we validated across diverse targets. On 9 novel protein targets (no similar bound complexes in PDB), testing ≤15 designs/target yielded nanomolar binders on 6/9 for nanobodies (67%), mirrored by protein designs. We also show functional peptide designs (including disordered regions) and support for challenging modalities like disulfide-stabilized peptides and small-molecule binders. The entire stack is open-source (MIT): weights, training & inference, and a full design pipeline. 🚀 Model & code: lnkd.in/e5zKCDsy 🤗 Slack community: lnkd.in/ecmfsNJ3 🧠 Blog: boltz.bio/boltzgen 📄 Manuscript: lnkd.in/eXteVX4c Join us for live presentations, demos, and discussions: • MIT (Cambridge) — Thu Oct 30: luma.com/7474iho2• London — Thu Nov 6: luma.com/l2zgvfwt It’s been a huge pleasure working with the Boltz team, and special thanks to @HannesStaerk for leading the project! Original post from Hannes: x.com/HannesStaerk/s…

@PetarV_93 Great explanation @PetarV_93 That paper is interesting but IMO the situation considered is a bit constrained. In practice, good features are all you need for message passing GNNs to solve DP etc

This is a very cool paper! However, if I understood it correctly, it doesn't invalidate the GNN-DP alignment result of @KeyuluXu et al. [33]. Rather, it shows a very interesting DP unsolvability result over arbitrarily-initialised features. See thread -- happy to discuss. 1/4

@aoaXBT Thank you in advance 🥺

Our paper shows neural networks can learn surprisingly good representations with noisy labels. Pre-training and fine tuning with a good architecture structure is the key #Neurips2021 arxiv.org/abs/2012.12896

Our following 3 papers on graph learning will appear at #ICML2021. Details will follow! Convergence and implicit acceleration of GNNs arxiv.org/abs/2105.04550 GraphNorm for accelerating training arxiv.org/abs/2009.03294 Graph adversarial networks (GAL) arxiv.org/abs/2009.13504

3/3 GNNs are vulnerable to information leakage attacks on social networks etc. We propose GAL with minimax games to protect the sensitive information. arxiv.org/abs/2009.13504 w/ @LiaoPeiyuan @hanzhao_ml @rsalakhu @StefanieJegelka

2/3 We evaluate and understand what normalization methods work the best for GNNs. Our GraphNorm consistently improves the training and generalization of GNNs. arxiv.org/abs/2009.03294 w/ @tianle_cai

1/3 GNNs are implicitly accelerated by skip connections and more depth. We also show gradient descent training linearized GNNs converge to global minima. arxiv.org/abs/2105.04550 w/ Kenji Kawaguchi, @mozhi_zhang @StefanieJegelka

2つの『ポケットモンスター』シリーズ最新作、『ポケットモンスター ブリリアントダイヤモンド・シャイニングパール』と『Pokémon LEGENDS アルセウス』の発売が決定！ シンオウ地方を舞台に、異なる冒険が始まるよ！ pokemon.co.jp/ex/bdsp_legend… #ポケモンBDSP #PokemonLEGENDS

Today's notes on "How Neural Nets Extrapolate: From Feedforward to GNNs" (arxiv.org/abs/2009.11848) by @KeyuluXu, @StefanieJegelka Ken-ichi Kawarabayashi @SimonShaoleiDu @mozhi_zhang @jingling_li. Read to learn why and when GNNs extrapolate. notion.so/How-Neural-Net…

#CVPR2021 Reviewer 3: “What is label noise? It’s impossible that noise can exist in regression. Strong reject!!!”

Our "how neural networks extrapolate" will appear at #ICLR2021 @iclr_conf as an oral presentation. openreview.net/forum?id=UH-cm… w/ @StefanieJegelka Ken-ichi Kawarabayashi @SimonShaoleiDu @mozhi_zhang @jingling_li

無限列車

欲しいいいい！！！ Nintendo Switchに新カラー「マリオレッド×ブルー」登場　1月25日より予約受付開始 nlab.itmedia.co.jp/nl/articles/21… @itm_nlab

ICLR's h5-index is 203 and ranks 17 among all scientific publication venues (ahead of NeurIPS, ICCV, ICML). 203 ICLR papers of the last 5 years have gathered more than 203 citations. Pretty amazing for a conference that started in 2013. scholar.google.com/citations?view…

【拡散希望】前作の発売から半年、なんとこの度、さらにやばい本を姉妹書として上梓しました！！（詳細は画像をご覧下さい） 本当にスーパー面白いので買って下さい！！（ド直球） 『分析者のためのデータ解釈学入門〜データの本質をとらえる技術』Kindle版もすぐ出ます amazon.co.jp/dp/4802612907/…

大統領選などしょうもない話で盛り上がってる間にmultilingual T5が出てたのだけれども、mBERTからの精度の上がり方を見ると、やはり我々NLPerは仕事をするよりも新しい事前学習モデルが出るのを待ちながらあつ森でもやってた方がいいんじゃないかという思いがある。 arxiv.org/abs/2010.11934

【今日の一曲：11/2】Carla Bruni - Un ange  youtu.be/2pBABLy1Xew  前の前のフランス大統領夫人です

学習したNNが学習分布外データの予測（外挿）をできるか。MLPは分布外では原点からの方向毎に線形になり、学習目標が非線形な場合は外挿できない。GNNは対象問題が動的計画法の形で解け、適切な非線形がアーキテクチャもしくは特徴に組み込まれている場合、外挿に成功する。arxiv.org/abs/2009.11848