Lars Schaaf
@lars__schaaf
Followers
502
Following
712
Media
23
Statuses
96
PhD student @Cambridge_Uni • ML on Graphs for Molecules & Materials
Joined May 2022
❓Struggling to capture non-local effects using message passing? We present Matrix Function Networks, a new architecture that captures non-local effects in a structured and topology aware manner. https://t.co/9VYEHTNN05 🧵👇(1/10)
arxiv.org
Graph Neural Networks (GNNs), especially message-passing neural networks (MPNNs), have emerged as powerful architectures for learning on graphs in diverse applications. However, MPNNs face...
Excited to present Matrix Function Networks (MFNs), our new non-local GNN architecture that parameterizes many-body non-local effects using the spectrum of learnable graph operators. MFNs' structured non-locality outperforms global attention. https://t.co/JrcuDN6NZ0
1
14
79
We sought to make proteins both potent and FAST. We used #proteindesign to design precise control over protein interaction lifetimes, enabling us to construct rapid-response circuits, biosensors, and switchable cytokines. Now published @Nature! Links to paper and tutorial below.
11
99
459
La-Proteina open source now, excited to see what people build with it!
🧬La-Proteina🧬 The first generative model demonstrating accurate co-design of fully atomistic protein structures (sequence + side-chains + backbone) at scale, up to 800 residues, with state-of-the-art atomistic motif scaffolding performance - has just made its code open-source!
2
23
186
AtomWorks is out! Building upon @biotite_python, we built for a toolkit for all things biomolecules and trained RF3 with it. All open-source, test it via `pip install atomworks`! AtomWorks: https://t.co/cNl31hMzws RF3: https://t.co/CmIXV29FXA Paper: https://t.co/Cc3lB8BCmm 1/6
3
47
167
Check out La-Proteina, our new model for all-atom structure generation at scale! Was a very fun project to work on with @tomasgeffner and the rest of the team. Two things I am particularly excited about in 🧵 1/n
Presenting La-Proteina! A new model for scalable, all-atom protein design 🧬 Backbone + sequence + side-chains, indexed and unindexed atomistic motif scaffolding, scalable up to 800 residues, and more… A thread 🧵
4
23
66
Matbench Discovery is out in Nature Machine Intelligence @ Paper: https://t.co/BtKuHy2fPp Leaderboard: https://t.co/lg3btmXUM1 No better time to thank all my co-authors @RhysGoodall, @PhilippBenner2, Yuan Chiang @cyrusyc_tw, @Bowen_D_, Mark Asta, Gerbrand Ceder @cedergroup,
4
13
60
Super practical tool for placing adsorbates on surfaces! Quick and exhaustive. Great to see it out @efssh
Started with some slides for AutoAdsorbate-ended up with an animated deep dive into surface site detection. It works on any surface, no bulk symmetry needed. Here’s what this one-liner really does: s=Surface(slab211) code: https://t.co/p0nfGeDNzc paper: https://t.co/3ZdXPgqo3Q
1
0
12
Check out our latest work on a new quantum primitive to compute scalar products with phase information, with a reduced circuit depth - preprint available! 😊 https://t.co/ChSBNCO2Yt
@ThomGroupCam @QuantinuumQC @ChemCambridge
arxiv.org
The measurement of scalar products between two vectors is a common task in scientific computing and, by extension, in quantum computing. In this work, we introduce two alternative quantum circuits...
0
3
7
Come checkout *BoostMD*, a new approach for accelerating machine learning force fields, by leveraging infromation from previous time steps. Today at MLSB @ Neurips (Room E-11). paper: https://t.co/EwH1FjEnxF
1
17
98
Generative models often create molecules that are hard to synthesize. Tomorrow, we present SynFlowNet, a GFlowNet designed to generate synthesizable molecules using documented reactions and purchasable materials. Learn more at the MLSB workshop: https://t.co/jOL5zfaHGz
@jasonhartford "SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints" Where: Machine Learning in Structural Biology workshop in Easy Meeting Room Rooms 11 & 12 When: Dec 15th at 8:15 AM PST https://t.co/jOL5zfa9R1
@MirunaCretu2 @folinoid
0
17
99
Check out OMat24 for inorganic materials! I'm super excited to see what the community does next with this dataset and models. Get in touch if you have questions or feedback!
Introducing Meta’s Open Materials 2024 (OMat24) Dataset and Models! All under permissive open licenses for commercial and non-commercial use! Paper: https://t.co/vYSutPJT7L Dataset: https://t.co/nDZUnSiwL6 Models: https://t.co/MMPq0zKeGi 🧵1/x
3
13
96
There's lots of choice out there of for algorithms to perform simulations at constant-temperature 🌡️. If you want to make an informed choice, there's now a #cookbook recipe ready to try 🧑🍳 https://t.co/QfWtD4ihAl
1
13
58
Honored to have been part of the TechScale workshop on single-atom catalysis. Thanks for the invitation to talk about ML force fields for heterogenous catalysis. And thank you for all the inspiring talks by @gianpacc, @MichalOtyepka, TBenešová and others.
Thank @lars__schaaf for great talk at #TECHSCALE workshop in @ChateauLiblice! There is a bright future of ML FFs! @CatrinUP @CUCAM8 @PumeraGroup
0
1
11
There's been many new Geometric GNNs in the past couple weeks combining 3D equivariance and topological ideas (simplices, cells). I think folks should really try their best to compare to MACE ( https://t.co/hy0Wo46wOb), if possible, b/c its a way to get similar capabilities.
arxiv.org
Creating fast and accurate force fields is a long-standing challenge in computational chemistry and materials science. Recently, several equivariant message passing neural networks (MPNNs) have...
5
16
108
First ever ab-initio solvation free energies AND they match experiment 🤯. Super cool stuff by @jhmchem!
In our latest preprint, we introduce an efficient, alchemical free energy method compatible with maching learned forcefields, enabling calculation of hydration free energies with first principles accuracy! (1/5)
0
0
10
Excited to share a new update to gRNAde code! https://t.co/DaAOGCvelM You can now forward-fold designed RNAs in 2D and 3D (RhoFold) to measure structural self-consistency metrics + cool visualisations! Hope we develop our own version of 'designability' in the RNA world ;)
1
13
78
Come with us about enzyme engineering with AI and ultra high throughput screening at our poster at @gembioworkshop at #ICLR2024 with @MaxGantz_
2
4
51
Curious what’s hard about non-compact group equivariance? Interested in ML for polynomial optimization? Come to Mitchell Harris' and my poster #275 on Friday morning, 10:45am - 12:45pm, to learn more (including a surprising impossibility result)! Paper: https://t.co/6zKJChKBKq
1
9
55
Excited for #ICLR with @IlyesBatatia and Felix Faber. Come checkout Matrix Functions NN tomorrow or DM to have a chat. See how we outperform global attention to capture structured non-locality. 😎 Spotlight: tomorrow 10:45 (Halle B) Updated paper: https://t.co/4yvNXJhA8y
0
5
42
Matthias Wissuwa @UniBonn @PhenoRob presents new variety 'Mavitrika' which strives in Madagascar's low phosphorus conditions and shows increased zinc in the grain in presence of Madagascar's prime minister Christian Louis Ntsay and ministers of Agriculture and Higher Education
Last preparations by Matthias and his team to present the new variety to the high profile audience...
0
4
8
Experiment-driven atomistic materials modeling. MLPs and on-the-fly ML-XPS prediction generate atomistic structures that match the experimental XPS. On arXiv: https://t.co/cmOgNMPC0G Esp. thanks to Tigany Zarrouk, as well as Rina Ibragimova, @apbartok, @SuomenAkatemia, @CSCfi
9
28
193