Leo Klarner (@leoklarner) header
Leo Klarner (@leoklarner) profile picture

PhD Student, Clarendon Scholar at University of Oxford. @OPIGlets & @oxcsml. Working on data-efficient generative models for early-stage drug discovery.

Joined January 2021
Photo by Leo Klarner (@leoklarner)Photo by Leo Klarner (@leoklarner)Photo by Leo Klarner (@leoklarner)Photo by Leo Klarner (@leoklarner)
📣Come work with us 🤗! We're recruiting a postdoc fellow in AI for small molecule drug discovery. A special opportunity to benefit from huge data efforts @thesgconline and have your predictions tested in the lab. apply here 👉 tinyurl.com/25fdka28 deadline 🗓️ 31 Dec '24
PDF position in Toronto at the interface of deep learning with @cjmaddison at @VectorInst and #compchem in my group at @thesgconline. Develop novel approaches in hit-finding and hit-to-lead opt. Predictions are tested at the bench at @thesgconline bit.ly/3SvHXCU
2
21
2
50
18,355
So many great new papers in diffusion space over the last few weeks, really highlights the flexibility of this modelling paradigm. Some key papers --> 1/ Diffusion Forcing: can introduce independent noise on particular dimensions (e.g. time). train with independent noise, then can inference with structured noise (e.g. causally) 2/ DisCo-Diff: mixing diffusion models with discrete latents. this allows mixing high level discrete latent variables to potentially improve learning of data-level diffusion modelling 3/ Context Guided Diffusion: leverage labelled for conditional generation, together with unlabelled data to provide regularisation and improve OOD generalisation Paper links - Diffusion Forcing arxiv.org/abs/2407.01392 - DisCo-Diff arxiv.org/abs/2407.03300 - Context Guided Diffusion arxiv.org/abs/2407.11942
4
45
2
299
31,156
Leo Klarner retweeted
💥 Introducing PINDER & PLINDER With existing evals saturating without clear advances in real life downstream tasks, current progress in AI x Bio is primarily rate-limited by better datasets & evals. In two back-to-back preprints, we address this via fantastic academic-industry collaborations by @vant_ai @NVIDIA & the group at @TorstenSchwede that co-organizes CASP, @MIT: Read the release blog posts at: plinder.sh/blog/plinder & pinder.sh/blog/pinder - 🧵(1/6)
6
101
17
348
98,483
Leo Klarner retweeted
Context-Guided Diffusion for Out-of-Distribution Molecular and Protein Design | ICML 24' - Introduce a regularizer that uses unlabeled context data to encourage high uncertainty and smooth gradients in OOD regions - Construct a covariance matrix from model embeddings to control prediction variance and enforce smoothness in guidance signals - Evaluate on Small Molecules, Materials and Protein Sequences (Antibody) Preprint: arxiv.org/abs/2407.11942
11
1
58
6,450
I am super excited about this #ICML2024 paper! Context-Guided Diffusion is extremely easy to use and worked out-of-the-box when we applied it to molecular and protein design problems!
A semi-supervised method to improve the out-of-distribution performance of guided diffusion models. @leoklarner @timrudner @OPIGlets @yeewhye
4
24
1,795
In our #ICML paper, we develop Context-Guided Diffusion, a **plug-and-play method** that can be applied to any trained diffusion model **at test time**. Context-Guided Diffusion is **embarrassingly easy to use** and derived from Bayesian principles. 🚀 📃arxiv.org/abs/2407.11942
Generative models for molecular optimization & protein design often rely on data-driven guidance functions for conditional sample generation. Our new #ICML2024 paper presents a simple but effective approach to improve their performance in OOD settings. arxiv.org/abs/2407.11942
2
11
67
9,127
A semi-supervised method to improve the out-of-distribution performance of guided diffusion models. @leoklarner @timrudner @OPIGlets @yeewhye
2
20
1
145
13,080
Generative models for molecular optimization & protein design often rely on data-driven guidance functions for conditional sample generation. Our new #ICML2024 paper presents a simple but effective approach to improve their performance in OOD settings. arxiv.org/abs/2407.11942
5
35
1
183
23,036
Our method is easy to implement, applicable to any diffusion process & causes no overhead at sampling time. We show that it consistently outperforms existing conditional sampling and guidance techniques across applications from drug discovery, materials science & protein design.
1
2
457
Working on this project with @timrudner @OPIGlets and @yeewhye has been a lot of fun! Looking forward to discussing it during our poster session on Thu 25 Jul 11:30am-1pm! Paper: arxiv.org/abs/2407.11942 Code: github.com/leojklarner/conte… Poster: icml.cc/virtual/2024/poster/…
1
1
4
1,732
Generative models for scientific data can often benefit from using domain-informed constraints to reduce their problem space. In our new #NeurIPS2023 paper, we simplify constrained (Riemannian) diffusion models to make them better, faster & easier to use. arxiv.org/abs/2307.05439
2
31
3
207
25,788
We provide a theoretical justification for our method's use in diffusion models and demonstrate its empirical performance across problems from geospatial modelling, robotics and protein design using both convex and non-convex constraints.
1
577
🚀Join us at TODAY's 2pm poster session, where @leoklarner and I will be presenting our #ICML paper Drug Discovery under Covariate Shift with Domain-Informed Prior Distributions over Functions -> 2pm, Hall 1 #124 📄Paper: proceedings.mlr.press/v202/k… 💻Code: github.com/leojklarner/Q-SAV…
1
2
9
1,556
Can we use scientific prior knowledge to make models more robust in low-data, OOD drug discovery? In our new #ICML2023 paper, we use prior knlg of drug-like chemical space to regularize the function space of neural networks, improving OOD generalization. proceedings.mlr.press/v202/k…
2
20
2
106
11,205
This framework also presents researchers with a transparent and probabilistically principled way to encode arbitrary domain-informed modeling preferences, such as compound availability, patentability, likely adverse side effects, etc.
1
1
397
Working on this with @timrudner, Torsten Schinder, Michael Reutlinger, @gmm, Charlotte Deane, and @yeewhye has been a lot of fun - thank you! Paper: proceedings.mlr.press/v202/k… Code: github.com/leojklarner/Q-SAV… Poster: icml.cc/virtual/2023/poster/…
2
6
745