Tomas Geffner
@tomasgeffner
Followers
287
Following
155
Media
7
Statuses
36
(ex) tennis, (trying) ML. Research Scientist at NVIDIA, views are my own. Previously, intern at DeepMind, VantAI, Microsoft research, and Amazon AWS.
Joined July 2019
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 🧵
2
40
183
RT @KevinKaichuang: Partially-latent flow matching enables sequence-structure codesign of large proteins and functional motif scaffolding.….
0
13
0
RT @anthonycosta: Awesome work from a great collaborative team at @nvidia @NVIDIAHealth! Be on the lookout very shortly for code release!.
0
5
0
RT @NVIDIAHealth: Hybrid explicit-latent flows are the new foundation models for protein structures. La-Proteina shows that one network ca….
0
23
0
RT @Mattmcpartlon1: These guys are crushing all of the benchmarks 🦾. Congrats to @tomasgeffner, @DidiKieran and team!.
0
4
0
In summary, La-Proteina is a versatile and scalable generative model that produces high-quality, full-atom protein structures, enabling more advanced and practical protein design tasks.
1
0
5
La-Proteina also masters unindexed scaffolding. The model finds the motif's place in the sequence, a harder & more practical task.
1
0
2
La-Proteina excels at atomistic motif scaffolding 🛠️.🔹 All-atom: The model gets the full motif structure (backbone & side-chain). 🔹 Tip-atom: Only critical atoms at the side-chain tips of the motif are provided. (Below, motif overlayed in red.)
1
0
2
Generated structures are physically realistic. They score better on MolProbity analysis & capture natural side-chain conformations (rotamers).
1
0
2
Results? SOTA co-designability & diversity 🚀 La-Proteina generates valid proteins up to 800 residues, where other models often collapse.
1
0
6
This simplifies the problem by avoiding mixed data types and variable data. We can then use standard flow matching in continuous spaces to jointly generate the backbone and latents.
1
0
5
Generating full-atom proteins is hard. It's a mix of discrete sequences & continuous coordinates with variable length side chains. We propose a partially latent representation for proteins ✨ modeling backbone explicitly, while encoding sequence & side-chain in latent variables.
1
0
4
Project Page: Paper: Code and weights: Coming soon!. Work by: @tomasgeffner, @DidiKieran, @ZhonglinJC, Danny Reidenbach, @Oxer22, @sacdallago, Emine Kucukbenli, @karsten_kreis, @ArashVahdat.
arxiv.org
Recently, many generative models for de novo protein structure design have emerged. Yet, only few tackle the difficult task of directly generating fully atomistic structures jointly with the...
2
3
11
RT @vant_ai: Announcing Neo-1: the world’s most advanced atomistic foundation model, unifying structure prediction and all-atom de novo gen….
0
376
0
RT @NaefLuca: The NVIDIA bio generative modelling folks @tomasgeffner @DidiKieran @karsten_kreis @ArashVahdat & co (and their interns :))….
0
5
0
RT @vant_ai: 📢 Join us next week for a talk with @karsten_kreis and @tomasgeffner on their recent paper Proteina: Scaling Flow-based Protei….
0
14
0
Check out Proteina! Paper, code, and nice visuals on the project page
research.nvidia.com
Proteina is a novel flow-based protein backbone generative model. It is trained with flow matching, leverages a scalable and efficient transformer architecture, and offers hierarchical fold class...
📢📢 "Proteina: Scaling Flow-based Protein Structure Generative Models". #ICLR2025 (Oral Presentation). 🔥 Project page: 📜 Paper: 🛠️ Code and weights: 🧵Details in thread. (1/n)
0
6
23
RT @chaidiscovery: Chai-1 has always been available for commercial use via our server. Today, we're also making Chai-1(r) code and weights….
github.com
Chai-1, SOTA model for biomolecular structure prediction - chaidiscovery/chai-lab
0
61
0
📢Annoucing EDLM, our brand-new Energy-based Language Model embedded with Diffusion framework!. Key results:.1. We (for the first time?) almost match AR perplexity. 2. Significantly improved generation quality. 3. Considerable sampling speedup without quality drop. 🧵1/n
0
2
11