Introducing Neo-1: We're thrilled to announce Neo-1—the world's most advanced atomistic foundation model, unifying all-atom structure prediction and de novo generation to decode and design life's molecular structures. Neo-1 employs the latest latent diffusion techniques to reason simultaneously about molecules and their structures, replacing multiple specialist models with one versatile framework. Design: Neo-1 generates diverse, drug-like small molecules and natural proteins from sequences, properties, or partial inputs. This enables completely novel applications, such as rational molecular glue design directly from sequences. Decode: Neo-1 achieves state-of-the-art accuracy in all-atom structure prediction, excelling particularly in complex ternary systems. NeoLink: NeoLink is VantAI's experimental platform using cross-linking mass spectrometry (XLMS) to capture structural constraints across the entire proteome, which Neo-1 integrates into precise atomic models. Programmability: Highly programmable, Neo-1 supports molecular generation, protein design, structure prediction, and molecular inpainting. Neo-1 can be guided using partial structures, sequences, or desired molecular properties. Drug Discovery & Optimization: Beyond hit discovery, Neo-1 leverages experimental data, accelerating iterative optimization in drug discovery. Applications already include molecular glues, PROTACs, and antibodies, significantly streamlining the path from computational design to experimental validation. Learn more: https://www.vant.ai/neo-1 Interested in pioneering the next frontier in biology and AI research? We're hiring—contact recruiting@vant.ai. https://lnkd.in/ekF8nfPK
Neo-1: Decode & Design the Structures of Life
https://www.youtube.com/
Kudos to all the amazing folks in the VantAI team, amazing work yet again 👏
Cool stuff
Hi, amazing work! However, one thing remains unclear to me: how exactly was the train set created? Did you start from the PLINDER/PINDER train splits and then shorten them according to the time cutoff, or did you use all available structures from these datasets and subsequently split them based on the time cutoff to create train/val/test?
I do not grasp all you present here, but I am very proud of you and your accomplishments. Congratulations! You are always welcome.
Congratulations, Chalada!
Senior researcher @ Microsoft Research
8moGreat work guys!