We are happy to Announce our Latest #SilverSponsor Nostrum Biodiscovery for our Digi-Tech Pharma & AI 2023 Conference to be held on the 10th & 11th of #may this year at #London, UK! #DTP #Digitechpharma Check out the Latest Agenda & Speakers: corvusglobalevents.com/conference/dig…

Our latest work at Nostrum Biodiscovery is out! We trained a protein language model of only 14M parameters, a huge downsize in comparison with the latest pLMs. Gen. seqs. have similar properties to the natural ones, show great folding and are stable in MD. Stay tuned for more!

Our team is thrilled to share our latest work: the development of a new approach to protein sequence generation using a Small Scale Protein Language Model (SS-pLM). We invite you to read our arxiv: lnkd.in/dZt8-qvK

Sign in for a more detailed view on AI driven HTVS and on how we leverage generative models to surpass IP crowded spaces in relevant targets. Thanks CCPBioSim for having me!

Are Protein Language Models Compute Optimal?. arxiv.org/abs/2406.07249

Scoreformer: A Surrogate Model For Large-Scale Prediction of Docking Scores. arxiv.org/abs/2406.09346

Are Protein Language Models Compute Optimal? | ICML 24' Workshop - Investigate the compute efficiency of pLMs (encoder-only models) by adapting NLP scaling laws to determine the optimal ratio between model parameters and training tokens - Identify a training loss plateau,

If you want to work on AI projects where pretraining isn’t dead, join a techBio drug discovery company 😜 CC: Ilya Sutskever

ADDRESSING MODEL OVERCOMPLEXITY IN DRUG-DRUG INTERACTION PREDICTION WITH MOLECULAR FINGERPRINTS 1. This study challenges the trend of increasingly complex DDI prediction models by showing that simple, interpretable molecular representations—like Morgan fingerprints (MFPS)—can

Heading to ICLR'25! We'll be presenting 5 different papers on learnable tokenization in biomolecular language models, binding site prediction, pocket opening with flow matching, drug-target and drug-drug interaction prediction. Wide range of topics we'll be happy to chat about!

Active Learning-Guided Seq2Seq Variational Autoencoder for Multi-target Inhibitor Generation １．This paper introduces a structured active learning (AL) framework built around a Seq2Seq Variational Autoencoder (VAE) to generate molecules with high predicted affinity to multiple

Geodesic-guided diffusion for more accurate protein–ligand docking In drug discovery, predicting how a small molecule binds to a protein is essential—but even advanced generative methods like DiffDock can misplace ligands or produce unrealistic poses. Raúl Miñán and coauthors

SESAME: OPENING THE DOOR TO PROTEIN POCKETS 1. SESAME is a novel generative model that efficiently predicts protein conformational changes from apo to holo states, addressing a key challenge in drug discovery. Traditional methods like molecular dynamics are computationally

Evolutionary and Geometric Signatures Reveal Ligand-Binding Sites Across Proteomes 1. A new deep learning model called PickPocket has been introduced to identify ligand-binding residues in proteins at a proteome scale. This model combines evolutionary information from protein