Keep calm and apply Bayes' Rule.

I enjoyed giving a seminar at Stanford today, thanks for all the attendees for allowing yourselves the distraction from your ICLR reviews to listen to some GenAI for science work.

Today we’re releasing VariantFormer — a 1.2B-parameter DNA language model built to encode and interpret human genetic variation. Blog: https://t.co/Rnrtt9KkXN Model: https://t.co/MOdUgpJoHj Preprint:

Both models share the idea that we want to study biology virtually. VF allows us to generate counterfactuals such as with in silico mutagenesis to see “what would happen to liver tissue if allele X was mutated in this person” which are impossible to observe in practice. Models

Why does scLDM matter? Why does scLDM matter? Conditional cell generation is a key capability for virtual cell modeling - enabling virtual experiments such as generating counterfactual cells based on patient identity, cell type, perturbation, or experimental condition. scLDM

How well does scLDM work? scLDM advances the state of the art in generative modeling of transcripts, enabling conditional gene expression generation with unprecedented fidelity. Its capabilities are demonstrated in challenging tasks such as conditional generation of cell atlas

What is scLDM? scLDM is a deep latent diffusion model that combines a transformer variational autoencoder with score matching in latent space. It adapts ideas from generative models for other modalities (such as images and molecules) to generate cellular transcriptomes -

2. scLDM: single cell latent diffusion model This deep generative model of gene expression represents an evolutionary step for our transcript models. Work led by @g_palla1, @jmtomczak and the scLDM team Preprint: https://t.co/bjk2kfhiU2

Model assets:

How well does it work? Across a wide range of tasks - including cross-individual and cross-tissue gene expression prediction, somatic mutation effect generalization, in-silico mutagenesis, variant effect prediction, variant prioritization, and disease-risk interpretation (in

1. VariantFormer: from human genomes to populations VariantFormer strives to elucidate how human cells differ as a consequence of individual variation based on personalized genomes. What is VariantFormer? VariantFormer is the first genomic foundation model that bridges

Releasing two new models today: VariantFormer and scLDM Our goal at Biohub is to advance virtual biology by merging frontier biology with frontier AI models. Here's what's new from my team:

Science article on AI x Bio for cells Excited to share this article in science discussing what’s happening in AI x Bio for cells across a representative set of different teams across industry, philanthropy, and academia, also featuring our efforts at CZI. It was a privilege to

It was great to spend time with my colleagues and chart the future directions for the interface of probabilistic/Bayesian reasoning and generative AI.

Here’s also a link to the new and updated version of the TranscriptFormer paper: https://t.co/EuUshMIfFZ

Time article discussing our work We recently had the opportunity to talk to TIME magazine about the burgeoning landscape of virtual cell research across the frontier AI and biology teams, which I see as a position piece about the future potential of this vision. Also fun to see

Multiple ML/AI roles available at https://t.co/jgDolM2K8Z focused on building atomistic foundation simulation models. If you’re in the market and interested in AI-infra\research engineering, or GenAI-research roles reinventing our understanding of matter and reasoning about

Really interesting work thinking about the impact our scientific foundation models can have

We are disappointed that some community members have been sending threatening and disrespectful messages to our organizers after NeurIPS decisions. While we always welcome feedback, we wish to remind all that our organizers are volunteers who work tirelessly on behalf of the