Chan Zuckerberg Initiative

Latent diffusion model for generating single-cell gene expression profiles, pairing a permutation-invariant autoencoder with a diffusion transformer.

Single-cell latent diffusion model fine-tuned on 14.5 million CD4+ T cells to simulate transcriptomic effects of single-gene perturbations.

Hierarchical transformer with 1.2 billion parameters that predicts personalized gene expression from diploid genomes for variant effect prediction.

Open-source Apache-2.0 reproduction of AlphaFold3 that predicts all-atom structures of proteins, RNA, DNA, small molecules, and their complexes.

Diffusion model for multichannel fluorescent cell microscopy, generating morphologically plausible images aligned to OpenPhenom phenotypic embeddings.

Reasoning language model post-trained on virtual cell simulations, answering questions about gene perturbations and their effects in natural language.

Single-cell transcriptomics foundation model that encodes gene regulatory network structure into self-attention through graph signal processing.

Generative single-cell foundation model trained on 112 million cells from 12 species, autoregressively modeling gene identities and expression counts.

Cryo-ET particle picking ensemble of three 3D segmentation models predicting particle-center heatmaps with ResNet50d and EfficientNetV2-M backbones.

Cryo-ET particle picking model, an ensemble of 3D U-Nets with EfficientNet encoders that finds protein complexes in tomograms by heatmap segmentation.

Cryo-ET particle picking model that averages tiny, medium, and large 3D U-Nets pretrained on simulated tomograms and fine-tuned on experimental data.

Cryo-ET particle picking model that localizes six protein complexes in tomograms using an ensemble of lightweight 3D U-Nets.

Cryo-ET segmentation framework adapting SAM2 to vesicles and membrane-bound compartments in tomograms and 2D micrographs, zero-shot or fine-tuned.

Cryo-ET particle picking model that localizes and classifies multiple protein complexes in a tomogram with a single 3D U-Net forward pass.

Variational autoencoder that learns interpretable representations of protein subtomograms from cryo-ET, trained on 5.8 million synthetic particles.

Vision transformers trained on Human Protein Atlas fluorescence microscopy for subcellular protein localization and cell morphology representation.

Single-cell perturbation prediction model that adds gene-level language embeddings from NCBI, UniProt, and Gene Ontology to scGPT representations.

Diffusion model translating in both directions between protein sequences and fluorescence microscopy images to predict subcellular localization.

Self-supervised contrastive model embedding cell and organelle dynamics from time-lapse microscopy for cell-state analysis without manual labels.

Variational autoencoder pretrained on 74 million human single-cell transcriptomes from the CELLxGENE Census for batch correction and cell typing.