Harvard Medical School

Pan-cancer clinico-genomic model for treatment response and survival prediction, transferring zero-shot to unseen hospitals and cancer types.

Medical imaging foundation model unifying pathology and radiology, serving classification and segmentation on 2D, 3D and gigapixel inputs.

Intracortical speech brain-to-text decoder jointly pretrained across six BCI users, with over 50% lower relative word error than per-user models.

Spatial proteomics foundation model for multiplex immunofluorescence, with a 268-marker vocabulary and marker-conditioned 768-dimensional embeddings.

Vision-language model for neuroblastoma pathology that reads H&E slides with their reports to grade tumors, infer biomarkers and stratify risk.

Transformer that predicts protein-protein interactions at residue resolution, spanning mutations, PTMs, peptide-MHC binding, and disease variants.

RNA language model that predicts secondary structure of internal ribosome entry sites from sequence alone, trained on roughly 50,000 IRES sequences.

Prime editing efficiency prediction from pegRNA sequence, with every biochemical step of the editing mechanism modeled as its own learned rate.

Vision-omics finetuning that aligns pathology foundation models with spatial transcriptomics so morphology features predict local gene expression.

Full-atom SE(3)-equivariant diffusion model that inpaints binding interfaces to design proteins that bind DNA, RNA, and small molecules.

Native 3D vision transformer self-supervised on unlabeled fluorescence microscopy volumes, segmenting subcellular structures without voxel labels.

Siamese protein language model whose embedding distances approximate TM-score and lDDT, enabling alignment-free protein structure comparison.

Predicts single-cell scRNA-seq coverage and scATAC-seq insertion profiles from DNA sequence, adapting the Borzoi trunk with a cell-specific decoder.

Peptide-MHC class I binding predictor that scores force-field energy terms from modeled pMHC structures, holding precision on rare HLA alleles.

Multimodal diffusion transformer for de novo protein design, jointly generating sequence and structure conditioned on 465 Gene Ontology functions.

Modality-agnostic foundation model for human brain imaging that runs five core neuroimaging tasks across uncalibrated CT and MRI without retraining.

Generalist MRI vision-language foundation model that handles reconstruction, segmentation, abnormality detection, and report generation in one model.

Attention-based multiple instance learning heads for whole-slide pathology, pretrained on a 108-way pan-cancer slide classification task.

Histopathology image synthesis from a latent diffusion model conditioned jointly on unpaired diagnostic text reports and cell-type masks.

Self-supervised medical imaging foundation model pretrained on 3.3 million CT, X-ray, ultrasound, pathology, OCT, fundus, and dermoscopy images.

Protein fitness prediction with end-to-end differentiable homology search, replacing MSA construction with vector search over 62M UniRef50 sequences.

Spatial proteomics foundation model, marker-aware and panel-agnostic, pretrained on 47 million multiplexed tissue-imaging patches from 175 markers.

Distribution-level representation learning that embeds whole cell populations, perturbation responses, and sequence sets, not individual data points.

Promptable 3D medical image and video segmentation foundation model fine-tuned from SAM 2.1, cutting lesion annotation time by up to 92%.

2D/3D X-ray to volume registration for surgical navigation, aligning live fluoroscopy to preoperative CT or MRI after a five-minute fine-tune.

Clinical imaging encoder multitask-pretrained across X-ray, mammography, dermoscopy, fundus, ultrasound, CT, and histopathology for few-shot transfer.

Drug-combination safety prediction that fuses molecular structure, pathway knowledge, cell viability, and transcriptomic response to perturbation.

PET/CT foundation model pretrained by cross-modal masked autoencoding on whole-body scans, for tumor lesion segmentation and lymphoma staging.

Pan-tumour CT foundation model pretrained on 30,000 synthetic 3D scans carrying lesion masks and structured reports across ten organ systems.

Slide-level pathology foundation model that encodes a whole-slide image of any size into one embedding, supervised by paired sequencing data.

Protein language model that jointly embeds a set of sequences and reconstructs phylogenetic trees without alignments or guide trees.

3D vision foundation model for computed tomography, contrastively pretrained on 148,000 scans for segmentation, triage, retrieval, and concept search.

Histopathology foundation model pretrained on 200 million H&E and immunohistochemistry tiles from more than 350,000 whole-slide images.

Genetic language model predicting disease risk and cell-type-specific expression changes from up to 88 megabases of an individual's genome sequence.

Vision-language foundation model for precision oncology, pretrained on 50M pathology images and 1B text tokens via unified masked modeling.

CRISPR-Cas PAM specificity prediction directly from Cas protein sequence, plus computational evolution of Cas9 variants toward a chosen PAM.

Multiomic foundation model for zero-shot in silico perturbation, predicting gene regulation and cell fate transitions from DNA and ATAC signal.

Multimodal foundation model integrating protein sequence, structure, and natural language to model and generate protein phenotypes across scales.

Histopathology patch encoder turning 512x512 tiles into 768-dimensional features, trained with a CoCa objective on 1.26 million captioned images.

Self-supervised vision foundation model for structural brain MRI, providing a reusable encoder for brain age, survival, and image classification.

Slide-level pathology foundation model turning whole-slide images into reusable embeddings for classification, retrieval, and report generation.

Histopathology foundation model aligned to spatial transcriptomics by a cross-modal ranking loss, embedding H&E patches without gene input.

Cell-type-specific gene expression prediction from DNA sequence, mapping Enformer epigenomic features to pseudobulk expression for cell-resolved TWAS.

Conditional autoregressive genomic language model trained on 13.6M mammalian promoters, scoring promoter variants, including indels, zero-shot.

DNA methylation foundation model over 49,156 array CpG sites. Imputes missing values, embeds samples, and predicts epigenetic age and disease risk.

Light-weight rigid protein-ligand docking model adapting the AlphaFold2 architecture, with a binding affinity module for virtual screening.

Joint embedding space for common SNPs and free-text clinical concepts, aligned by contrastive learning over GWAS, biobank and knowledge-graph pairs.

Convolutional ECG foundation model trained on expert annotations spanning 150 diagnostic categories, with 12-lead and single-lead wearable variants.

Manufacturing-aware generative sequence models whose parameters are DNA synthesis reaction conditions, so designs are made in vitro at petascale.

Weakly supervised histopathology foundation model pretrained on 60,530 whole-slide images for cancer detection, prognosis, and molecular prediction.

Slide-level pathology foundation model that learns whole-slide embeddings by aligning multiple stains of the same tissue during pretraining.

Multi-omics transformer generating transcriptomic, methylation and proteomic signatures for a given tissue, disease, age group, sex and compound.

Multimodal vision-language copilot for pathology that answers open-ended questions about histology images and reasons about differential diagnoses.

DeepLabV3 segmentation model that separates tissue from glass background in H&E and IHC whole-slide images, as used by the HEST-Library.

Protein language model that predicts which of twelve subcellular compartments and biomolecular condensates a human protein partitions into.

Histopathology vision-language foundation model pretrained on 1.17 million image-caption pairs with contrastive and captioning objectives.

Computational pathology foundation model (ViT-L/16, DINOv2) pretrained on over 100 million H&E tiles from more than 100,000 whole-slide images.

Bulk tumor transcriptome model ensembling hundreds of variational autoencoders into interpretable cancer-specific latent spaces for 18 cancers.

Self-supervised 3D CT foundation model that extracts general-purpose tumor representations for cancer imaging biomarker discovery and prognosis.

Blind protein-ligand docking that transfers to binding domains absent from training, scoring 22.6% top-1 on DockGen and 50% on PoseBusters.

Multimodal medical vision-language model for few-shot visual question answering, learning new imaging tasks from in-context examples at inference.

Enzyme optimum pH prediction from sequence, ensembling a light-attention network and a support vector regression over frozen ESM-1v embeddings.

Medical image segmentation model that solves unseen segmentation tasks in context from a few labeled examples, with no retraining or fine-tuning.

Autoregressive protein language model scoring variant effects zero-shot, blending sequence likelihood with homolog statistics retrieved at inference.

Mass spectrometry proteomics foundation model that jointly embeds MS/MS spectra and peptides in one space, for open and error-tolerant search.