Carnegie Mellon University

Protein language model that supervises embedding geometry with inter-residue contacts, so representation distance tracks physical distance.

Hybrid framework that predicts ribosome location profiles from mRNA sequence alone, pairing a structure-aware TASEP simulation with a Mamba polisher.

Protein foundation model for de novo enzyme design that co-designs sequence and 3D structure under small-molecule ligand guidance, at 730M parameters.

EEG foundation model pretrained by spectrogram reconstruction that improves online directional motor-imagery brain-computer interface control.

Tokenizer-free genomic foundation model that adaptively chunks raw nucleotides, enabling zero-shot variant fitness and gene essentiality prediction.

Single-cell foundation model running self-attention across all 27,874 human genes, with Gene Ontology priors injected through a graph network.

Structure-based drug design model that unifies de novo generation, docking, conformer generation, and pharmacophore conditioning via flow matching.

Spatial transcriptomics language model that reads tissue as spatial sentences to simulate cell profiles and run in silico perturbations.

Generative framework that reconstructs missing spatial transcriptomics regions by jointly predicting cell locations, cell types, and gene expression.

Retrieval-augmented inverse folding model that fuses structural motif retrieval with a hybrid attention decoder to design sequences for a backbone.

Spectroscopy-grounded molecular foundation model that reads NMR, IR, and mass spectra as text, elucidating structures and generating 3D conformers.

Spatial transcriptomics foundation model pretrained on 22 million cells, encoding each cell with its neighbors for niche and density prediction.

Diffusion model that co-designs binder sequence and backbone for arbitrary protein targets, pretrained on 706,360 protein-protein complexes.

Ligand-binding protein design driven by a natural-language function description plus a ligand SMILES string, in 1B and 3B parameter variants.

Protein structure retrieval model aligning 3D structures with functional text via contrastive learning, for zero-shot search of PDB and cryo-EM maps.

Bidirectional DNA foundation model with a Mamba-attention-mixture-of-experts design, reading 1 million base pairs at single-nucleotide resolution.

Pharmacophore-conditioned diffusion model generating 3D molecular graphs that satisfy a given pharmacophore hypothesis without a target structure.

Pocket-based molecular docking with an SE(3)-equivariant diffusion transformer. Places 77.65% of top-1 poses within 2 Å RMSD on PoseBusters.

Machine-learned interatomic potential supplying quantum-quality geometric restraints for refining cryo-EM and X-ray protein structures in Phenix.

Protein-conditional RNA design model that generates binding RNA sequences for any target protein, with no post-generation optimization step.

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

De novo protein sequence design conditioned on target vibrational dynamics, pairing a diffusion designer with a predictor that ranks candidates.

Long-context generative genomic foundation model with a 98k-nucleotide window, trained on 386 billion bases of eukaryotic DNA for sequence design.

Autoregressive transformer pretrained on 2,000 hours of intracortical spiking activity, decoding motor intent across subjects, species, and tasks.

Adenylation domain substrate specificity prediction from frozen ESM-2 embeddings, with zero-shot calls on substrates absent from training.

Cell type annotation from multiplexed tissue images, using a pretrained Vision Transformer ensemble that runs on new panels without fine-tuning.

Protein structure tokenizer that discretizes backbones into 512 discrete tokens and reconstructs all-atom structures, including side chains.

Vector-based virtual screening model that co-embeds proteins and small molecules so a drug-target interaction reduces to a single dot product.

De novo peptide generation from four ProtGPT2 fine-tunes, one per design goal: hemolytic, non-hemolytic, non-fouling, and soluble sequences.

Multimodal large language model that interprets 12-lead electrocardiogram images, answering open-ended clinical questions and generating ECG reports.

Multi-modal protein foundation model aligning 3D structure and literature text to a sequence anchor through contrastive pretraining.

Transformer-based single particle tracker for fluorescence microscopy, using multi-hypothesis attention to link particles at low SNR and high density.

Multi-modal self-supervised transformer for regulatory genomics, pre-trained on DNA sequence together with transcription factor binding matrices.

Self-supervised molecular graph model contrastively pretrained on ~10M unlabeled PubChem molecules, then fine-tuned for property prediction.