University of Illinois Urbana-Champaign

Conformational ensemble generation between two anchor structures, mixing inverse-folding probabilities to prompt a frozen structure predictor.

Cryo-EM density enhancement for protein-ligand binding sites, sharpening weak ligand maps with a 3D Swin-Conv UNet trained on 6,511 complexes.

Variable-length generative protein design across structure, sequence, motif scaffolding, and peptide co-design via a generalized Poisson flow.

All-atom protein co-design model that generates sequence and structure together in one unified diffusion process, aimed at hard binder design.

Protein backbone design model pairing flow matching with a Mamba state-space backbone, generating long proteins in linear time with exact geometry.

Predicts residue-residue dynamic contact maps from a single sequence, matching molecular dynamics ensemble methods at a fraction of the compute.

Inverse folding model refined by online reinforcement learning against folding and stability rewards, cutting design failure rates by 36-48%.

Promptable medical image segmentation trained only on procedurally generated synthetic images, then applied zero-shot to CT, MRI, and ultrasound.

Full-atom flow matching model that generates a ligand and the induced-fit holo pocket together, starting from an apo binding site.

Single-channel EEG tokenizer that learns a discrete vocabulary of time-frequency motifs, usable as a front end for existing EEG foundation models.

Photoplethysmography foundation model pretrained on raw wearable signals from a field study, transferring across lab and field health tasks.

Molecular docking framework that poses several ligands sharing one protein pocket at once, using their consistency to sharpen each prediction.

Motion foundation model for wearable accelerometry, trained with relative contrastive learning on 1B segments from 87,376 participants.

Target-conditioned peptide binder design model that samples hot-spot residues from an energy-based density, then extends fragments autoregressively.

ECG foundation model that learns discrete rhythm tokens from noisy real-world recordings for arrhythmia classification and anomaly detection.

Single-cell transcriptomic aging clock predicting immune age for CD8+, CD4+ T and NK cells, and transferring to bulk whole-blood RNA-seq.

Cellular senescence prediction from protein sequence, pairing ESM-2 embeddings with a hybrid BiLSTM-CNN classifier at 86.43% test accuracy.

Lasso peptide language model that adapts ESM-2 to threaded RiPP core sequences, supplying embeddings for cyclase substrate and activity prediction.

Reprograms a frozen single-target diffusion model for dual-target drug design by composing SE(3)-equivariant messages across two aligned pockets.

Molecular docking and design foundation model that unifies structure-based drug design and peptide design at the atom level in one checkpoint.

Chemical language model that tokenizes each atom by its functional group, giving transferable embeddings for molecular property prediction.

Full-atom peptide binder design against a target pocket, generating backbone frames, side-chain torsions and residue types in one joint flow.

Structure-based drug design that generates 3D ligands for a protein pocket entirely in the continuous parameter space of a Bayesian flow network.

3D molecule generation that models atom coordinates and element types as distribution parameters updated by Bayesian inference, not by denoising.

Structure-based drug design by SE(3)-equivariant diffusion over 3D atom coordinates and types, with the same frozen network scoring binding affinity.

Protein language model that annotates intrinsically disordered regions per residue from sequence alone, without MSAs or biophysical features.

Protein language model over byte-pair-encoded amino acid tokens, fine-tuned for protein family classification and binary interaction prediction.