Southern University of Science and Technology

EEG foundation model that corrects low-frequency bias by reconstructing band-standardized time-frequency targets. State of the art on 24 of 41 tasks.

Protein language model that predicts which of eight lipid categories a protein binds from sequence alone, plus binding sites and mutation effects.

Single-cell perturbation response prediction using dual conditional diffusion bridges that link unpaired control and perturbed populations.

RNA language model that reads full-length transcripts up to 10,000 nucleotides, pairing bidirectional state space layers with multi-head attention.

RNA foundation language model pretrained on mammalian and viral genomes, fine-tuned to predict translation efficiency, half-life, and splice sites.

Nanopore basecaller for fully 5-hydroxymethylcytosine-substituted DNA, reading raw ion current from strands that standard basecallers cannot resolve.

Multimodal ECG-language model aligning 12-lead waveforms with clinical report text for conversational cardiac diagnosis and report generation.

Mitochondrial ultrastructure segmentation in electron microscopy volumes, resolving membranes and cristae with cross-sample domain adaptation.