University of Tokyo

Reasoning LLM that predicts antimicrobial susceptibility of clinical bacterial isolates and supplies mechanistic explanations for each prediction.

Enzyme function prediction that scores whether two sequences catalyze the same reaction, via attention pooling over frozen ESM Cambrian embeddings.

Reasoning-guided foundation model for de novo antibody CDR design, pairing a multimodal LLM understanding expert with a Boltz-1 diffusion expert.

GPT-style generative language model for mRNA coding sequences, pretrained across bacteria, eukaryotes, and archaea for de novo CDS design.

Cross-species brain spatial transcriptomics foundation model pretrained on 133M cells from human, macaque, marmoset, and mouse whole brains.

Hierarchical immune cell type annotation for scRNA-seq, rendering expression as images for a CNN. 93.2% mean subtype accuracy over seven datasets.

Domain-level language model treating Pfam protein domains as tokens to predict and design bacterial and fungal biosynthetic gene clusters.

Medical vision-language model trained with reinforcement learning for generalizable reasoning across eight imaging modalities and five question types.

Protein language models evotuned on influenza A hemagglutinin, with a pLM entropy metric scoring per-site conservation from a single input sequence.

Sequence-only TM-score prediction pairing frozen ProtT5 embeddings with a bidirectional GRU and multi-scale convolution for protein homology search.

Protein backbone generation with a diffusion model whose noise schedule is derived from the renormalization group rather than heuristically tuned.

Tumor microenvironment segmentation on H&E slides, labeling 13 tissue and cell components from a single model in semantic or panoptic form.

Structure-to-sequence protein design network inverting trRosetta, predicting coevolution features from a backbone to generate stable sequences.

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

EEG foundation model that pretrains a distance-weighted electrode graph ahead of its convolutional encoder to capture inter-channel relationships.

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

Chromatin accessibility prediction across 29 human immune cell types, using in-silico saturated mutagenesis to score constrained regulatory regions.

Chromatin-state language model pretrained on ROADMAP annotations from 127 human cell types to find chromatin-state motifs and predict gene expression.