Teaching models to follow explicit written rules: turning guidelines into environments a model can act in, optimizing prompts for expert annotation, and grounding multimodal reasoning in domain protocols.
Research Areas
Selected publications per topic. See the PI homepage for the full list.
Aligning and fusing heterogeneous sources — images, text, molecules, clinical records — into shared representations, with attention to the geometry of the embedding space and to what happens when modalities are missing.
Whole-slide image analysis, nuclei segmentation, survival prediction and screening from medical images, with a focus on learning under weak labels and on predictions clinicians can interpret.
-
IEEE TNNLS'25Segment Any Cell: A SAM-based Auto-prompting Fine-tuning Framework for Nuclei Segmentation -
-
MICCAI'24PathM3: A Multimodal Multi-Task Multiple Instance Learning Framework for Whole Slide Image Classification and Captioning -
-
-
MIDL'26Guideline-Informed MLLM Reasoning for Pathology-Aware Post-Operative Prostate CTV Segmentation -
-
-
-
Language models and graph methods for genomic sequence and gene annotation: motif-oriented DNA pretraining, Gene Ontology-informed representations, and predicting the phenotypic consequences of gene knockouts.
Representation learning for protein sequence and structure — secondary structure and structural property prediction, self-supervised pretraining from tertiary structures, and enhancing evolutionary profiles — together with the immune receptor side: BCR and TCR repertoires, antigen binding affinity and B-cell epitope prediction.
Pretraining and robust training for molecular property prediction, including learning from limited labels and handling the distribution shift that comes with new chemical space.
Work in Progress
Ongoing projects, with internal notes linked where available.
Quantum features for epitope prediction
In progressWhole-chain electronic structure of antigens as additional features for epitope prediction.
Gene-perturbation
In progressIntegrating multimodal gene representations for predicting transcriptional responses to single and combinatorial gene perturbations.
KEGG
In progressLearning transferable molecular relationships from biological pathway graphs for downstream interaction prediction across pathways and databases.
Shared-Residual World Models
In progressA multimodal world model that splits each modality into a shared summary plus a modality-specific residual instead of forcing one aligned latent. Preliminary ALFRED results: better next-state prediction and native missing-modality inference, with no alignment loss.
Modality-Aware Fusion for AD Conversion Prediction
In progressPredicting progression from MCI to Alzheimer's disease using ADNI imaging, PET, protein, genetic and clinical data. Our modality-aware fusion model matches the strongest tabular baseline on complete data and pulls significantly ahead once input modalities go missing — the setting it would actually be deployed under.
People
Principal Investigator
Ph.D. Students
Mascot
Bunny
Lab Mascot
Opportunities
Welcome to join our team!
We are always looking for motivated students and collaborators who are interested in machine learning, computer vision, and AI for biology and medicine. If our work sounds exciting to you, we would love to hear from you.
Contact
- yguo30@kent.eduDr. Yuzhi Guo, Principal Investigator
- Address
- Department of Computer Science
Mathematical Sciences Building (MSB)
1300 University Esplanade
Kent State University, Kent, OH 44242 - Students
- Prospective students: see Opportunities.























