— 01
Foundation models for life sciences
Large-scale models for genomic and biomedical data, with biology-grounded interpretability that goes beyond anecdotal evidence.
Shasha Zhou
PhD Candidate · University of Exeter · advised by Dr. Ke Li
Hi, I'm Shasha — a Computer Science PhD candidate at the University of Exeter, advised by Dr. Ke Li. I want foundation models in life sciences to be as accountable as the science they support.
My research builds toward this from three directions: models grounded in domain knowledge, agents whose answers trace back to evidence, and evaluation that survives scrutiny beyond held-out test sets.
I expect to graduate in Sep 2027 and am exploring postdoctoral fellowships and industry research roles in AI for Science, foundation models, and trustworthy machine learning.
📍 Exeter, UK · ✉️ sz484@exeter.ac.uk
News latest
recent milestones from my research, collaborations, and academic activities.
Apr 2026
grantAwarded a second Modal Research Grant to support large-scale genomic FM experiments.
Apr 2026
paperOur position paper on genomic model interpretability (with Mingyu Huang) accepted as ICML 2026 Spotlight (top 5%). Looking forward to Seoul.
Mar 2026
paperOur paper on PlantScience.ai, a live AI co-scientist for plant biology, accepted to Molecular Plant (Cell Press, IF 24.1). Joint work with Prof. Ding's group (John Innes Centre / The Sainsbury Lab) and Prof. Zhang's group (Northeast Normal University). Try the live agent.
Dec 2025
grantAwarded a Modal Research Grant to support work on genomic FM evaluation.
Nov 2025
paperOur paper on factual evaluation of medical LLMs with knowledge graphs accepted as an Oral in the AAAI 2026 AI for Social Impact (AISI) Track.
Sep 2025
paperBig congrats to lead author Mingyu Huang — our work on landscape features for biological fitness benchmarks selected as NeurIPS 2025 Datasets and Benchmarks Spotlight (top 2.2%).
Research themes
three threads that I keep braiding together.
— 02
LLM agents for scientific discovery
Knowledge-graph-grounded agents combining structured retrieval with multi-hop reasoning — medicine, plant science, literature.
— 03
Trustworthy & robust ML
Stress-testing modern ML through adversarial evaluation and multi-objective optimization — code intelligence and language models.
Publications
* indicates equal contribution.
Software open source
tools I built or co-led.
OmniGenBench
COLA-Laboratory/OmniGenBench
Modular benchmarking platform for genomic foundation models — RNA structure, gene function, multi-species generalization.
EvalXAI
COLA-Laboratory/EvalXAI
Interpretability toolkit for genomic FMs — attention viz, attribution methods, motif overlap, counterfactual perturbations.
PlantScience.ai
COLA-Laboratory/autoSKG
LLM-powered virtual scientist for plant biology, backed by a 4M+ entity / 10M+ edge knowledge graph from 11 databases.
EMOC
COLA-Laboratory/EMOC
C++ framework for multi-objective evolutionary optimization: 30+ algorithms, 80+ benchmarks, cross-platform GUI. Co-led development.