Bo Li 李波

Bo Li 李波

Ph.D. candidate · University of Macau
Visiting Ph.D. student · National University of Singapore

Bo Li 李波 — Virtual Cells, Intervention Design & Self-Improving Models

Scientific AIVirtual-Cell World Models · Closed-Loop Design · Self-Improving Models
From Virtual Cells to Self-Improving Scientific AI

I study how AI systems learn from observations, guide scientific interventions, and improve through new evidence. My research spans multimodal cellular modeling, genetic and drug perturbation prediction, measurement-aware evaluation, and intervention design. In parallel, I investigate how pretrained models can acquire, retain, and accumulate new capabilities through weight-space programming. My long-term goal is to develop adaptive scientific AI systems that connect observation, experimentation, and continual learning to support better scientific decisions.

University of Macau (advisor Bob Zhang; co-advisor Qianqian Song) Visiting National University of Singapore (host Yang Zhang)
Open to postdoctoral and faculty opportunities starting in 2027, spanning virtual-cell world models, scientific intervention design, and adaptive AI.
View CV Email

News

2026.11 🧠 WeightCode released its paper, code, data and executable weight programs. Code · Paper
2026.11 🔁 WeightLoop preprint released on arXiv: weight-space programming driven by experimental feedback (co-author).
2026.10 🎉 MVCBench was accepted in principle by Nature Communications. Project · Preprint
2026.09 🎯 Released VCDesign, a framework for finite-budget intervention design in virtual cells. Project · Code
2026.09 🧬 Released PertResolve, using split-half reproducibility to distinguish measurement and model limitations in perturbation prediction. Project · Code
2026.09 💊 Released DrugDis, a component-resolved framework for interpreting general and context-specific drug-response prediction. Project · Code · Data
2026.08–09 📘 Three collaborative papers published: incomplete multi-view multi-label learning in IJCV and IEEE TPAMI, and PSSD for histology-to-gene-expression prediction in Bioinformatics.
2026.06 🇸🇬 Started a one-year visit to the National University of Singapore, hosted by Prof. Yang Zhang.
2026.05 🤖 CellScientist preprint released on arXiv (co-first author).
2026.04 🧫 MVCBench preprint released on bioRxiv.
2026.01 🎉 One paper accepted at ICLR 2026 (co-author).
2025.12 🎉 PhenoProfiler published in Nature Communications.
2025.12 🧬 Collaborative spatial-omics paper HAST published in Communications Biology.
2025.08 🧬 SpaIM published in Nature Communications.
Earlier news
2024.10 📄 HGGEP published in Briefings in Bioinformatics.
2024.08 🧬 Collaborative paper AntiFormer published in Briefings in Bioinformatics for antibody binding-affinity prediction.
2024.08 🎓 Started my Ph.D. at the University of Macau.
2024.07 🏅 Graduated from Beijing University of Technology as a Top 100 Graduate and Beijing Outstanding Graduate.
2024.03–06 🔬 Published MHFAN, Lite-UNet, and EDTM across Pattern Recognition and EAAI.

Research

Research loop in five parts: (a) cellular state learned from image, omics and spatial measurements; (b) a virtual cell predicts image and expression responses to drug and gene perturbations; (c) evaluation of which predicted differences measurements can resolve; (d) intervention design that ranks candidates by their predicted shift from a source state to a target state and tests the top-K within budget B; experiments on the selected candidates give measured outcomes, and (e) self-improving models (WeightCode, WeightLoop, CellScientist) turn them into lasting model updates that return to the virtual cell.
From Multimodal Virtual Cells to Self-Improving Models. (a) Learn cellular state from multimodal measurements. (b) Predict responses to genetic and drug perturbations. (c) Assess which distinctions the measurements can reproducibly resolve. (d) Select interventions within a finite budget. (e) Turn measured outcomes into lasting model updates that feed the next round of prediction.
  1. (a) · CELLULAR STATE

    What representations of cellular state can be learned from imaging and omics measurements?

    PhenoProfiler SpaIM HGGEP
  2. (b) · RESPONSE PREDICTION

    Do learned cell-state transitions under intervention generalize across modalities, compounds and contexts?

    MVCBench
  3. (c) · EVALUATION & RESOLUTION

    When does a predicted gain reflect biology that the assay can reproducibly resolve?

    PertResolve DrugDis MorphoSuff
  4. (d) · EXPERIMENTAL CHOICECurrent focus

    Planning with the model: given a target cellular state and a finite budget, which interventions are worth testing next?

    VCDesign
  5. (e) · SELF-IMPROVING MODELS

    Closing the loop: can measured outcomes keep adding capabilities to a model, and do the gains improve its next decisions?

    WeightCode WeightLoop CellScientist
Next directions
  • Updatable cell world models. Turn new assay results into persistent, transferable model updates, not retraining.
  • Does capability growth pay off? Test whether each update improves intervention choices on held-out experiments.
  • Accumulating capabilities. Reusable model edits that add perturbation and cell-context skills without forgetting.

Selected Publications

† Equal contribution  ·  * Corresponding author

Published & in press

(b) Response Prediction · Multimodal Benchmark
MVCBench: benchmarking drug-molecular and gene representations for drug-induced virtual cell phenotypes

MVCBench: A multimodal benchmark for drug-induced virtual cell phenotypes

Bo Li†, Qing Wang†, Shihang Wang†, Bob Zhang*, Yuzhong Peng, Pinxian Zeng, Chengliang Liu, Mengran Li, Ziyang Tang, Xiaojun Yao, Chuxia Deng, Qianqian Song*

Nature Communications In press (2026)  ·  Project  ·  Preprint  ·  Code  ·  Data  · 

TL;DR: Benchmarks image- and expression-based perturbation prediction under compound, cell-line, plate and dataset shifts.

(a) Cellular State · Image-based Profiling
PhenoProfiler: end-to-end phenotypic profiling of high-content cell images

PhenoProfiler: Advancing phenotypic learning for image-based drug discovery

Bo Li, Bob Zhang*, Chengyang Zhang, Minghao Zhou, Weiliang Huang, Shihang Wang, Qing Wang, Mengran Li, Yong Zhang, Qianqian Song*

Nature Communications 17, 793 (2026)  ·  Paper  ·  Webserver  ·  Code  ·  arXiv  · 

TL;DR: Learns phenotypic profiles directly from high-content microscopy images, without hand-crafted feature-extraction pipelines, for compound profiling.

(a) Cellular State · Spatial Omics
SpaIM: style-transfer imputation for single-cell spatial transcriptomics

SpaIM: Single-cell spatial transcriptomics imputation via style transfer

Bo Li†, Ziyang Tang†, Aishwarya Budhkar, Xiang Liu, Tonglin Zhang, Baijian Yang*, Jing Su*, Qianqian Song*

Nature Communications 16, 7861 (2025)  ·  Paper  ·  Code  ·  Data  · 

TL;DR: Treats imputation of unmeasured genes in spatial transcriptomics as style transfer from single-cell references.

Current work · under review and preprints

(e) Self-improving Models · Weight-Space Programming
WeightCode: executable weight-space programs install new capabilities into trained networks, rerun on other networks and support successive updates

WeightCode: Installing new capabilities into neural networks via weight-space programming

Bo Li†, Taoyong Cui†, Bob Zhang*, Mengran Li, Chengyang Zhang, Yang Zhang, Pheng Ann Heng, Wanli Ouyang, Zhenfei Yin*, Yingcheng Wu*, Ling Yang*

Under review 2026  ·  Paper  ·  Code  ·  Data  · 

TL;DR: Executable weight programs install capabilities without gradient training: SCAN exact match 0%→99.8%, and BabyAI pickup success 0%→98.4% from eight demonstrations. Replaying programs on five independently trained ViTs retains 83–110% of source-network gains, and subsequent programs add capabilities while preserving earlier ones.

(d) Experimental Choice · Intervention Design
VCDesign: finite-budget intervention design for virtual cells

VCDesign: Finite-budget intervention design for virtual cells

Bo Li†, Lin Wang†, Bob Zhang*, Mengran Li, Zhenchao Tang, Chengyang Zhang, Minghao Sun, Chengliang Liu, Zhiyuan Liu, Yang Zhang*

Under review 2026  ·  Project  ·  Manuscript  ·  Code  ·  Data  · 

TL;DR: Ranks candidate interventions toward a target state under a fixed budget, scores the selected set against independently measured held-out outcomes, and transfers evidence to candidates with no measured response.

(c) Evaluation & Resolution · Measurement Limits
PertResolve: measurement resolution for fine-grained perturbation prediction

Distinguishing measurement and model limitations in perturbation prediction

Bo Li†, Chengyang Zhang†, Mengran Li†, Bob Zhang*, Lin Wang, Zhenchao Tang, Jun Liu, Chengliang Liu, Chen Wei, Yuhao Yi*, Jiancheng Lv, Yang Zhang*

Under review 2026  ·  Project  ·  Manuscript  ·  Code  ·  Data  · 

TL;DR: Uses split-half reproducibility across 470 coding-variant conditions and 31 public perturbation configurations to separate detection, identification and held-out response prediction, exposing measurement limits in TP53/KRAS and a model gap in JAK1.

(c) Evaluation & Resolution · Drug-response Components
DrugDis: disentangling general and context-specific effects in drug-response prediction

DrugDis: Disentangling general and context-specific effects in drug-response prediction

Bo Li†, Chengliang Liu†, Yuzhong Peng†, Bob Zhang*, Qing Wang, Pinxian Zeng, Mengran Li, Shenghui Huang, Chuxia Deng*, Yang Zhang*

Under review 2026  ·  Project  ·  Manuscript  ·  Supplement  ·  Code  ·  Data  · 

TL;DR: Decomposes drug-response prediction into general drug, general sample and drug–sample interaction effects, showing that aggregate accuracy can hide weak recovery of the context-specific part.

Other co-first and co-corresponding work: PopRetrieve (when single-cell response distributions improve drug retrieval; co-corresponding author, under review) and MorphoSuff (when label-free morphology suffices for targeted cellular measurements; co-first author, bioRxiv preprint).
Full publication record on Google Scholar ↗

Software & Resources

Selected research software and open-science resources.

Project What it is Stars
WeightCode Executable weight programs for capability installation, reuse and accumulation 10
VCDesign Finite-budget intervention design and decision-aligned evaluation 5
MVCBench Multimodal benchmark for drug-induced virtual-cell phenotypes 10
PhenoProfiler Cellular phenotypic learning from high-content microscopy 11
SpaIM Spatial-transcriptomics imputation using single-cell references 24
Nature-Paper-Skills Agent skills for scientific writing, revision and review 671
DrugClaw Agentic RAG for drug research 118
Awesome Virtual Cell Virtual-cell literature, resources and interactive research map 337
Awesome Weight-Space Programming Weight-space methods, programmable models and research landscape 8

Stars as of Nov 13, 2026; open each project for current counts.

Education

  • 2026.06 – 2027.06: National University of Singapore

    - Visiting Ph.D. Student. Host: Prof. Yang Zhang

  • 2024.08 – Present (expected 2027.06): University of Macau

    - Ph.D. in Computer Science, Department of Artificial Intelligence, Full Scholarship. Advisor: Prof. Bob Zhang; Co-advisor: Prof. Qianqian Song, Purdue University

  • 2021.09 – 2024.07: Beijing University of Technology

    - M.Eng. in Electronic Information. Advisors: Prof. Yong Zhang, Prof. Baocai Yin

  • 2017.09 – 2021.07: Beijing Information Science & Technology University

    - B.Eng. in Robotics Engineering. Advisor: Prof. Hongbo Huang

Selected Honors & Awards

  • 2024: Ph.D. Full Scholarship, University of Macau
  • 2024.07: Top 100 Graduates, Beijing University of Technology (top 100 of 6,331)
  • 2024.07 & 2021.07: Beijing Outstanding Graduate
  • 2023.10: Xiaomi Scholarship
  • 2022.10: National Scholarship, Ministry of Education of China
  • 2022, 2023: First-Class Academic Scholarship, Beijing University of Technology
  • 2021: First Prize, Science & Technology Innovation Scholarship, Beijing Information Science & Technology University

Patents

Co-inventor of two granted Chinese invention patents on cell localization: CN115457547B (2026) and CN115810046B (2026).

Academic Service

Journal reviewer: Science Advances, IEEE TPAMI, IEEE TIP, IEEE TNNLS, Medical Image Analysis, Bioinformatics, Briefings in Bioinformatics, Pattern Recognition, BMC Biology, Engineering Applications of Artificial Intelligence, and Expert Systems with Applications.

Collaborations: Purdue University, Cornell University, University of Florida, National University of Singapore, Sun Yat-sen University, Sichuan University, Beijing University of Technology, and Macao Polytechnic University.

Beijing Information Science & Technology University Beijing University of Technology University of Macau West China Hospital, Sichuan University National University of Singapore