AI for biology · Biological Foundation Models · Reinforcement Learning

Building AI that
learns from life.

I’m Lanqing Li, an incoming postdoc at Stanford Medicine and a researcher working at the intersection of machine learning, biomedicine, and scientific discovery. I build AI systems that are principled, trustworthy and aligned with scientific reality.

RESEARCH CONSTELLATION23°N  113°E
Research constellation Connected nodes representing protein design, trustworthy learning, and scientific discovery. SCIENTIFIC DISCOVERY REINFORCEMENT LEARNING AI AGENTS REPRESENTATION LEARNING TRUSTWORTHY AI FOUNDATION MODELS AI for Science
01Where learning
meets science
↗
AI for Biology · Trustworthy ML · Learning to ActSCROLL TO DISCOVER
( 01 )THE THROUGHLINE

A scientist’s
curiosity.

My work revolves around a central question: how can we make learning systems more reliable and actionable when real-world data are complex, sparse, and biased?

From offline meta-reinforcement learning to protein engineering, I build AI models that are generalizable, controllable and interpretable, with the ultimate goal to accelerate scientific discovery, making translational impact especially in biomedicine.

( 02 )RESEARCH AREAS

Three lenses,
one practice.

I work across disciplines to develop next-generation AI systems to empower scientific discovery and invetion.

01 / 03

AI for Biology

Designing and aligning biological foundation models for therapeutic discovery, protein engineering, and delivery.

Explore the projects ↗
02 / 03

Trustworthy ML

Building learning systems with reliable representations, robust objectives, and aligned with human/scientific prior.

Explore the projects ↗
03 / 03

Learning to Act

Developing reinforcement learning and AI agents that can generalize beyond tasks and narrow data distributions.

Explore the projects ↗
( 03 )SELECTED WORK

Ideas that
travel.

A few projects that show how the questions evolved—from decision making to molecular modeling.

Conceptual reward-guided protein language model landscape A conceptual, nonquantitative wireframe landscape with an intrinsic reward path in lavender and an extrinsic reward path in peach. Both paths converge at a luminous point, suggesting label-free guidance for a protein language model. The letters M A K L V G below are an illustrative amino-acid sequence, not an experimental result. INTRINSIC REWARD EXTRINSIC REWARD PROTEIN LM M A K L V G LANGUAGE MODEL × RL01LABEL-FREE FINE-TUNINGFOR PROTEIN LANGUAGE MODELSOriginal conceptual SVG illustration. Scientific source: https://arxiv.org/abs/2606.18961. Not a measured molecular structure or quantitative plot.
2027 · PreprintNew manuscript

Label-Free Fine-Tuning

A post-training method that leverages intrinsic and extrinsic reward signal—without requiring labeled preference data.

Read the paper ↗
DeepDelivery — symmetry-inspired protein nanocage and mRNA cargo A conceptual human-derived protein nanocage shown as a hollow three-dimensional shell of repeated folded slate-blue ribbon subunits. Pentagonal and hexagonal pores reveal a dimmer rear shell and a single continuous coral mRNA cargo strand within the cavity, partly occluded by the front protein layer. The architecture is inspired by published symmetric nanocages, not a measured DeepDelivery molecular structure or a quantitative diagram. HUMAN-DERIVED SCAFFOLD mRNA CARGO AI × PROTEIN DISCOVERY02HUMAN-DERIVED SCAFFOLDSFOR mRNA DELIVERYOriginal conceptual SVG illustration for DeepDelivery. Symmetry and porous-shell architecture inspired by Fig. 2 of "Hierarchical design of pseudosymmetric protein nanocages", Nature; DOI: 10.1038/s41586-024-08360-6. The paper's bacterial-derived designed assemblies are used only as visual architectural inspiration and are not represented as human-derived proteins or DeepDelivery data. Repeated folded ribbon motifs, relative scale, cargo placement, and subunit arrangement are illustrative. Not a measured molecular structure, an experimentally established DeepDelivery assembly, or a quantitative plot. Existing DeepDelivery context source: https://www.nature.com/articles/s41467-026-79039-x.
2026 · Nature CommunicationsIn vitro validated

DeepDelivery

AI-driven discovery and engineering of human-derived protein scaffolds for mRNA delivery. The platform surfaced 12 of 15 novel materials with delivery activity in vitro.

Read the paper ↗
ERAM — an enzyme as a learned molecular relation Artistic scientific schematic of a compact generic enzyme protein ribbon, with three shaded alpha helices and five adjacent arrow-shaped beta strands connected by narrow loops, above schematic substrate and product graphs. A restrained amber connection represents the enzyme-mediated relation, while a faint constellation below suggests alignment in a shared molecular and knowledge-graph representation. The protein fold is illustrative and does not depict a real PDB structure; the chemical graphs and embedding positions are not a reaction formula or measured result. ENZYME SUBSTRATE PRODUCT SHARED REPRESENTATION ENZYME × REPRESENTATION LEARNING03MULTIMODAL RELATIONAL LEARNINGFOR ENZYMATIC REACTIONSOriginal conceptual illustration for Accurate and Task-Agnostic Modeling of Enzymatic Reactions through Multimodal Relational Learning. Source: https://pmc.ncbi.nlm.nih.gov/articles/PMC13366380/. Shapes and curves are schematic and do not encode measured experimental data.
2026 · Acta Pharmaceutica Sinica BEditor’s Highlight

ERAM

A task-agnostic framework that connects substrates, enzymes, and products in a shared molecular representation space. Multimodal relational learning supports enzyme retrieval, substrate prediction, and binding-site analysis.

Read the paper ↗
UNICORN — information from offline experience Conceptual scientific illustration. Three flowing bundles of offline trajectories enter a suspended latent constellation. Distinct task clusters and the mutual-information objective I(Z; M) evoke informative task representations, with a branching policy at the right. This is explanatory artwork, not a measured result or a literal architecture diagram. OFFLINE CONTEXT TASK REPRESENTATION POLICY I(Z; M) OFFLINE META-RL04AN INFORMATION-THEORETIC FRAMEWORKFOR DECISION MAKINGOriginal conceptual SVG illustration. Scientific source: https://proceedings.neurips.cc/paper_files/paper/2024/hash/8a30aba6514b56d02976f49797f6338a-Abstract-Conference.html. Not a measured molecular structure or quantitative plot.
2024 · NeurIPS SpotlightTop 2.47%

UNICORN

An information-theoretic framework that unifies context-based offline meta-RL and paves a principled path towards foundation models for decision making.

Read the paper ↗
CC-SAM — class-conditional neighborhoods in a broad loss basin Conceptual, nonquantitative illustration of class-conditional sharpness-aware minimization for long-tailed recognition. An unnumbered descending class-frequency histogram feeds into an oblique topographic loss basin. Three class points have different dashed parameter-perturbation neighborhoods within a broad, flat region. Sizes do not encode measured values or a claimed ordering by class frequency. This illustrates the first-stage class-conditioned training idea, not the complete two-stage architecture or an experimental result. LONG-TAILED CLASSES CLASS-CONDITIONAL SCALE FLATTER MINIMA ROBUST OPTIMIZATION05CLASS-CONDITIONAL OPTIMIZATIONFOR LONG-TAILED RECOGNITIONOriginal conceptual illustration for Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition. Source: https://openaccess.thecvf.com/content/CVPR2023/html/Zhou_Class-Conditional_Sharpness-Aware_Minimization_for_Deep_Long-Tailed_Recognition_CVPR_2023_paper.html. Shapes and curves are schematic and do not encode measured experimental data.
2023 · CVPRLong-tailed recognition

CC-SAM

A robust optimization method for long-tailed visual recognition. PAC-Bayesian analysis motives class-conditional parameter perturbations that encourage flatter minima in the loss landscape.

Read the paper ↗
( 04 )PUBLICATIONS

Selected
writing.

Peer-reviewed work, current manuscripts, and the ideas in between.

Google Scholar ↗
01

Be Your Own Teacher: Steering Protein Language Models via Unsupervised Reward Optimization

Lanqing Li†, Shentong Mo, Yang Yu, Pheng-Ann Heng

Preprint2027
↗
02

AI-Driven Discovery and Engineering of Human-Derived Protein Scaffolds for mRNA Delivery

Dongmin Xun*, Lanqing Li*†, et al.

Nature Communications2026
↗
03

Accurate and Task-Agnostic Modeling of Enzymatic Reactions through Multimodal Relational Learning

Yuansheng Huang*, Lanqing Li*, et al.

Acta Pharmaceutica Sinica B2026 · Editor’s Highlight
↗
04

Unified and Explainable Molecular Representation Learning for Imperfectly Annotated Data from the Hypergraph View

Bowen Wang, Junyou Li, Donghao Zhou, Lanqing Li, et al.

Nature Communications2025
↗
05

Scrutinize What We Ignore: Reining In Task Representation Shift Of Context-Based Offline Meta Reinforcement Learning

Hai Zhang, Boyuan Zheng, Tianying Ji, et al. Lanqing Li†

ICLR2025
↗
06

Towards an Information Theoretic Framework of Context-based Offline Meta-Reinforcement Learning

Lanqing Li*, Hai Zhang*, et al.

NeurIPS Spotlight2024 · Top 2.47%
↗
07

Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition

Zhipeng Zhou*, Lanqing Li*, Peilin Zhao, Pheng-Ann Heng, Wei Gong

CVPR2023
↗
08

On the Pitfall of Mixup for Uncertainty Calibration

Dengbao Wang, Lanqing Li†, Peilin Zhao, Pheng-Ann Heng, Minling Zhang

CVPR2023
↗
09

ImGCL: Revisiting Graph Contrastive Learning on Imbalanced Node Classification

Liang Zeng, Lanqing Li†, Ziqi Gao, Peilin Zhao, Jian Li

AAAI2022
↗
10

Local Augmentation for Graph Neural Networks

Sontao Liu, Rex Ying, Hanze Dong, Lanqing Li†, et al.

ICML2022
↗
11

FOCAL: Efficient Fully-Offline Meta-Reinforcement Learning via Distance Metric Learning and Behavior Regularization

Lanqing Li, Rui Yang, Dijun Luo

ICLR · Seminal Paper on Offline Meta-RL2021
↗
( 05 )THE JOURNEY

A computer scientist and physicist
by training.

Physics taught me to seek nature’s underlying structure and principles. Machine learning gave me the tools to test them.

I’m an incoming postdoctoral scholar in Bioengineering at Stanford Medicine, working with Prof. Le Cong and Prof. Mengdi Wang (Princeton). Before that, I led AIDD research at Zhejiang Lab and Tencent AI Lab. I was fortunate to complete my PhD under the supervision of Prof. Pheng-Ann Heng at CUHK.

Read the full CV ↗
2026—now

Stanford Medicine

Postdoctoral Scholar · Bioengineering

2023—2026

Zhejiang Lab

Principal Investigator · Life Sciences Computing

2019—2023

Tencent AI Lab

Senior Research Scientist

2018—2019

Infervision

Tech Lead & Machine Learning Engineer

EDUCATION

Ph.D. CSE · CUHK 2026

EDUCATION

M.S. Physics · UChicago 2017

EDUCATION

B.S. Physics · MIT 2015

HONOR

International Physics Olympiad Gold Medal· 2011

( 06 )CONTACT

Have a research
problem?

I’m always interested in thoughtful conversations about AI for science/biology, post-training and recursive self-improvement of scientific foundation models, reinforcement learning and what comes next. Research collaborations and inquiries are welcome. We are also hiring remote research interns and visiting students, please feel free to drop an email.

llq1993@stanford.edu ↗ lanqingli1993@gmail.com ↗