AI for Science
Methods, benchmarks, and reproducible software releases.
Methods
Benchmarks
Packages
AI for Science · Science for AI
Zeyu Fu
Ph.D. Candidate · methods, systems, and research tooling
I work on two axes: AI for Science (methods, benchmarks, and research software) and Science for AI (harnesses, orchestration, telemetry, and evaluation). Domain papers are listed as published case work.
13
peer-reviewed papers highlighted on this homepage
10
published software projects with code or package distribution
7
public sites and tools linked from this homepage
Start here
See all tools
Linked pages
Use SCPortal for broad browsing, benchmark pages for detailed comparisons, and project pages for methods, datasets, or companion materials.
Hub
Primary Route
SCPortal
Canonical public discovery hub for datasets, benchmarks, models, and related tools.
Open SCPortal →
Utility
Primary Route
mRNA Intersection
Interactive mRNA intersection analysis and visualization tool.
Launch →
Microsite
Secondary Route
LAIOR Benchmarks
Focused benchmark destination for model summaries, datasets, and metric detail pages.
Open LAIOR Benchmarks →
Public Surface
Secondary Route
iAODE Pages
Public-facing iAODE destination for datasets, explorer pages, and project overview content.
Open iAODE Pages →
Benchmark
Secondary Route
scCCVGBen
Centroid-coupled VAE benchmark across 200 single-cell datasets (100 scRNA + 100 scATAC) with per-dataset metadata cards.
Open scCCVGBen →
Microsite
Secondary Route
GAHIB Companion Site
Companion site for GAHIB, the graph-attention information-bottleneck framework, with method overview and online figures.
Open GAHIB Companion Site →
Documentation Atlas
Secondary Route
scCCVGBen Atlas
Reference atlas for scCCVGBen datasets, methods, metrics, and the companion application.
Open scCCVGBen Atlas →
Suggested route
Start broad, then open the detail you need
A simple path from the homepage overview to broader browsing and then benchmark-specific detail.
Overview
Browse
Compare
01
02
Homepage: overview
Research themes, publication highlights, profile links, and released tools in one place.
SCPortal: browse the catalog
Compare datasets, models, benchmarks, and related tools across the public work.
Open SCPortal → 03LAIOR: compare benchmarks
Model, dataset, and metric pages when you need benchmark-specific detail.
Open benchmarks →About
Direct statement of work: AI for Science and Science for AI.
I work on two axes. AI for Science: machine learning methods, benchmarks, and open software. Science for AI: harnesses, orchestration, telemetry, and evaluation. Single-cell and other domain papers are published case work, not the full scope of the page.
AI for ScienceScience for AIMethodsBenchmarksAgent HarnessesOrchestrationTelemetryOpen Releases
On this site
AI for Science
Published methods, benchmarks, packages, and companion sites.
Science for AI
Agent harness work across Claude Code, Codex, Grok Build, Antigravity, OpenCode, and the oh-my-* series: skills, plugins, runtime loops, telemetry. Upstream contributor to oh-my-openagent (33 merged PRs on team mode, tmux subagent runtime, and runtime fallback; ranked 11th of 313 contributors by commits as of Sep 2026).
Published case work
Earlier single-cell papers remain listed as published case work.
Profiles
GitHub, ORCID, Scopus, and Web of Science for code, citations, and authorship tracking.
Research focus
Match focus to software
Three direct lines
Methods work, agent-harness work, and published domain papers.
Science for AI
Harnesses, orchestration, telemetry, and evaluation for research and engineering agents.
Harness
Orchestration
Telemetry
Published case work
Single-cell and other domain papers are listed as published case work.
Case work
Domain papers
13 papers
Publications
# Equal contribution * Corresponding author 13 peer-reviewed papers
1
Islands and bridges: Momentum contrastive coupling unifies discrete and continuous structure in single-cell omics
Fu, Z.#,*, Chen, C.#, Zhang, K. · Biomedical Signal Processing and Control 122, 110376 · 2026
2
CCVGAE: A centroid-coupled variational graph attention autoencoder for stable and interpretable single-cell representation learning
Fu, Z.#,*, Liu, Y.#, Wang, J., Wang, S. · Array 30, 100808 · 2026
3
CLOP-DiT: Structured-metadata-conditioned single-cell latent generation via contrastive language-omics pretraining and Diffusion Transformers
Fu, Z.#,*, Liu, Y.#, Wang, J.*, Wang, S.* · Array 30, 100934 · 2026
Open-Source Software
10 published tools and code releases (7 on PyPI)
MCCVAE
BSPC · 2026
Momentum contrastive coupling for discrete and continuous structure in single-cell omics.
CCVGAE
Array · 2026
Centroid-coupled graph attention VAE for stable, interpretable single-cell embeddings.
CLOP-DiT
Array · 2026
Structured metadata-conditioned latent generation via language-omics pretraining and DiT.
iAODE
pip install iaode
Neural ODE-VAE for scATAC-seq benchmarking and continuum modeling.
iVAE
pip install iVAE
Interpretable VAE enhancing clustering for single-cell data.
LiVAE
pip install livae
Lorentz-regularized VAE for transcriptomic and epigenomic embeddings.
GNODEVAE
pip install gnodevae
Graph ODE-VAE with GNN-enhanced clustering and dynamics.
scRL
pip install scrl-fatedecision
RL for evaluating cell fate decisions in single-cell data.
scFocus
pip install scfocus
SAC-based branching probability detection for lineage focusing.
CODE
pip install scCODE
Correlated latent space learning for single-cell RNA continuum modeling.
Locate-then-edit
Open the map
Maps and measurement sites
Public sites for key geometry, edit federation, noise swamping, and prospective admission. Science names only.
Edit federation map
Gain-screened two-regime map of rank-one merge interference.
Map →Key geometry
Closed-form key-cosine / S×C vs locate-then-edit collateral.
Site →Noise swamping
INT8 / NF4 damage-rank survival and base-model noise.
Site →Prospective admission
Geometry-ordered merge admission; mixed predicates, no cross-model rule.
Site →
Released tools
Public tools
Public pages for browsing datasets, checking benchmarks, reading companion materials, and using focused utilities.
Browse
Benchmarks
Utilities
Start here
Benchmark and companion pages
LAIOR Benchmarks
SecondaryMicrosite
Focused benchmark destination for model summaries, datasets, and metric detail pages.
Open LAIOR Benchmarks →iAODE Pages
SecondaryPublic Surface
Public-facing iAODE destination for datasets, explorer pages, and project overview content.
Open iAODE Pages →scCCVGBen
SecondaryBenchmark
Centroid-coupled VAE benchmark across 200 single-cell datasets (100 scRNA + 100 scATAC) with per-dataset metadata cards.
Open scCCVGBen →GAHIB Companion Site
SecondaryMicrosite
Companion site for GAHIB, the graph-attention information-bottleneck framework, with method overview and online figures.
Open GAHIB Companion Site →scCCVGBen Atlas
SecondaryDocumentation Atlas
Reference atlas for scCCVGBen datasets, methods, metrics, and the companion application.
Open scCCVGBen Atlas →More linked pages
Local infrastructure
Protocol
Model Router
Local-only model routing infrastructure; no public browser API.
Read public protocol →
Local-First
Local-First boundary
Kept off the homepage because it is a local-first workspace rather than a public hosted app surface.
Landing-Only
Landing-Only boundary
Maintained as a landing-only reference surface, so the homepage does not present it as an actively browsable public tool.
AI Usage & Stats
Live AI interaction statistics for the Science for AI harness loop behind methods, benchmarks, and pages.
Tokscale AI Profile
Track the AI-assisted build loop across Claude Code, Codex, Grok Build, Antigravity, and OpenCode, coordinated by the oh-my-* harness series.
View Full Stats →Contact
For collaborations, code, profiles, and public research references.