Work
Each entry gives the dates, my role and what AI did, the limits of the result, and links to evidence you can check. Where nothing is public, the entry says so.
AmadeusBio.ai
The company’s research systems are not public. What is public is listed here.
Oct 2026
Molecular Animation
An open framework that turns a scientific brief into a molecular film with Blender. Every film comes with a production record that states, beat by beat, what is experimental and what is illustrative. The channelrhodopsin film took about two hours from idea to final render.
Scenes are built from PDB structures by a tested, pure-NumPy science layer: mmCIF assemblies, atom-matched states, constrained morphs that keep bond lengths and angles, pathway profiles, and procedural membranes.
Since v0.2.0, molecules move by physics on one clock per film: Stokes–Einstein diffusion, elastic-network normal modes, side-chain turns, and worm-like-chain DNA.
Builds are idempotent and report their numerical checks. After a refactor, the case study’s freshly rendered posters matched the delivered ones within 1/255.
Case study: channelrhodopsin, the light-gated channel behind the 2026 Nobel Prize in Physiology or Medicine, from PDB 9GO1 and 9GO2. In 21 seconds, a photon twists retinal from all-trans to 13-cis, Lys132 swings toward Glu129 (9.0 → 2.66 Å), and the cation pathway opens. Each beat is tagged experimental, illustrative, or symbolic.
That film took about two hours from idea to final render. By hand, curating assets, animating, adjusting, and iterating on a film like it usually takes days.
Second case study: the ART locus of Staphylococcus phage SA1, a recently reported system of array-associated reverse transcriptases, from GenBank MW218148.1 and AlphaFold 3 predictions. No complex of its RNA, reverse transcriptase, and partner protein was predicted with confidence, so the 31-second film keeps them apart and marks how they work together as unknown.
The channelrhodopsin film: the light-gated channel behind the 2026 Nobel Prize in Physiology or Medicine, built from PDB 9GO1 and 9GO2, with each beat labelled on screen. It was released two days after the prize was announced. The membrane is procedural, not simulated. Describe the film
A 21-second silent film of channelrhodopsin C1C2, built from PDB structures 9GO1 and 9GO2. It opens on the protein dimer in a lipid membrane as photons arrive, dives into the retinal pocket, and shows retinal twisting from all-trans to 13-cis, an illustrative morph between the two crystal states. Lys132 then swings toward Glu129 to form a salt bridge. A section through the cation pathway first shows it interrupted, as in the crystal, then open in an illustrative conducting-state model as sodium ions cross the membrane. Labels on screen mark each beat as experimental, illustrative, or symbolic.
Role. I lead the project at AmadeusBio.ai. It is built with AI agents and designed so that an agent (Claude Code through Blender MCP) can drive it, though it also works by hand.
Limits. No outside lab has used it yet. The two-hour figure is my own timing of one film. In the channelrhodopsin film, the retinal twist is a morph between two crystal states, and the open pathway is an illustrative model rather than a deposited structure.
May 2026
Bioinformatics plugins for coding agents
The first AmadeusBio.ai tools: a Claude Code plugin that operates PyMOL in natural language and retrieves sequences and structures from databases such as UniProt, and a plugin for Galaxy bioinformatics workflows.
These plugins came before the current research systems, which are not public. They remain available as historical work.
Role. I built them with AI coding agents, building on skills from the open SciAgent-Skills collection.
Independent projects
Personal work, on my own GitHub.
Jul–Sep 2026
Human-gated music agents
Claude Code skills that compose and arrange electric-guitar tablature. The model writes the tab, a validator parses and lints it, and I audition every chunk before it is accepted.
The arranger keeps the melody and root motion of the source piece, enforced by a fidelity check, and reports every other departure as information rather than as an error.
A companion skill produces commit-pinned, evidence-linked implementation plans for public GitHub repositories, for handoff to a coding agent.
Role. Built with Claude Code. No arrangement is final until I have listened to it.
Wet-lab research
Undergraduate research at UC San Diego. The data belong to the labs and are unpublished.
2025
Neuronal morphology genetics in C. elegans
Undergraduate research on genes that maintain the shape of touch-receptor neurons, using genetic crosses, PCR and Sanger genotyping, and fluorescence microscopy. First-author poster at the 2025 International Worm Meeting.
Poster: Yangyi Liu, Yixiao Xiao, Ting Zhao, Seungmee Park, Yishi Jin. “LNKN-1 as a Novel Regulator of Neuronal Morphologies in C. elegans.” International Worm Meeting, 2025.
Role. Supervised by Yishi Jin, with mentors Seungmee Park, Ting Zhao, and Yixiao Xiao.
Limits. The underlying data are unpublished, so there is no public link.
2024
A plant transcription factor in engineered yeast
Expressed the Arabidopsis transcription factor BBX31 in campesterol-producing Saccharomyces cerevisiae and measured its effects on growth, stress tolerance, and cell size. First-author poster at the 2024 ASCB Annual Meeting.
Poster: Yangyi Liu, Zixu Chen, Zechun Wang, Alan R. Zhang, Ethan E. Chin, Kaibiao Wang, Veronica Gomez-Godinez, Linda Shi, Yanran Li. “Investigating the Function of Arabidopsis thaliana Transcription Factor BBX31 in Saccharomyces cerevisiae.” ASCB Annual Meeting, 2024.
Four of the coauthors were high-school interns in UC San Diego’s OPALS program, where I was a research mentor in the summer of 2024.
The lab’s main project was strigolactone synthetic biology, and my Eleanor Roosevelt College Honors Project came from it: protein expression in E. coli and yeast consortia, ELISA with purified D14 strigolactone receptors, and LC-MS.
Role. Supervised by Yanran Li.
Limits. The ASCB poster is not online.