Session 3.3 · Live demo · Eric Ma

The IRED-88 Research Copilot.

One enzyme, four kinds of data, and one question ladder. A deep mutational scan, a crystal structure, a structure prediction, and a small literature knowledge base answer one scientific question together, in seven notebooks you can read here or run yourself.

Start with the question Live notebooks running on our server: nothing to install · Static copy if the live one is slow

The question

Where do activity-improving mutations of the enzyme IRED-88 come from, and could we have seen them coming by combining DMS data, structure, prediction, and literature?

IRED-88 is an imine reductase from an enzyme engineering campaign at Novartis. The campaign measured 4,720 single mutants (81% of every possible one), the enantioselectivity of 310 engineered variants, and a 1.9 Å crystal structure. Around that data sits a knowledge base of six paper notes, built the way John Damask showed in 3.2. Here the knowledge base is not only read by an agent: it is used inside the analysis, next to the data.

The question ladder

Each notebook answers one question; together they answer the big one. Every claim traces back to a cell, a data file, or a knowledge base note. Each card opens the live notebook, which runs fresh when you open it, so give it a few seconds. Read-only copies of all seven are in the static snapshots.

Start hereThe overarching questionThe enzyme, the four kinds of data, and how the notebooks fit together.00 Q1 · DMSWhich mutations improve activity?The activity map: which single mutations help, and where they sit along the sequence.01 Q2 · StructureAre the hits where the structure says?Beneficial positions against their distance to the NADP cofactor in PDB 7OG3.02 Q3 · LiteratureWhat does prior work already know?The top hits, checked against six paper notes in the knowledge base.03 Q4 · PredictionWhat does prediction see that the crystal cannot?ESMFold fills in the residues the crystal leaves out, with a confidence score.04 Q5 · CombinationsAdditive or epistatic?Doubles and triples of the top mutations: does the knowledge base’s claim survive the data?05 Q6 · EverythingWould we have noticed a hole?Hide part of the scan where the literature matters, and see what the other data recover.06

Four kinds of data, four jobs

Data	What it is good at
Deep mutational scan and enantioselectivity	Discovery: finding hits nobody guessed
Crystal structure (PDB 7OG3)	Mechanism: ruling out bad guesses
Knowledge base (six papers)	Explanation: naming what worked, and why
Structure prediction (ESMFold)	Blind spots: reaching what the crystal misses, with a confidence score

Run it yourself

The live notebooks above let you read the code and use the widgets, but not change the code. To run and edit the notebooks, you need uv and git. In a terminal:

Get the code and the data

git clone https://github.com/UMMS-Biocore/ired-88-research-copilot
cd ired-88-research-copilot
uv run scripts/fetch_data.py

The last line downloads the paper’s data from Figshare and the structure from the PDB, and checks every file.

Open the notebooks

uv run --with marimo marimo edit notebooks/

Your browser opens with the notebook list. notebooks/live/ holds empty versions with the same questions and no code, if you want to build the analysis yourself.

Or hand it to the agent you set up in 3.0: open the folder in it and ask it to read README.md and AGENTS.md, then run a notebook and explain the result.

Credit

Made by Eric Ma for this bootcamp; the source is ericmjl/ired-88-research-copilot, mirrored at UMMS-Biocore. The data are public, from the supporting information of the paper below, and the deep mutational scan is Eric’s own work from his time at Novartis. Cite the paper for any use beyond teaching.

Ma EJ, Siirola E, Moore C, et al. Machine-directed evolution of an imine reductase for activity and stereoselectivity. ACS Catal. 2021;11(20):12433–12445. doi:10.1021/acscatal.1c02786. Crystal structure: PDB 7OG3.