The Mission

Aiming AI at the diseases we can’t yet predict.

Honest, verifiable contributions to understanding Alzheimer’s and ALS — built in the open, with calibrated uncertainty as the rule.

Read the build log → See the work so far
This work is personal — why →
Why this can work

A built-in answer key.

There’s a corner of this science where a laptop is genuinely enough to make a checkable contribution that stands on its own. When a mutation changes a protein, scientists measure the effect in the lab — and many of those measurements are public. That measurement is an answer key. If a model is wrong, the data says so immediately. No lab required. That built-in check is what makes a careful solo effort do real work here instead of just writing reports.

The personal why →

The role of AI

AI does the work. The answer key keeps it honest.

Something changed in the last few years: AI can now do the work — propose models, generate and test variants, work through the biology step by step, not just summarize it. That shift is what makes it possible now for one person to take a serious run at this.

But it’s trustworthy in exactly one place: where it can be caught being wrong. So the rule is strict — AI is aimed only at questions with a built-in verifier, a lab measurement or a held-out outcome, and nothing it produces is claimed until that answer key confirms it. On problems that can be checked, generative AI does real work. Aimed anywhere else, it’s confident-sounding noise, and this work doesn’t go there.

Why aim AI this way — the vision →

The work so far

A first result, honestly stated.

Every number below is regenerated by a script from checksum-locked public data — never typed by hand. If one ever drifts, it’s visible. Full detail, number by number, is in the build log.

0.82

Reproduced the field’s model

A recent, strong model (CANYA) predicts protein aggregation from sequence. Rerun on its own held-out data, it scored an AUROC of 0.82 — matching the paper’s 0.809. The pipeline is faithful, and there’s a concrete baseline to improve on.

~6x

Fixed its overconfidence

A model can rank well and still misrepresent how sure it is. This one’s calibration error was 0.068; a standard recalibration cut it roughly six-fold, to 0.011, with its ranking preserved. Calibrated uncertainty is the point, not a footnote.

0.62vs 0.20

The disease result

Alzheimer’s has a handful of inherited mutations that cause the early-onset form. Trained on Aβ data and tested on 468 held-out single mutations it had never seen, a purpose-built model’s correlation with the lab measurements was 0.62 — versus 0.20 for the general model applied cold. On telling the inherited mutations apart, it reached the same discrimination as the lab assay itself (~0.90).

↕

Then a wall worth hitting

Pointing what works on Alzheimer’s Aβ at the ALS protein TDP-43 didn’t just fail — it failed backwards: the relationship between aggregation and disease flips sign between the two proteins (harmful in one, protective in the other). That isn’t a bug — it’s consistent with known biology, and it emerged from the pipeline rather than being put in by hand. Which means there are years of genuine open problems here, not a weekend’s worth.

What I will not claim
  • I didn’t “beat” the measurement — matching ~0.90 is within its noise. “Matched the ceiling,” not “surpassed it.”
  • The headline rests on only 8 inherited mutations, so the result I actually trust is the 0.62 vs 0.20 correlation across 468 points.
  • It’s not a fair fight, and I’ll say so: my model was trained on Aβ data; the general one wasn’t. The honest claim is “a protein-specific model far outperforms a transferred general one,” not “my architecture is better.”
  • The method is standard. The contribution is doing it carefully, honestly, and in the open — not inventing something new.
  • The data and the comprehensive molecular picture belong to the Lehner/Bolognesi lab (Aβ/TDP-43 DMS atlases; Seuma et al., Sci. Adv. 2025). The familial-mutation discrimination result is itself their own published finding (Seuma et al., eLife 2021) — reproduced here, not discovered here. This is a calibrated reproduction and extension on their foundation, not a competing discovery.
— from the build log, “The molecular track”
See the full research →
How it fits together

One spine, several pillars.

Everything orbits the verified work. Everything here is anchored to it — nothing claims more than a checkable result can support.

Live

The Work

The verified molecular results — reproducing, calibrating, and pushing variant-effect models on the disease metric. The clinical prognosis track opens when data access lands.

See the research →
Live

Motivation

Why this exists — two parents, two diseases, and a modeler’s conviction that the right tools, honestly applied, can contribute something that holds up.

Read the why →
Live

Methods

The right-fidelity discipline underneath all of it — match the model to the decision, quantify what you don’t know, never ship false certainty.

The methodology →
Starting

Build Log

The work documented in public as it happens — including what I won’t claim. The honest record is the credibility.

Read the first entry →
Preview

Health-AI Watch

Tracking how frontier AI is really doing on clinical and biological reasoning — rigorously, not by press release. A live preview is up now.

See the preview →
On the roadmap — what’s coming
Planned

For Families

A plain-language version for patients and caregivers — but only once a forecaster has earned it on the answer key. No family-facing tool ships until the numbers say it’s honest enough.

What this will be →
Build Log

Watch the work happen.

I’m doing this in the open — the results and the caveats both. Chapter 01 walks through twelve investigations, what they found, and why I’m pivoting.

Read the build log →
Where this is going

Aimed at the diseases that take who we are.

It begins with Alzheimer’s and ALS, but the longer arc is calibrated, honest prognosis — and, over years, the wider family of these diseases. Aimed only where a measurement can prove it wrong.

Read the vision →
Get involved

This is built in the open. Here’s how to be part of it.

This work is personal — here’s the why. There are several ways to connect, depending on where you stand.

Right now this mission is one person — me, Michael Key. I’m building toward a board and collaborators. If you’re a clinician, a researcher, or someone affected by ALS or Alzheimer’s who wants to help — reach out.

Follow the work

The build log is a running record of every result, every caveat, and every dead end as it happens. No press releases. No polished summaries. Just the honest record of what the science is actually saying, week by week.

Read the build log →

If you research or practice

Clinicians, neuroscientists, biostatisticians: your scrutiny is welcome. If you see a flaw in the method, I want to hear it. If you’re working on the clinical side of Alzheimer’s or ALS and want to talk about collaboration — especially on the prognosis track — reach out. How data is stored, accessed, and protected is laid out in data stewardship.

michael@rightfidelity.ai →

If your family has been affected

You’re why this exists. Families living with Alzheimer’s or ALS deserve honest tools, not confident-sounding predictions with hidden assumptions. You can follow along through the build log, or reach out directly — I’d be glad to hear from you.

Reach out →

Support the mission

A nonprofit structure to house this work long-term is forming. If you’d like to help fund or anchor it — financially or institutionally — I’d welcome that conversation. This isn’t a commercial project. The goal is research that families can actually trust.

Get in touch →