Building the digital twin of HER2 — the protein first, everything else second
HER2T is an autonomous research platform with a single focus: HER2/ERBB2, a protein that drives several of the most aggressive cancers. Our mission is to build the most complete, honest digital twin of HER2 — by reading the entire literature, finding where the science disagrees, and deeply modeling how the protein actually behaves. Drug-testing and experimental validation build on top of that foundation. The twin comes first.
HER2T Laboratories
Twin first — Computational & Experimental Research for HER2
Think of it as one research institution, not a pile of separate tools. HER2T is the mind — the digital twin and autonomous research system that asks the scientific questions and absorbs every answer. HER2T Laboratories is where those questions get tested, across two labs:
Computational Lab
Live · ExpandingThe reading and reasoning is live today: it reads the HER2 literature, extracts and formalizes claims, finds where the science disagrees, and forms testable predictions. The Compute Lab runs molecular dynamics on a GPU node — the simulations that build the twin’s understanding of how HER2 moves and mutates. It is currently paused under a safety freeze, so no simulations are running. Drug-docking comes next, once the twin is solid enough to trust those results. Stages that need it are shown as explicitly blocked, never faked.
Experimental Lab
Partner · FutureWhere the strongest predictions get tested against reality. HER2T has no in-house wet lab yet, so physical experiments go to qualified partner laboratories today; a dedicated HER2T laboratory is the long-term goal. No physical experiment has been run yet, and reality has the final say.
The loop in one breath
HER2T predicts → the Compute Lab simulates → a partner laboratory measures → HER2T learns. Then it asks a sharper question, and the loop tightens. (A dedicated HER2T lab is the goal; partner laboratories carry the physical testing until then.)
The HER2T Compute Lab is intentionally paused under a safety freeze. No simulations are being sent to any GPU node.
No molecular-docking engine is attached. Docking stages stay blocked, never faked.
No FEP / binding-free-energy engine is attached. Those stages are not run.
Physical experiments go to qualified partner laboratories today; a dedicated HER2T lab is the goal. No physical experiment has been run yet.
The problem we are solving
There are roughly 70,000 published papers on HER2/ERBB2, and more arrive every week. No human team can hold all of that in their head at once. Crucial findings sit buried in papers nobody has connected; studies quietly contradict each other; and the open questions that matter most get lost in the sheer volume. The knowledge to make progress often already exists — it is just scattered across more literature than any group can read.
HER2T attacks that bottleneck directly: it reads at machine scale, connects findings across papers, and surfaces exactly where the field disagrees and what to test next.
The plan: Understand → Simulate → Discover
The platform is built in three stages, and the site is always honest about which one is actually live today:
UNDERSTAND — live today
Read and organize everything the field already knows about HER2, and find where it conflicts with itself.
SIMULATE — paused · safety freeze
Model how HER2 actually moves, folds, and responds to mutations. The HER2T Compute Lab (GPU molecular dynamics) is built but currently paused under a safety freeze. Drug-docking is the natural next capability once the twin’s foundation is solid — stages that need it are shown as blocked, never faked.
DISCOVER — planned
Propose genuinely new, testable predictions and drive them toward real-world validation.
You can follow this live on the Journey page.
How the work actually flows
The heart of the platform is a discovery pipeline — a team of specialized AI agents that works a research question in sequence. You can watch it happen live on the Discovery Console:
A Chief Scientist agent breaks your question into specific, answerable sub-questions.
Five domain specialists — Literature, Genomics, Protein, Drug, and Pathway — investigate those questions in parallel, each reporting real findings and open gaps.
A Hypothesis Synthesis agent combines everything the specialists found into candidate ideas.
A Skeptic agent attacks those ideas and throws out the ones too weak to survive.
An Evidence agent independently grades each surviving idea against what the published literature actually supports.
An Experiment Designer turns the ideas that held up into concrete experiments a lab could really run.
Alongside this runs a literature belt: it mines individual papers into clean, testable claims, then collides claims from different papers against each other. Where two papers genuinely disagree, that contradiction becomes a ranked open question — and can seed a brand-new hypothesis.
The data we create — and why it is the real asset
Every step above writes a permanent, traceable record. That accumulating dataset is the real value being built here. Concretely, the platform is producing:
- Formalized claims — extracted from papers, each tagged with exactly where it came from, so nothing is a floating assertion.
- Contradictions — a catalog of precisely where the HER2 literature disagrees with itself.
- Hypotheses — the platform’s own ideas — each ranked, skeptic-reviewed, and evidence-graded.
- Experiment designs — concrete, runnable tests for the ideas that held up.
- A knowledge graph — all of this connected into a map of how findings relate to each other.
- A prediction ledger — frozen, timestamped predictions that can later be checked against reality — this is what keeps the system honest over time.
- An open-questions agenda — everything HER2T does not yet know, ranked by how much fresh disagreement the literature shows.
What we do with that data
1. Turn it into a prioritized research agenda.
The contradiction-driven agenda and triage system sort every open question by how much genuine conflict exists and what can actually be done about it right now — execute now, recover missing inputs, wait for compute, hand to a wet lab, or send to a human expert.
2. Package the strongest leads for expert review.
The best unresolved contradictions — after passing skeptic, evidence, and novelty checks — become dossiers a human specialist can grade. That is the bridge from “AI found something interesting” to “a real scientist should look at this.”
3. Track each question up an honest ladder.
The lifecycle view follows any question from unknown → investigated → explained → computed → decisive experiment → actually measured → model updated. Each rung it reaches cites a real record; each rung it has not reached states plainly what is still missing.
4. Show all of it publicly and live.
The activity feed, the What HER2T Knows scoreboard, and the Scale of the Work page all draw from this same real data — no secrets, no invented numbers — so anyone can see exactly how much of the field we have processed and what it produced.
The honesty rule
HER2T never pretends to know something it has not actually worked out. Any step that would need real lab equipment or a compute backend we have not yet connected (for example molecular docking) is shown as explicitly blocked, not faked — and the Compute Lab's GPU molecular dynamics is currently paused under a safety freeze. The AI does the reading, reasoning, contradiction-finding, and experiment design — it does not claim to have run experiments it cannot run.
Right now HER2T has processed a small fraction of the ~70,000-paper field. That is exactly why the Scale page exists — to make the honest case for going faster.