Discovering cancer's hidden rules from raw data.
A general intelligence, and a new way to find cures.
Given only a table of raw cancer measurements — about 18,000 genes switched off across 1,100 cancer cell lines, the public DepMap screen — with no labels, no expert features, and no hints, Heather rediscovered known cancer biology, redrew the cell's internal map, and proposed roughly 1,900 gene-combination leads to investigate.
It pulled out textbook machinery it was never told about, and two of its highest-ranked drug-combination predictions are real targets already in human trials. Heather reached all of this blind, before anyone looked anything up, which is how we know what it saw was real structure and not noise. This report is two things at once: evidence that a general intelligence can read a disease's own rules from raw data, and a first look at what that could mean for finding cures.
Heather is AGI
Heather is Aphorion's general intelligence. It perceives raw information directly, discovers the hidden laws inside it, recalls by content, and knows when something is genuinely new, abstaining rather than guessing. It learns from a single example and does not forget. The same machine that reads a cancer screen reads a physics trajectory or a market. It is a general instrument, and cancer is where its first proof and its largest stakes happen to coincide.
The two AGI claims this proves
A general intelligence rests on abilities most systems fake with hand-built machinery. Two of them Heather has, and the cancer result proves both, because in cancer the right answers are already known and each claim can be graded against reality.
Raw perception
An intelligence must perceive raw information directly, with no person first engineering the features for it. Most AI skips this: a vision model is handed pixels only after human-designed preprocessing, a language model is handed tokens rather than raw text. Heather takes the raw stream, here the unprocessed gene-effect table, and finds what matters in it with nothing supplied.
Hidden-law discovery
Storing information is not understanding it. The leap is to recover the rules that generated the observations: the quantities that stay fixed and the things that move together. From the raw cancer numbers Heather recovered the laws of the cell. It is the same operation that, given raw motion, recovers a conservation law in physics.
The future of cancer research
Cancer is not solved because the search is vast and slow. Thousands of genes, hundreds of tumour contexts, an astronomical number of possible combinations, and every promising pair costs months at the bench. Progress is gated less by ideas than by how quickly hypotheses can be generated, triaged, and trusted.
A general intelligence that reads raw data and discovers its laws changes that arithmetic. Point it at every screen, every measurement type, every tumour, and it returns the couplings worth testing — ranked, traceable to the data, and stripped of the obvious confounders — in minutes rather than years. The two axes it rediscovered here are already in the clinic; the roughly 1,900 it proposed are waiting. On data no one has yet mined, the same instrument could surface the combinations no one has seen.
This is what the future of cancer research could look like: not a sharper microscope, but an intelligence that reads the disease's own rules and says where to strike. Dataset by dataset, that may be how cancer is finally solved.
The data and the signal
The DepMap screen knocks out each of about 18,000 genes, one at a time, across 1,100 human cancer cell lines, and records how much each knockout slows or kills each line. A gene's entire description, to Heather, is its column: the vector of fitness costs across all 1,100 lines. No gene name carries meaning, no pathway is annotated, no feature is computed.
Two genes are co-dependent when their columns move together: lines that need gene A also need gene B. Co-dependence is the footprint of shared function — the two genes sit in the same molecular machine, the same pathway, or in a synthetic-lethal relationship where the loss of one forces reliance on the other. Heather measures it as the correlation of the two raw columns, computed for every pair of genes.
The discovery procedure
A bare correlation is cheap and mostly spurious. A co-dependency is reported only when it satisfies every condition below. Each is a property the link must hold, and each removes a distinct class of false positive.
| Condition | What it requires | What it removes |
|---|---|---|
| Coupling | raw correlation ≥ 0.30 | weak, incidental pairs |
| Unrecorded | absent from the STRING interaction database | links already known, isolating candidates |
| Reproducible | holds on a held-out 70 / 30 split of the cell lines | flukes of the particular sample |
| Context-invariant | within-tissue correlation keeps ≥ 55% of strength | tissue-of-origin artifacts |
| Mutual | each gene sits in the other's top neighbours | one-sided hub effects |
| Cross-module | the two genes' neighbour sets do not overlap | trivial within-complex pairs |
| Active | both genes expressed in the lines carrying the link | couplings on genes a cell never switches on |
The last condition is mechanistic rather than statistical: a cell cannot depend on a gene it does not transcribe, so a link between two silent genes is noise and is discarded, however strong the correlation. "Absent from the database" flags candidates but does not certify them — the database is incomplete, so a link can be unrecorded yet real. Final novelty is settled only by a literature review, then by experiment.
Discover blind, then confirm
Heather is run with no biological input and returns ranked co-dependencies. Only afterward is each top result checked against the published literature. The order is the safeguard: the answer is found before any annotation is consulted, so agreement with the textbooks cannot be circular. The largest source of false co-dependency is tissue of origin — two genes can correlate simply because both are essential in, say, melanoma. The control measures correlation within each tissue and requires the link to survive. Its trustworthiness is established on two pairs whose answers are already known.
| Pair | Raw r | Within-tissue | Verdict |
|---|---|---|---|
| BRAF · MAPK1 — same pathway | +0.679 | +0.573 | survives (real) |
| BRAF · MITF — melanoma lineage | +0.347 | +0.164 | collapses (artifact) |
A known functional pair keeps most of its strength; a known lineage artifact loses most of it. The control passes the true link and rejects the false one, which is why its verdicts on the novel links can be trusted.
The two confirmed axes
Of the links Heather ranked highest, two were checked exhaustively against the literature. Both are real, both are druggable, and both were reached blind. They are not new biology; they are the proof that the perception is veridical, and the template for the leads still to be tested.
METTL14 · MRE11 — RNA tagging meets DNA repair
Heather's strongest novel signal places METTL14, the partner subunit of the m6A RNA-methylation writer, with the DNA-repair nuclease MRE11. The internal check is immediate: METTL14's single nearest neighbour is METTL3, the enzyme it is physically bound to (+0.58), exactly what the biology demands. The link reproduces on held-out lines (train +0.52, test +0.54), and under the tissue control it does not weaken at all (within-tissue +0.534 against a raw +0.523), cleaner than the BRAF–MAPK1 positive control. The literature confirms the axis, and a METTL3 inhibitor (STC-15) is in Phase 2 trials — which turns a METTL3-plus-PARP combination into a concrete, testable hypothesis.
BRD9 · KAT2A — remodeller meets acetyltransferase
The second flagship places BRD9, a subunit of the ncBAF chromatin-remodelling complex, with KAT2A, the catalytic enzyme of the SAGA acetyltransferase complex. BRD9 is a validated dependency: synovial sarcoma becomes addicted to ncBAF, and BRD9 degraders are already in trials. The link survives the tissue control (within-tissue +0.31 from a raw +0.38) and is reciprocal: asked from the other side, KAT2A's own top co-dependants include BRD9 and the BAF subunit SMARCD1.
Field validation
We shared the METTL14 finding with researchers in medicine and pharmaceutical R&D. The reading from the bench matches what Heather found from the numbers alone — and turns it into a concrete drug strategy.
The correlations make sense. They tell us two parts are involved — METTL14 and METTL3 — where the former depends on the latter to synergistically deactivate a critical step in DNA repair, which is mediated by RNA. We already have a lead compound, STC-15, that inhibits METTL3, which essentially renders METTL14 of no effect because it needs its counterpart. That is why the report shows their activities, with respect to this specific pathway, so highly positively correlated. So it makes sense to select one target — METTL3, an enzyme in this case — and attack it with a molecule like STC-15 to limit its activity and automatically inhibit METTL14, just because we know they are highly correlated. It's very relevant medically. We also have PARP inhibitors that could be added to the mix to increase the lethal load. It's good and practical, and could help speed up computational drug discovery.
The clinical reading also explains the strength of the signal. Because a single molecule acting on METTL3 collapses the activity of METTL14 as well — the partner cannot function without it — the two genes' dependencies track each other almost exactly, which is the correlation Heather ranked first. One target, one drug, two genes silenced; a PARP inhibitor stacks a second hit on the same DNA-repair vulnerability. This is one expert reading, to be confirmed with others, but it places the finding squarely in practical, druggable territory.
The cell's architecture, recovered
The two confirmed axes are single links. The deeper result is structural. Apply the same correlation operation to Heather's own output, treating each recovered machine as a new object, and a hierarchy appears: 18,000 genes condense into 58 molecular machines, and those into 8 functional systems. The machines are recognisable protein complexes — SWI/SNF, SAGA, the Fanconi-anemia repair group, ATP synthase — recovered with no annotation. The 8 systems are the cell's job description: making energy, copying and repairing DNA, building proteins, processing genetic messages, holding shape, signalling, moving cargo, and dividing. No single rule dominates: the largest pattern, roughly "how fast the cell grows," explains only about a tenth of the variation. What results is close to the organisational chart of a living cell, assembled from raw numbers by one repeated operation.
Two machine-produced outputs, each with its own page describing what it is and the actual data it contains.
Co-dependency leaderboard
The ranked gene-pair co-dependencies that cleared every gate yet are absent from the interaction databases — the combination-target leads, with the actual output.
View output →Dataset · the cell mapLaws of the laws
The recovered hierarchy — 18,000 genes into 58 molecular machines and 8 functional systems — and the module-to-module laws between complexes.
View output →The honest catch
"Needed together" is not "a cure." A reported pair means two genes matter to the same cells. Whether hitting both does better than hitting one is a laboratory question that data alone cannot answer. The vision of solving cancer is exactly that, a vision; this report proves a capability, not a treatment.
The output is bounded by the input. The test ran on public data the whole field has studied for years, so most of what surfaced was already known. That is what makes it a clean proof of the two claims, and it caps how novel the output can be on this dataset. The genuinely new biology, and the cures, would come from data not yet mined.
The novelty lives in the data; the capability lives in Heather. The same two abilities — raw perception and hidden-law discovery — carry to any field that produces data to perceive.
Bottom line
Heather perceives raw information and discovers the laws inside it, with no hints and no neural networks. On cancer it rediscovered real, drugged biology and redrew the cell's map, proving a general intelligence and, in the same stroke, a new way to search for cures. The proof is here; the cures are what the next datasets are for.
Reproduce — input: public DepMap 24Q4 release (CRISPRGeneEffect, Model, OmicsExpression) and the STRING interaction database; all public. Confirmations — Wainberg et al., Nat Genet 2021 (co-essential modules); Michel et al., Nat Cell Biol 2018 (ncBAF / BRD9 / synovial sarcoma); eLife 2021 and Nature Cancer 2022 (m6A, MRE11, DNA-repair, PARP); STORM Therapeutics, ASCO 2024 (METTL3 inhibitor STC-15 in trials).