Oncofusion

start / graph / ana-q204-failure-mode-atlas-90-modes-over-110-layer-by-mechanism-cells-and-the-eight-modes-the-corpus-has-no-entry-for

Analysisweak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.king

A failure-mode atlas for drugging the KMT2A::AFF1 axis: 90 modes over a 10 x 11 grid of layers and mechanisms, 46 cells argued not applicable, and eight modes the corpus has no entry for

also called failure mode atlas · what can go wrong atlas · pre-mortem atlas KMT2A::AFF1

We made a checklist of the ways treating this leukaemia can go wrong, built so that it cannot quietly skip a category: ten places something can fail (the drug, the target protein, the proteins around it, the DNA packaging, the RNA, the cell, the cancer cell population, the body, the clinic, and our own knowledge) crossed with eleven ways it can fail (a relative steps in, another route takes over, the target changes, the drug lets go, it creates something new, it harms something else, the population evolves, the cell changes identity, it springs back, other drugs interfere, or we mismeasure). That gives 90 specific failure modes. Eight have nothing about them anywhere in our existing entries. The list is only as complete as the grid, and the 46 cells we ruled out are claims that could be wrong.

The same thing, for a scientist

Atlas built as ten layers (drug, target molecule, network, chromatin/locus, RNA, cell state, clone, tissue/organism, clinical system, knowledge) by eleven mechanisms (compensation, bypass, target escape, state dynamics, gain of function, loss of other function, selection/induction, plasticity, reversal/memory, interactions, mis-measurement), with a meta level 0-4 on each mode (the failure; failure of the countermeasure; non-detection; failure of our model; wrong decision frame). 64 of 110 cells hold a mode; 46 carry a written reason with a pointer to the mode that covers the nearest real concern; the validator asserts that every cell is accounted for. Corpus coverage was screened by regex and then read: the keyword hit counts are noisy (broad terms matched unrelated entries; one regex mistook 'bivalent chromatin' for bivalent binding) and are not coverage. After reading, eight modes have no corpus entry on the mechanism: RNA-protein sensing leading to mutation (FM-041), drug-induced mutagenesis (FM-061), autoinduction of clearance (FM-102), tissue depots (FM-103), cis-regulatory escape at HOXA/MEIS1 (FM-113), drug-induced aberrant splicing (FM-119), clonal cooperation (FM-128), and publication bias hiding resistance (FM-142). Molecular rebinding (FM-001) is partial. Two analyses were run on the first examples (see the linked entries).

Go deeper
How it is built and what the grid can and cannot guarantee

Each mode has a layer, a mechanism, a meta level, a generic description, the specific form it takes here, a regex used to find corpus coverage, the evidence class (patient, model, inference, speculative) as the project's own classification, what existing data could test it, what a model could generate, a design response, what to prepare if it is inevitable, and ordinal severity, likelihood and detectability priors used only for ranking 1. The grid makes the enumeration exhaustive over layers x mechanisms, not over biology: a failure that fits no cell, or fits one nobody filled, is invisible. The 46 not-applicable cells are claims with reasons and pointers; if an event cannot be mapped to a cell, that is a finding and the atlas is extended 1.

The five worked examples from the request

A similar protein compensates (FM-020, FM-021) is covered by the paralog analysis entry. Off-target and on-target neo-interaction (FM-004, FM-005, FM-023) are partly in the corpus (JUND liberation is the documented on-target case) and open for chemical off-target and neo-surface. Dissociation then rebinding or mutation (FM-001, FM-010, FM-060) is covered for mutation by the reachability entry and open for kinetics. RNA-protein sensing leading to mutation (FM-041) is speculative with no support anywhere in the corpus; the entry records which test could discriminate it (genome-wide mutation spectra before and after drug, or a fluctuation test) and that the observed escape variants cannot, because they are selected 1.

Facts and figures
scriptanalyses/q204-failure-mode-atlas/atlas.py
modes90
layers10
mechanisms11
cells110
cells with a mode64
cells argued not applicable46
unaccounted cells0
corpus nodes at run3128
modes with no corpus entry
[
  "FM-041",
  "FM-061",
  "FM-102",
  "FM-103",
  "FM-113",
  "FM-119",
  "FM-128",
  "FM-142"
]
partial
[
  "FM-001"
]
How this could bite us back
  • A long list can look like preparedness. The risk is that the 90 modes are read as complete and the unmapped failure is the one that happens; the atlas states its own grid limit for that reason.
  • The severity, likelihood and detectability priors are my ordinal judgements and will mis-rank if used as probabilities. They are for ordering the reading, not for dosing or trial decisions.
Still open
  1. Which of the 46 not-applicable cells are wrong? Each is a one-line claim; a reviewer who knows a mechanism that fits one should add it.
  2. Can an adverse event or relapse sample in the literature be mapped onto a cell without forcing it? Cases that cannot be mapped extend the atlas.
  3. Which eight no-entry modes merit an experiment first, given that FM-041 and FM-061 are the hardest to detect?
Every source behind this entry (1)
  1. E1weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.database

    The atlas holds 90 failure modes over 64 of 110 layer-by-mechanism cells and 46 argued not-applicable cells with no cell unaccounted for; after reading the nearest corpus hits, eight modes have no corpus entry on the mechanism.

    What this evidence cannot show

    The enumeration is mine; the evidence classes and ordinal priors are unverified judgements except where a corpus entry or an analysis supports a mode. The corpus screen is a regex followed by reading of the nearest hits, so a mode judged absent may be covered under wording the regex missed (this happened twice in this analysis and was caught by reading).

    Evidence that argues against this (1)
    • Several regexes were too broad or too narrow in the first pass: they mistook a similar word for the mechanism and missed an existing entry whose title used different wording. The coverage claim therefore rests on reading, not on the hit counts.
Connected to