Sequencing phase 2: the order to do things in, ranked by information per unit cost
also called phase-2 running order · what to do Monday · information per dollar
This is the running order for phase 2: the first five things cost almost nothing and several could change what we do next, so they go first. Then come three cheap laboratory experiments that each settle a whole programme, with the expensive work last and five items we recommend not doing at all.
The same thing, for a scientist
The ordering criterion is expected information per unit cost: the probability that a result changes a decision, multiplied by the number of decisions it changes, divided by what it costs. Applied to this portfolio it produces an order that is not the intuitive one. The five highest-ranked actions are all free or nearly free, and three of them are administrative or arithmetic rather than experimental. The single highest-ranked item is the breakpoint prognostic reanalysis, because the sequences already exist, the analysis is desk work, and a positive result would change how infants are risk-stratified - an unusually high ratio of consequence to cost. The second is submitting the St. Jude Cloud data-access application, which costs nothing, takes a median of seven days and unblocks four analyses already written into this graph. Doing that late is the most expensive cheap mistake available.
Tier 0: free or near-free, do these first
1. Breakpoint prognostic reanalysis (ana-interfant-breakpoint-prognostic-reanalysis). The sequences exist in Interfant and Frankfurt hands; the analysis is desk work. A positive result changes risk stratification for every infant with this disease. Nothing else in the portfolio has this ratio.
2. Submit the St. Jude Cloud PanALL and PCGP applications (ana-phase2-dataset-stjude-cloud). Free, median seven days, one in five rejected, and approval is durable at the Data Access Unit level. Four analyses already in this graph are blocked on it. An application submitted when an analysis is ready is a week late, and a month late if refused.
3. Junction-class coverage calculation (target-phase2-fusion-junction-dossier). Pure arithmetic on published breakpoint distributions. Decides whether any junction-directed immunotherapy is a platform or a per-patient product, before a single experiment is funded.
4. ChEMBL row audit for KMT2A (ana-phase2-dataset-chembl, target-phase2-cxxc-dna-dossier). One day. There are 577 potent rows against KMT2A and this review believes none engage the CXXC domain. If that belief is wrong, a feasibility verdict flips.
5. Boltz positive controls, campaigns 1 and 2 (ana-phase2-boltz-cofolding-plan). Negligible credit cost, and every subsequent co-folding result depends on whether these pass.
Tier 1: cheap wet-lab, each settles a whole programme
6. LEDGF IBD displacement assay plus a HOXA9 readout (target-phase2-psip1-ibd-dossier). This is the recommended first laboratory programme. The interface is solved, the hotspots are named, the dependency profile is the best in the graph and the chemistry field is empty. One assay and one qPCR decide whether that emptiness is an opportunity or a graveyard.
7. ENL degrader versus occupancy probe, head to head (target-phase2-enl-yeats-dossier, target-phase2-degrader-modality-dossier). Both reagents exist. The result determines whether ENL is a degrader target or not a target, and it generalises to every reader domain in this interactome.
8. Menin KMT2A-versus-JUND displacement assay pair (target-phase2-menin-mbm1-dossier). Builds the assay the field does not have, and it is the prerequisite for every partner-selectivity claim anyone in this project wants to make.
Tier 2: expensive or slow, fund only after tier 1 reports
9. KAT7 MYST-family counter-screen plus mark-and-programme test. Good chemistry, modest selectivity, no physical contact with the fusion, and a missing knockout phenotype. Worth doing, not worth doing first.
10. DOT1L combination time-to-resistance panel. Restart condition only; the endpoint is time to resistance, never depth of response, because depth was already measured and was insufficient.
11. Compound-specific MEN1 base-editor resistance screen. Replaces the resistance model that cannot be trained. Real cost, real value, and it differentiates between menin inhibitors on a basis nobody currently uses.
12. Immunopeptidomics for a junction peptide. Only if the tier-0 coverage calculation says a junction class covers a useful fraction of infants.
Not funded this phase, and why that is a decision
- CBX8 programme. DepMap mean gene effect is positive, co-dependency is at negative-control level, 15 ChEMBL rows at 120 nM. No-go, with one named reopening condition.
- CXXC orthosteric small molecules. Selectivity against a dozen-member family on a protein-DNA surface. Route to degrader instead.
- Fusion junction small molecules. Verified structural absence. Negative feasibility finding, not a gap.
- A menin degrader intended to defeat resistance. The observed resistance substitutions sit at the warhead's own binding surface.
- Resistance-mutation and patient-response prediction models. Six and roughly thirty-five labels respectively. Not trainable; the entries say so with the counts attached.
Declining five things is the part of a readiness review that saves the most money, and it is the part that is hardest to publish.
Why this order and not the intuitive one
The intuitive order starts with the most exciting target and buys reagents. This order starts with an administrative form and a spreadsheet, because both have a high probability of changing several downstream decisions and cost essentially nothing.
The formal criterion is expected information per unit cost. Three of the top five items are not experiments, which is an uncomfortable but honest output: at this stage the portfolio's biggest constraints are data access and unexamined assumptions, not laboratory capacity. The wet-lab work that does appear early was chosen because a single readout terminates a programme in either direction, which is the property the go/no-go register was written to enforce.
| criterion | expected information per unit cost = P(result changes a decision) x (number of decisions changed) / cost |
|---|---|
| uncomfortable output | Three of the top five actions are not experiments. At this stage the portfolio's binding constraints are data access and unexamined assumptions, not laboratory capacity. |
tier 0 free — the numbers behind the chart
[
{
"rank": 1,
"action": "Breakpoint prognostic reanalysis",
"node": "ana-interfant-breakpoint-prognostic-reanalysis",
"cost": "desk work on sequences that already exist",
"why_first": "a positive result changes risk stratification for every infant with this disease"
},
{
"rank": 2,
"action": "Submit St. Jude Cloud PanALL and PCGP data-access applications",
"node": "ana-phase2-dataset-stjude-cloud",
"cost": "free, median 7 days",
"why": "unblocks four analyses already written into this graph; approval is durable"
},
{
"rank": 3,
"action": "Junction-class coverage calculation",
"node": "target-phase2-fusion-junction-dossier",
"cost": "arithmetic on published breakpoint distributions",
"why": "decides platform versus per-patient before any immunotherapy spend"
},
{
"rank": 4,
"action": "ChEMBL row audit for KMT2A (577 rows)",
"node": "target-phase2-cxxc-dna-dossier",
"cost": "one day",
"why": "could flip the CXXC feasibility verdict"
},
{
"rank": 5,
"action": "Boltz positive controls, campaigns 1 and 2",
"node": "ana-phase2-boltz-cofolding-plan",
"cost": "negligible credits",
"why": "every downstream co-folding result depends on these passing"
}
]tier 1 cheap wetlab — the numbers behind the chart
[
{
"rank": 6,
"action": "LEDGF IBD displacement assay plus HOXA9 readout",
"node": "target-phase2-psip1-ibd-dossier",
"why": "best dependency profile in the graph against an empty chemistry field; one assay and one qPCR settle it"
},
{
"rank": 7,
"action": "ENL degrader versus occupancy probe head to head",
"node": "target-phase2-enl-yeats-dossier",
"why": "both reagents exist; generalises to every reader domain in the interactome"
},
{
"rank": 8,
"action": "Menin KMT2A-versus-JUND displacement assay pair",
"node": "target-phase2-menin-mbm1-dossier",
"why": "builds the assay the field lacks; prerequisite for every partner-selectivity claim"
}
]tier 2 expensive — the numbers behind the chart
[
{
"rank": 9,
"action": "KAT7 MYST counter-screen plus mark-and-programme test",
"node": "target-phase2-kat7-dossier"
},
{
"rank": 10,
"action": "DOT1L combination time-to-resistance panel",
"node": "target-phase2-dot1l-dossier",
"constraint": "endpoint is time to resistance, never depth of response"
},
{
"rank": 11,
"action": "Compound-specific MEN1 base-editor resistance screen",
"node": "ana-phase2-model-resistance-mutation",
"why": "replaces a model that cannot be trained"
},
{
"rank": 12,
"action": "Immunopeptidomics for a junction peptide",
"node": "target-phase2-fusion-junction-dossier",
"precondition": "only if the tier-0 coverage calculation is favourable"
}
]not funded
[
{
"item": "CBX8 programme",
"reason": "positive DepMap gene effect, negative-control-level co-dependency, 15 ChEMBL rows at 120 nM",
"node": "target-phase2-cbx8-dossier"
},
{
"item": "CXXC orthosteric small molecules",
"reason": "protein-DNA surface, dozen-member family selectivity problem",
"node": "target-phase2-cxxc-dna-dossier"
},
{
"item": "Fusion junction small molecules",
"reason": "verified structural absence; negative feasibility finding",
"node": "target-phase2-fusion-junction-dossier"
},
{
"item": "Menin degrader intended to defeat resistance",
"reason": "the observed resistance substitutions sit at the warhead's own binding surface",
"node": "target-phase2-degrader-modality-dossier"
},
{
"item": "Resistance-mutation and response-prediction models",
"reason": "six and roughly thirty-five labels respectively",
"node": "ana-phase2-model-resistance-mutation"
}
]- An ordering by expected information assumes the probabilities are roughly right. Every one of them is a judgement, and the most confident judgement here - that the LEDGF interface is an opportunity rather than a graveyard - is the one with the least direct evidence.
- Front-loading free desk work can become a way of never starting the experiments. Each tier-0 item needs a date, not just a rank.
- Ranking by information per cost systematically favours questions that are cheap to ask over questions that matter most. The breakpoint reanalysis ranks first partly because it is cheap; a reader should check that it also ranks first on consequence, which this review believes it does.
- Does the lab actually have access to the Interfant or Frankfurt breakpoint sequences, or is the top-ranked action blocked on a collaboration that has not been asked for?
- What is the real cost of the LEDGF displacement assay in this lab's hands, including recombinant IBD production, and does that move it out of tier 1?
- Should the ordering be re-derived after every tier-0 result, given that each of the five could change a downstream decision?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Controlled-access approval at St. Jude Cloud takes a median of seven days with an 18.8% rejection rate and is granted durably at the Data Access Unit level, which is why submitting early dominates submitting when needed.
St. Jude Cloud: A Pediatric Cancer Genomic Data-Sharing Ecosystem · McLeod C, Gout AM, Zhou X, et al. · Cancer Discovery 2021What this evidence cannot show
2021 figures; current turnaround and rejection rate are unverified, and a rejection would require a different plan for the four blocked analyses.
Evidence that argues against this (1)
- A median seven-day approval with an 18.8% rejection rate describes the process for successful and unsuccessful applicants together, and the rejection rate is the planning-relevant number: nearly one in five applications does not proceed, which an information-per-cost ordering should price as schedule risk rather than as a flat latency. Cancer Discovery 2021
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The LEDGF IBD-KMT2A interface is solved with named aromatic hotspots and a genetic requirement for transformation, yet ChEMBL holds only three potent rows against PSIP1, all from antiviral chemistry - the combination of high dependency and absent chemistry that puts this programme first among the wet-lab items.
Validation and structural characterization of the LEDGF/p75-MLL interface as a new target for the treatment of MLL-dependent leukemia · Cermakova K, Tesina P, Demeulemeester J, et al. · Cancer Research 2014What this evidence cannot show
The functional evidence is murine MLL-AF9; no published work demonstrates that blocking this interface pharmacologically collapses the programme in a human KMT2A::AFF1 cell, which is precisely the gap the tier-1 experiment fills.
Evidence that argues against this (1)
- Twelve years without follow-on oncology chemistry is weak evidence that the interface is harder than it looks, and the tier-1 experiment is designed to find that out in months rather than years. source 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
DepMap per-cell-line data, on which the no-go and prioritisation decisions in this ordering rest, is reachable through the Open Targets GraphQL API and has already been used in two executed analyses in this project, one of which overturned an earlier claim made in this same graph.
Open Targets Platform GraphQL API, depMapEssentiality field, as used in analyses/depmap-lineage-conditioned · Open Targets / DepMap 2026What this evidence cannot show
DepMap release version unverified; the dataset is structurally blind to leukaemia-initiating-cell and fetal-cell-state requirements, which is the standing caveat on every prioritisation decision in this ordering.
Evidence that argues against this (1)
- The no-go decisions resting on per-line DepMap inherit the lineage confound this project has since documented. An ordering built on cost-per-bit is only as good as whether the bits answer the question, and for this disease the panel composition means some of them do not. Oncofusion knowledge graph (internal analysis) 2026
ranked first: desk work on sequences that already exist, changing risk stratification if positive
ranked second: free, median seven days, unblocks four analyses
the recommended first wet-lab programme
- supports →Phase-2 dossier: CBX8 - the candidate this review would kill, and the reasons are all numeric
not funded this phase; the no-go is a decision, not an omission
the ordering is only enforceable because each item has a pre-registered stopping rule
- depends on →Compute: which backend for which piece of phase-2 work, and the one constraint that matters for each
each ranked item is costed against the backend it would run on
the top-ranked action is the test of this hypothesis