Molecular Diagnostics · Product family
From marker to manufactured kit
A diagnostic is not a finding with a box around it. It is an assay with analytical performance data, a validation study that was powered before it ran, and documentation a reviewer can follow. We build all three, and we manufacture the result.
Detection performance · the only argument a reviewer accepts
- 421
- Span-anchored biomarker records across 184 targets and 86 diseases
- 0.835
- Macro recall against known clinical markers, four tumour types
- 6 of 6
- Published sample-size reference cases reproduced exactly
- 100%
- Provenance coverage - no claim without a traceable source record
Where diagnostic programmes actually fail
Two failure modes, and neither of them is the science
Failure one
The marker was the wrong class
It was selected from literature that could not distinguish a marker predicting drug response from one merely predicting disease course. Those are different products, with different claims and different trials behind them.
DiscoveredDiscovered at the interim analysis, when the enriched arm reads out no differently from the unenriched one.
Failure two
The study was designed after the assay
By the time anyone computed the power, the cohort had been recruited - and it turned out to be underpowered for the claim the sponsor actually wanted to make.
DiscoveredDiscovered at submission, when the claim has to be narrowed to whatever the data will carry.
The distinction that decides whether you have a product
Three markers, three different products - and only one of them is a companion diagnostic
A companion diagnostic is a predictive marker by definition. Confusing the three classes is the classic error in this field, and it is how a trial enriches its population and still fails. Switch between them and watch what happens to the treatment effect.
Tells you how the disease will go, regardless of treatment.
Marker negative
46%
Control
58%
Treated
Treatment effect +12 pts
Marker positive
22%
Control
34%
Treated
Treatment effect +12 pts
Enrichment changes who is in your trial. It does not change whether your drug works.
The treatment effect is identical in both subgroups - twelve points either way. The marker tells you which patients do worse, which is genuinely useful for stratification and for prognosis conversations. It tells you nothing about who benefits from this drug, so selecting patients on it buys you a smaller trial and exactly the same answer.
Tells you whether a specific treatment will work.
Marker negative
44%
Control
46%
Treated
Treatment effect +2 pts
Marker positive
21%
Control
49%
Treated
Treatment effect +28 pts
The treatment effect lives in one subgroup. This is what a companion diagnostic is made of.
Two points of separation in the marker-negative group and twenty-eight in the marker-positive one. That interaction between marker and treatment is the entire basis of a companion diagnostic claim - and it is established by a trial designed to detect it, not by a literature search.
Tells you whether the patient has the disease at all.
Against the reference standard
188
True positive
24
False positive
12
False negative
776
True negative
Sensitivity
0.940
Specificity
0.970
Answers a different question entirely - and says nothing about which therapy will work.
Detection performance against a reference standard, characterised as sensitivity and specificity with confidence intervals. A superb diagnostic marker can have no relationship whatsoever to treatment benefit, which is why using one to select therapy is not a shortcut but a category error.
Illustrative response rates and counts, drawn to make the distinction visible.
The five services
Marker, assay, kit, manufacture, dossier - on one contract
Each is independently available, and most programmes enter in the middle. Run end to end, they are the reason a marker can become a manufactured product without changing partners.
1Companion diagnosticsWhich patients should receive this drug - and can we prove it?Co-developed with the trial
A companion diagnostic developed against the drug it selects for, on the therapeutic programme’s timeline, with the validation study designed before it runs.
What you get
- Candidate marker assessment against the drug’s mechanism, with evidence traced to source
- Assay format selection and development
- A powered clinical validation study design, agreed before the study runs
- Co-development alignment with the therapeutic trial timeline
- Analytical and clinical validation data packages
- Documentation prepared for regulatory submission
What the platform contributes
Span-cited biomarker evidence
Our biomarker layer runs on 421 span-anchored evidence records across 184 targets and 86 diseases - every one carrying the quoted sentence and a character offset into a real paper. You read the sentence the claim came from instead of trusting a score.
Mechanism linkage, shown rather than asserted
A companion diagnostic exists to select patients for a specific drug, so the marker has to connect to that drug’s mechanism. We assemble that connection from Open Targets evidence, ChEMBL pharmacology and the knowledge graph, and we show the evidence paths that support it.
Regulatory and competitive context
FDA Orange Book approvals, exclusivity and listed patents, plus openFDA structured label data - indications, boxed warnings, mechanism - so the programme is scoped against what is actually approved and what protects it.
Enrichment impact modelling
A biomarker requirement is the single most restrictive thing you can put in an eligibility protocol. Our eligibility model quantifies the trade-off before you commit to it: a restrictiveness score weighted by criterion type, the eligible fraction of the diagnosed population, the resulting addressable pool, and a ranked list of which criteria remove the most patients for the least scientific value.
Measured performance
| Benchmark | Result |
|---|---|
| Biomarker recovery vs known clinical markers (top-20, 4 diseases) | Macro recall 0.835 - breast carcinoma 1.00, lung 0.875, colorectal 0.80, melanoma 0.667 |
| Evidence retrieval (4 diseases) | Macro recall@15 0.917 |
| Target dossier generation (6 targets) | Completeness 1.00 · citation coverage 1.00 across 72 cited items |
| Provenance coverage | 100% |
Also on the record: actionable-tag precision is 0.271. Retrieval works; the automatic "clinically actionable" tag over-fires roughly three to one, so we treat it as a triage hint and never as a claim.
2Liquid biopsyWhat belongs on the panel, and can we detect it reliably?Panel content · analytical performance
Panel content chosen against real population frequency and clinical classification, then characterised empirically in the matrix you will actually sample.
What you get
- Panel content selection with the rationale for every included variant or marker
- Assay development for the chosen matrix and analyte
- Analytical performance characterisation - sensitivity, specificity, limit of detection, precision, matrix effects
- Clinical validation study design and execution
- A protocol and transfer package written for your laboratory
What the platform contributes
gnomAD population frequency
Combined exome and genome allele counts and frequencies, dbSNP identifiers, VEP consequence and HGVSp protein change - so panel content is chosen against real population frequency rather than against a published list.
ClinVar with the review-status star rating preserved
A single-submitter call is never presented as equivalent to a reviewed-by-expert-panel one. The star rating travels with the classification into the panel rationale.
Ontology-resolved conditions
MONDO and EFO, so the same indication reported three ways collapses to one node and the intended-use statement stays consistent from the first design document onward.
Expression evidence with its assay and unit intact
For marker selection where the analyte is transcript or protein rather than DNA. RNA and protein readings are never pooled, because they are different molecules measured in different units.
The powering question for a liquid biopsy is usually a difference in proportions against a tissue comparator - and it should be settled before the assay is locked, not after.
3Diagnostic kit developmentCan this assay be made, shipped and run by someone else?Formulation · stability · transfer
A working protocol turned into a manufacturable product - formulation, controls, stability, labelling and a design history that was maintained as the design changed rather than reconstructed afterwards.
What you get
- Kit design: assay format, reagent composition, controls and workflow
- Reagent formulation and stability programme
- Lot-to-lot consistency data
- Instructions for use and labelling content
- Design history documentation, maintained as the design changes
- Transfer to manufacture
What the platform contributes
Analyte and target selection
The expression, variant and evidence work described above, applied to the question of what the kit should measure in the first place.
Control design informed by population frequency
A control that is itself polymorphic in the target population is a recurring and expensive failure mode - and an entirely avoidable one, because the frequency data exists before the control is designed.
Indication scoping against ontology-resolved definitions
So the claim the kit will eventually make is stated consistently from the first design document to the instructions for use.
Kit development is largely a laboratory and engineering discipline, and we describe it that way rather than attaching AI language to it. Formulation, stability, ruggedness and lot consistency are empirical work at the bench, and no model substitutes for any of it.
4OEM and white-label manufacturingWho makes it, at volume, with documentation we can hand to a regulator?Batch records · CoA · continuity
Manufacturing capacity, quality documentation and a single commercial relationship covering both the science that produced the kit and the plant that makes it.
What you get
- PCR and qPCR kits manufactured to your specification
- DNA/RNA extraction kits
- qPCR reagents and master mixes
- Custom-branded molecular diagnostic kits under your label
- Batch documentation, certificates of analysis and release testing
- Supply continuity planning
What the platform contributes
This service has no computational component at all
It is manufacturing, delivered through our scientific partners. We say so plainly, because a page that attaches an AI claim to reagent filling is a page that makes the rest of its claims harder to believe.
Two ways companies arrive here
They developed a kit with us and want it made - or they have a validated assay of their own and want a manufacturing partner who understands the science behind it rather than treating it as a filling job.
5Validation and regulatory documentationWill the evidence package support the claim we want to make?Powered before it runs
Validation designed for the claim rather than the claim trimmed to fit the validation - with every calculation reproducible and every document built for the partner who will submit it.
What you get
- Analytical validation plans and executed data packages
- Clinical validation study design, powered for the claim you intend to make
- Method comparison and agreement analysis against the reference method
- Design history and technical documentation
- Risk management and traceability documentation
- Submission-ready dossiers for your regulatory consultant or affairs team
What the platform contributes
Statistical study design, done before the study
Closed-form power and sample-size calculation for the three endpoint types that cover most diagnostic validation work, with one- or two-sided testing, configurable alpha and power, and dropout inflation.
Power computed in reverse
From a fixed sample size - which is the question a sponsor actually asks: what can I detect with the cohort I can realistically recruit?
Adaptive and staged designs where a fixed design would be wasteful
Six designs tested in simulation, with maximum Type I error deviation of 0.001 from the nominal 0.05 and minimum power 0.909.
Transparent rather than emitted
These are the same closed-form calculations a biostatistician would run, implemented so every number is reproducible and explainable instead of arriving out of a black box.
We prepare documentation and generate validation data. We are not a regulatory consultancy and we do not act as your authorised representative, notified body liaison or submission agent.
Enrichment impact modelling
What the biomarker requirement actually costs you, before you write it into the protocol
A biomarker criterion is the single most restrictive thing you can put in an eligibility protocol. The model quantifies the trade-off in advance: the eligible fraction, the addressable pool that survives it, and which criteria remove the most patients for the least scientific value.
Diagnosed population in scope
48,000
Confirmed histology, measurable disease
retains 92%
Disease definition44,160 remaining · −3,840ECOG performance status 0–1
retains 78%
Fitness34,445 remaining · −9,715Biomarker-positive by the candidate assayMost restrictive
retains 31%
Enrichment10,678 remaining · −23,767No prior systemic therapy for advanced disease
retains 64%
Line of therapy6,834 remaining · −3,844Adequate organ function
retains 88%
Safety6,014 remaining · −820
Addressable pool
Eligible fraction 12.5%
6,014
What the model is telling you
The biomarker criterion alone removes more than twice as many patients as any other line in the protocol. That may be exactly right - enrichment is the point of a companion diagnostic - but it should be a decision taken with the number in front of you rather than discovered during recruitment.
- Eligible fraction
- 12.5%
- Screened per patient enrolled
- 8
Honest scope on this model
Per-criterion retention uses typed heuristics from the trial-methodology literature, not patient-level screen-fail rates from real-world data - those require EHR or claims access we do not have. Every retention factor is an adjustable, visible assumption rather than a fitted number, and we will show you each one and argue about it with you.
Powered before it runs
The sample size is a calculation, not a negotiation
Normal approximation, per-arm sizing
Sensitivity or specificity against a comparator method
Detect 0.80 sensitivity against a comparator at 0.60, two-sided α 0.05, power 0.80
- Required
- 82 per arm
- Achieved power
- 0.804
- Recruit at 15% dropout
- 97 per arm
Normal approximation
Quantitative assay agreement
Detect a half-standard-deviation difference, two-sided α 0.05, power 0.80
- Required
- 63 per arm
- Achieved power
- 0.801
- Recruit at 15% dropout
- 75 per arm
Schoenfeld formula
Outcome-linked clinical validation
Detect a hazard ratio of 0.70, two-sided α 0.05, power 0.80
- Required
- 247 events
- Achieved power
- 0.800
Time-to-event designs size on events rather than patients, so recruitment and follow-up are planned from the event rate rather than inflated for dropout.
Enterprise AI for Life Sciences
Ready to Advance Your Drug Discovery Pipeline?
Partner with Prognica Labs to leverage enterprise-grade AI, computational chemistry, and molecular simulation technologies that accelerate discovery, reduce development risk, and improve R&D productivity.
From biotech startups to global pharmaceutical organizations, we help research teams make faster, evidence-driven decisions across every stage of early drug discovery.