flowchart TB
subgraph M[Mechanistic layer]
A[Vessel source geometry] --> B[Diffusion through matrix]
B --> C[Binding and internalisation]
end
subgraph E[Evidence layer]
D[Same-tissue measurement]
F[External validation]
G[Literature prior]
H[Missing / blocked]
end
M --> I{Evidence sufficient?}
E --> I
I -->|Yes| J[Relative target ranking]
I -->|No| K[NOT_COMPUTED]
reach-gap
Is the target physically reachable?
v0.8.0 · Research preview View repository
The problem
A tumour target can be highly expressed and still be difficult for an antibody to reach. Accessibility depends on far more than expression level:
- which vessels are structurally present, and which are actually perfused;
- how far target-positive cells lie from those vessels;
- how quickly an antibody moves through the extracellular matrix;
- how strongly and how densely it binds;
- whether an administered antibody reaches and engages the tissue at all.
These quantities are rarely measured together in the same sample. The common failure mode is quiet substitution: combining RNA, fluorescence, literature diffusion constants and vessel distances as though they were interchangeable.
The idea
reach-gap has two linked layers.
The mechanistic layer represents vascular geometry, diffusion, binding and internalisation in a simplified 2D reaction–diffusion model. The evidence layer records, for every required input, whether it was measured in the same tissue, externally validated, borrowed as a prior, missing, or explicitly blocked from transfer.
The second layer is the point. It makes questionable substitutions visible instead of letting them disappear into a single number.
What it returns
Depending on the evidence available:
- mechanistic simulation outputs with propagated uncertainty;
- relative target-rank probabilities and pairwise comparisons;
- evidence-readiness and measurement-priority reports;
- or an explicit abstention such as
reachable_fraction = NOT_COMPUTED.
The last of these is a feature. An absolute reachable fraction requires biology that usually has not been measured, so the tool declines to produce one rather than fabricating false precision.
Design choices
| Choice | Reason |
|---|---|
| Relative ranking rather than absolute fractions | The absolute quantity depends on unmeasured inputs; the relative comparison is defensible |
| Six vessel definitions instead of one | Vessel identity is itself a modelling decision, and the ranking should be shown to survive it |
| 20,000 uncertainty draws | Point estimates hide how wide the plausible range actually is |
Explicit NOT_COMPUTED outputs |
A missing answer is more useful than a confident wrong one |
| Auditable evidence graph | Every transfer between data types can be inspected and challenged |
Verified release
| Item | Current state |
|---|---|
| Automated tests | 117 |
| Coverage | 88.3% |
| Uncertainty draws | 20,000 per analysis |
| Vessel definitions compared | 6 |
| Containerised | Yes (Docker) |
| Licence | MIT |
Honest limitations
- The reaction–diffusion model is a simplified 2D representation, not a full pharmacokinetic simulation.
- Relative rankings are comparisons between targets in the same tissue context; they do not transfer to other tumours unchanged.
- Literature priors remain priors, however carefully labelled.
- Research use only: no prediction of patient response, treatment recommendation, or estimate of clinical efficacy.