Name a target. We return the count: how many active agents, from which classes, can act on it — and, for everything else, the specific reason it cannot. Most tools hand you a score for anything you ask. This one tells you when the honest answer is “no” or “not knowable yet”.
A target is whatever you want acted upon: an organism, one of its proteins, or a protein of your own body. Coverage differs by kind, and it is stated here rather than discovered by you halfway through.
Strongest case. Whole-organism metabolic models exist, essential genes can be computed, and the result can be checked against drugs that already work. Tuberculosis, H. pylori, Acinetobacter, Klebsiella, MRSA have all been run through unmodified.
Malaria and related organisms behave like bacteria for this purpose: they have their own metabolism, so the same reasoning applies without special cases.
Candida, Aspergillus and similar are reachable in principle, but fewer validated models exist and the selectivity question is harder: fungal and human cells are closer relatives than bacteria and humans.
A virus has no metabolism of its own, so the metabolic route does not apply. Two routes remain: viral proteins as direct targets, and host factors the virus depends on. We say which of the two we used — conflating them is a common way to be confidently wrong.
Human metabolic models exist and can be run, but the selectivity question dominates everything: the target sits inside the patient. We do not issue a therapeutic window.
Non-infectious targets — a receptor, an antagonist, a signalling protein. Here the question is structural, not metabolic: is there anything on this protein to bind at all. We have returned “no pocket” on our own project rather than force a design.
Anything requiring a human trial to answer: dosing in patients, safety, efficacy, drug–drug interaction. Also anything defined only as a disease name rather than a molecular target — “cure diabetes” is not a target.
Which mechanisms can be assessed with the available evidence? Supported packages distinguish ranked hypotheses, exclusions, scope mismatches and missing information.
Candidates ranked within a curated graph, with evidence coverage and known limitations. A score is not a measured effect or a probability of efficacy.
A candidate fails the stated criterion in the tested model and context. This does not establish that the target is impossible to affect or that the compound has no biological activity.
A number comes out, but not the one you asked for. Our own example: the model says an enzyme is dispensable for growth — while without it the organism cannot colonise the stomach at all. The number is correct. The question is wrong. This bucket catches the failure that quietly wastes the most money.
Refused, with which piece is missing: no model for this organism, no reference to check against, or a question the method is not entitled to answer. In the last full run this was the majority of candidates — and that is the honest state of the field, not a defect of the tool.
What you do not get: a probability of cure, a ranking that pretends the last two buckets are empty, or a number without the reasoning that produced it.
Seventy-four classes of intervention, each with its own vocabulary, gathered from open scientific sources. The bank grows with every target run: a class that had no candidates last month may have three today. Nothing here is restricted to what has already been tried on your target.
The collection started with nucleic bodies: oligonucleotides, siRNA, aptamers. They remain the anchor class because they can be designed against a sequence when no pocket exists on the protein — which is exactly where most target work stops. The header of this site says it in full: bank of nucleic bodies.
A search engine answers and forgets. A bank keeps. Every target that passes through leaves behind classified candidates, dead ends with their reasons, and vocabulary that makes the next target cheaper to assess. What is deposited is not compounds — it is the ability to act on a target, and the record of where that ability runs out.
In each case the value is the same: finding out early that a direction is closed, instead of finding out after the budget is spent.
You lose money on hypotheses nobody could kill. A properly produced negative result, early, is worth more than a survivor that was never tested hard.
You have the target and the bench. You do not have three months of a modeller's time to find out whether the target is reachable at all.
Before you commit a programme to one target, get the count of what could possibly act on it — and the list of what cannot, with reasons.
Independent assessment of somebody else's claim. Specifically useful when the pitch says “druggable” and you want to know on what basis.
When the standard regimen has failed, the question is what else exists at all — including classes outside the usual pharmacy.
Neglected diseases, rare organisms, unfashionable targets. The engine does not care whether a target is commercially popular.
Two of these numbers need explaining, and without the explanation they flatter us. Both explanations are below: a figure you cannot interrogate is not evidence.
The first run scored 12 of 12 and was worth nothing: the tasks were written after we could see how the engine behaved, and tested with the same compounds it had been tuned on. On compounds it had never seen it scored 5 of 7. The gap between those two numbers is the price of an honest test.
The 36-of-38 score is computed over 38 classes out of 74. For the other 36 there is no reference compound at all — nothing to check against. Half the engine is measured by nothing, and a high score does not apply to it. We say this before you ask.
Stated first, not buried. If any of these is a dealbreaker, better that we both find out now.
L1Literature graphL2Quantitative mechanistic modelwe are hereL3Model validated against known outcomespartlyL4Laboratory-confirmed twinnot reachedA prior-art document changed the research decision. The local dossier records the decision to stop drafting; this page does not assert a verified filing or withdrawal status.
The two primary metric files contain 110 and 50 designs, with maximum interface-confidence values of 0.67006 and 0.64171, below the project threshold of 0.70. Binding was not measured. An initial configuration issue limits interpretation of the first run.
Repurposing programmes lose money on the opposite behaviour: on people who cannot stop. That is the capability being sold here, and the bank is what makes it repeatable.
Scope, evidence and deliverables · Оценка мишени: что подготовить и что получить
The private workspace runs two curated packages: H. pylori mechanisms and SmuCA selectivity. Request access below. Other targets require a separate scope and evidence review.
No obligation on either side. If the bank has nothing useful to say about your target, you will be told that instead of being sold a report.