Reading an in silico result against the bench

A prediction is not a measurement. Every tool here returns a number that was graded against something, and what it was graded against decides what you can do with it. A pose scored against a deposited crystal structure tells you the geometry is plausible; it does not tell you the compound binds. A toxicity score with an AUC of 0.88 against a curated label set tells you the model separates known actives from known inactives; it does not tell you what happens in your assay.

Tools that make predictions carry a translatability child page — /docs/<tool>/translatability — that states this for that tool, with citations. This page explains how to read those pages, and indexes them.

There is no in silico to in vivo comparison, and that is not a gap in our sourcing

The literature does not contain studies that compare an in silico prediction directly against an in vivo outcome. Papers that pair the two are almost always sequential: a computational screen nominated a compound, and animal work followed. That is a workflow, not a concordance measurement — nothing in the design establishes how often the prediction was right.

Three comparisons do exist in quantity, and they are the ones the tool pages use:

1. Against other in silico methods — head-to-head on a shared benchmark, same task, same inputs.
2. Against the standard wet-lab approach — for structure and pose prediction, experimental structure determination, where agreement has a settled definition (heavy-atom RMSD, conventionally ≤ 2 Å); for property and toxicity prediction, the assay panel, reported as AUC or accuracy against assay-derived labels.
3. Against the combined tooling here — what the platform's own chain achieves, which is the only column we can measure ourselves.

Each is compared on time, cost, and accuracy. Anything past that — whether a confirmed pose becomes a compound that works in an animal, let alone a patient — is the part of drug discovery that no benchmark measures.

What the wet-lab cost figures mean

No commercial CRO publishes a list price for any of these techniques. Across roughly 110 vendors checked in October 2026 — including SARomics, Viva Biotech, Creative Biostructure, Proteros, Charles River, Evotec and WuXi — the count of public commercial prices for co-crystallography, ITC, MST, hERG patch clamp, hepatocyte stability and cryo-EM is zero. Everything is quote-only. Eurofins Discovery and Pharmaron both run per-item catalogs with fixed specifications and turnaround times where the price itself renders as "Login to View Pricing," so a list price demonstrably exists and is simply not published.

So the wet-lab cost cells on these pages quote academic core-facility rate cards at their industry or commercial tier, with the facility named and the rate card dated. Read them as recharge rates, not market prices — a core facility recovers its costs, while a CRO prices a service, and the gap between the two is unmeasured. Treat every wet-lab cost figure on these pages as a floor.

The figures used, verified against the published rate cards:

The runs column is what the same money buys here: the number of 1-credit runs (ADMET-AI, ProTox-3, AiZynthFinder, P2Rank) at $0.004 per credit, the subscription-plan rate, with euros converted at €1 = $1.1205 (ECB reference rate, 22 October 2026). The 1,000- and 2,000-credit packs cost $0.005 per credit, so the same money buys about a fifth fewer runs there. A tool that costs N credits per run gets N times fewer runs, and each tool page states its own figure. Runs are rounded down.

Platform costs are stated in credits, derived from public.job_pricing, and are the one column measured rather than estimated:

⊘ Job pricing changes, so re-derive these from public.job_pricing rather than citing this table.

Accuracy is never comparable across those rows

Read an accuracy figure cell by cell, never down a column. Crystallography's accuracy is definitional, because it is the thing the others are graded against. A docking percentage is the share of poses inside a geometric threshold. An affinity method's accuracy is an energy error in kcal/mol. An ADMET model's is an AUC against a label set. Nothing meaningful comes from comparing those to each other.

What the comparison does support is time and cost at a stated accuracy. That is the claim these pages are built to make.

Property predictors are graded one step removed from the bench

When a docking method is graded against a co-crystal, the crystal structure is the measurement. When a toxicity model reports an AUC, the comparison is against a label derived from an assay someone else ran — drawn from Tox21, ToxCast, ChEMBL, or an outcome set. The assay is still ground truth, but it reaches the model through a curation step: inclusion criteria, activity thresholds, and whatever chemistry happened to be screened.

A model's AUC is therefore a property of the model and of the label set. Figures in the 0.85–0.93 range on benchmark toxicity sets are routine, and they do not predict performance on a scaffold series absent from those sets. The same holds for off-target reporting: named off-targets can only be compared against panel screening that exists.

Ask what the label actually encodes. The sharpest failure in this class is a model that scores well because its label is a proxy for something easy. An antibody humanness model separating human from murine sequences looks strong at AUC 0.97 — but stratify the same set by species of origin and the correlation with clinical immunogenicity vanishes in every subgroup, because the score was distinguishing provenance, not predicting an outcome. One group put it directly: it "indirectly measures the ability of the model to distinguish between species." The reader's molecules all sit inside one of those subgroups, so the headline figure describes a comparison they will never make. Whenever a label is a category someone assigned rather than a quantity someone measured, check whether the model's apparent skill survives holding that category fixed — and expect the independent number to be far below the author-reported one. On BioPhi that gap is R² = 0.26 against R² = 0.02.

Training overlap is the first thing to check

Benchmark sets for structure and docking methods are drawn from the PDB, and the models were trained on the PDB. When a test complex resembles a training complex, the reported number is partly a lookup. The effect is measured: deep-learning docking methods degrade as sequence identity to their training data falls.

So when a page gives two numbers that look contradictory — strong on one benchmark, weak on another — the usual explanation is that one set overlaps training data and the other was built to avoid it. Both are real. The second describes a novel target.

Check the version, too. Published criticism often tests an earlier release than the one running here, and a figure quoted without its version is unusable.

What every translatability page contains

A short verdict block opens the page, for readers holding a result who want to know whether it is useful. Then six sections, in the same order on every page so tools can be compared side by side:

Sourcing rules. Figures are quoted with their source and metric. Where a tool has no published comparison for a section, the page says so rather than borrowing a figure from a different tool or version. A result from a multi-tool pipeline is attributed to the pipeline. Reviews and protocols provide context, never a concordance figure — the citation must be the primary comparative study.

Pages by domain

Pages exist where a tool makes a prediction that can be graded against an experiment. Lookup and retrieval tools — ChEMBL, PubChem, BLAST, UniProt and the other database tools — return records rather than predictions, so they have nothing to translate and carry no page.

Worked example: DiffDock

Pose prediction graded against co-crystal structures, including why the same method is reported at 12% and at 50%.

Adding a page for a tool

1. Create content/docs/<tool>/translatability.mdx with title, description, and tags including translatability.
2. Add "translatability" to that tool's meta.json pages array, after results or outputs.
3. Write the verdict block, then the six sections above in order. Keep a section and state that no published figure exists rather than dropping it.
4. Find the evidence by searching PubMed directly rather than relying on web search, which under-reports this literature. Resolve every PMID to a DOI before citing it.
5. Cite as NLM-style blockquotes with doi.org links, matching the other docs pages.
6. Move the tool's row in the table above to Published.