A viability figure that is trustworthy inside one laboratory can become ambiguous once it leaves the building. The receiving site may not know whether "viability" means live cells over total cells at thaw, or live cells relative to the cells that went into the freezer. It may not know how many cells were counted, which weights produced a quality index, or whether a decline rate was measured in hours or days. WIA-CRYO-010 Phase 2 settles these questions by fixing the arithmetic. Viability, recovery, effective viability, Wilson intervals, the Vitrification Quality Index (VQI), meta-analytic pooling and heterogeneity are each defined as explicit formulas. This chapter shows how those formulas support data exchange between institutions.
Three community specifications frame the discussion. The FAIR principles say what makes data reusable. DCAT 3.0 is a W3C vocabulary for describing catalogs, datasets and their distributions. The Common Workflow Language (CWL) is a declarative way to describe analysis steps so that different sites run the same computation. None of them is defined by WIA-CRYO-010, and none replaces it. They carry Phase 2 quantities between sites without changing their meaning.
For hands-on exploration, the reference material includes an interactive browser-based simulator for the protocol (wia-standards/standards/cryo-research/simulator/). Alongside it sits a TypeScript SDK skeleton and a POSIX shell client. All of these are starting points, not production systems. The simulator is useful for checking, before real data move, whether your reading of a formula matches the reference behaviour.
FAIR stands for Findable, Accessible, Interoperable and Reusable. In cryopreservation research, the hardest of these is usually reusability. A dataset is reusable only if a stranger can recompute its headline numbers from what travels with it. Phase 2 makes that possible, but only if the inputs of each formula are shared along with the outputs.
| Principle | What it means for a cryo dataset | Phase 2 hook |
|---|---|---|
| Findable | The dataset has a persistent identifier and a catalog record that names the cell type, preservation method and measured outcomes | Outcome names align with the quantities defined in Phase 2 (viability, recovery, VQI, success rate) |
| Accessible | Access conditions are stated, even where the data themselves are restricted | Summary statistics (counts, means, standard errors) can be public while sample-level records are controlled |
| Interoperable | Sites use the same definition of every derived quantity | Viability = viable / total × 100; Recovery = post-thaw / pre-freeze × 100; Effective Viability = Viability × Recovery / 100 |
| Reusable | Provenance, denominators, weights and units accompany every derived value | Sample size n for the confidence interval; w₁–w₅ for VQI; time units for decline rate |
Two details are easy to miss. First, a VQI value is meaningless without its weights. The defaults are 0.4, 0.3, 0.2, 0.05 and 0.05 for viability, metabolic activity, functional score, ice crystal score and apoptosis rate. A site that changed them has produced a different index even though the scale is still 0–10. Second, a decline rate of (V₀ − Vₜ) / t depends entirely on the unit of t. Shared records therefore have to state it.
DCAT describes data in layers. A Catalog holds Datasets. Each Dataset has one or more Distributions, which are concrete downloadable or queryable forms. A DataService describes an endpoint that serves data. Version 3.0 also adds dataset series, which suit longitudinal programs such as repeated storage-stability checks on the same cohort.
For cryo research, the natural split is by sensitivity. One distribution holds de-identified per-sample measurements, such as cell counts before freezing and after thawing. A second holds site-level summaries: viability with its 95% interval, mean VQI, and the number of samples. Summaries are what other sites usually need, and they can circulate more freely. The sketch below is illustrative Turtle using standard DCAT and Dublin Core terms. The identifiers are placeholders.
ex:post-thaw-panel a dcat:Dataset ;
dct:title "Post-thaw viability panel, site summaries" ;
dct:conformsTo ex:wia-cryo-010-phase2 ;
dcat:distribution ex:panel-summary-csv ,
ex:panel-counts-restricted .
ex:panel-summary-csv a dcat:Distribution ;
dct:description "Per-site viability, 95% CI, n" ;
dcat:mediaType "text/csv" .The dct:conformsTo link is the important line. It tells a harvester which calculation definitions the numbers follow, so a catalog consumer can tell whether two datasets are comparable before downloading either. Provenance of how a summary was derived can be expressed with W3C PROV-DM, which the standard lists in its informative cross-walk.
Moving biological material and moving its data are separate activities, but they have to stay synchronized. This book frames the exchange as six stages. The framing is a teaching device, not a normative state machine defined by the standard.
| Stage | Activity | Data that must exist at the end |
|---|---|---|
| 1. Request and scoping | The requesting site states the intended analysis and the outcomes needed | A scoped description of the requested measurements |
| 2. Eligibility and ethics review | Consent scope, institutional review and legal basis are checked | A recorded decision and its conditions |
| 3. Agreement | Use limits, publication terms and return of results are fixed | An executed agreement referenced from the catalog record |
| 4. Preparation and packaging | Samples are frozen or selected, with measurements taken before freezing | Pre-freeze cell counts, method of preservation, quality scores |
| 5. Transfer and receipt | Samples ship under controlled cold-chain conditions; data move through the agreed channel | Receipt record; chain-of-custody continuity |
| 6. Post-receipt reconciliation | The receiving site thaws and measures, then recomputes the shared quantities | Post-thaw counts, recovery, effective viability, comparison with the sender's figures |
The reason stage 4 matters so much is Phase 2's recovery formula, Recovery (%) = Cells Post-Thaw / Cells Pre-Freeze × 100. The numerator is measured at the receiving site, often months after the denominator was recorded at the sending site. If pre-freeze counts do not travel with the sample, recovery and effective viability cannot be computed at all. Only the cruder viability percentage remains. Stage 6 closes the loop. The receiving site recomputes the quantities from the raw counts and compares them with what the sender reported. Any discrepancy points to a counting or definition difference, not a handling failure.
CWL describes each analysis step as a tool with declared inputs and outputs, and a workflow as steps connected by those declarations. Steps are normally pinned to a container image, so the same description gives the same result at any site that runs it. For cryo research, the Phase 2 pipeline decomposes naturally: raw counts to viability, viability to interval, per-site summary, then pooled estimate and heterogeneity. Each arrow is a step whose output is exactly what the next site needs.
The core viability step in the reference material is short. This version returns percentages and a standard error computed from the normal approximation:
def calculate_viability(viable_cells, total_cells):
"""Calculate viability percentage with 95% CI"""
viability = viable_cells / total_cells * 100
se = np.sqrt(viability * (100-viability) / total_cells)
ci_lower = viability - 1.96 * se
ci_upper = viability + 1.96 * se
return {
'viability': viability,
'ci_95': [ci_lower, ci_upper],
'se': se
}A worked check shows why a workflow should also offer the Wilson interval from Section 2.2. Take 80 viable cells out of 100. The normal approximation gives a standard error of 4 percentage points and an interval of roughly 72.2% to 87.8%. The Wilson formula, with center = (p + z²/2n) / (1 + z²/n), gives about 71.1% to 86.7%. The two are close at this sample size. They diverge as n shrinks or as p approaches 0 or 100, where the normal interval can leave the 0–100 range. Because small per-site counts are common in multi-site studies, a shared workflow should document which interval it reports, so that sites do not compare unlike quantities.
Another place workflows need discipline is the VQI. The reference implementation clamps the result to 0–10 after applying the weights, so the exchanged index can never leave its declared range. Whether you use CWL or another tool, keep that clamping in the step itself rather than in downstream scripts.
Federated analysis keeps individual-level records at the site that collected them and exchanges only aggregates or model updates. Many Phase 2 methods are already federated in this sense. A random-effects or inverse-variance meta-analysis needs only an effect estimate θᵢ and standard error SEᵢ from each site. The pooled mean is θ̂ = Σ(wᵢθᵢ) / Σwᵢ with wᵢ = 1/SEᵢ². Heterogeneity is I² = 100% × (Q − df) / Q, where Q is Cochran's statistic. A compact illustration:
def pool(theta, se):
w = [1 / s**2 for s in se]
est = sum(wi * t for wi, t in zip(w, theta)) / sum(w)
q = sum(wi * (t - est)**2 for wi, t in zip(w, theta))
df = len(theta) - 1
i2 = max(0.0, 100 * (q - df) / q) if q > 0 else 0.0
return est, i2Flooring I² at zero is the usual convention when Q is smaller than its degrees of freedom. The standard marks I² above 75% as high heterogeneity. In practice that is a signal to stop and ask why sites disagree before reporting a single pooled number. Common causes in this field are different viability assays, different cryoprotectant protocols, and different post-thaw timing.
| Phase 2 method | What each site must contribute | Sharing note |
|---|---|---|
| Viability, Wilson interval | Viable count and total count | Counts can be summed across sites, but per-site results should be kept to detect heterogeneity |
| Two-sample t-test, Cohen's d | Mean, standard deviation, n per group | Enough to compute t and the pooled standard deviation without raw values |
| ANOVA F | Group sums of squares, group sizes | MS_between and MS_within combine across sites only if groups are defined identically |
| Kaplan-Meier | Events dᵢ and number at risk nᵢ at each time | Event tables are aggregates, but very small risk sets can still reveal individuals |
| Logistic regression | Iterative updates to coefficients | Needs several rounds of exchange |
| Random forest, neural network | Model parameters or updates | Neural networks (Adam, learning rate 0.001) suit parameter averaging; tree ensembles are awkward to merge |
Federation does not by itself guarantee privacy, because aggregates and model updates can still leak information. The standard's open-questions annex records privacy-preserving aggregation, including differential privacy and secure aggregation, as future work without prescribing an algorithm. Treat any claim of "privacy by federation" with caution until a site has reviewed what its outputs reveal.
National biobank networks, such as the one abbreviated KBN in Korea, face a cataloguing problem that single laboratories do not. Member institutions hold collections described in different local vocabularies. An integrated catalog works by harvesting standardized records from members, here DCAT descriptions, and presenting them as one searchable index. It does not copy the underlying data.
For cryopreserved material, a catalog entry is useful to researchers only if it exposes comparable facets. The most valuable are the preservation method, the outcome measures with their Phase 2 definitions, the sample count behind each summary, and a data quality indicator. Phase 2 provides one: DQ = (Completeness + Consistency + Accuracy + Timeliness) / 4, each component scored from 0 to 1. A catalog can display DQ so that a researcher sees at a glance whether a collection's records are complete and current. One limitation is that DQ is an average, and a high mean can hide one poor component. Show the four components alongside the score.
The catalog should also resist the temptation to pool. Surfacing a single network-wide viability figure from member summaries invites the heterogeneity problem of Section 5.6. A better design lists member summaries side by side, with I² computed on request. Sample-level identifiers stay at the member sites and are released only after the review in Section 5.8.
Before a dataset or summary leaves an institution, a short review catches the problems that most often make shared cryo data unusable or unsafe.
On the compliance side, the standard defines three conformance tiers (self-declared, third-party assessed and accredited). A site claiming Tier 2 or Tier 3 must pass the published test vectors and disclose its tier in its OpenAPI description. A partner can therefore ask for that evidence before accepting a site's numbers.
Mapping the workflow onto Korean institutions mostly means identifying who decides at each stage. The items below are pointers for planning, not legal advice. Each site should confirm the details with its own institutional review board and legal office.
| Need | Korean-context element | How it fits |
|---|---|---|
| Ethics review of human-derived material | The Bioethics and Safety Act and institutional review boards | Stage 2 decision; its conditions are recorded and linked from the catalog entry |
| Handling of personal data | The Personal Information Protection Act, including its provisions for pseudonymized data used in research | Governs what per-sample distributions may contain and who may receive them |
| Cross-institution discovery | The national biobank network catalog (KBN) | Receives DCAT-style records and indexes member collections |
| Cold-chain transfer | Site logistics and receiving-laboratory procedures | Stage 5 receipt record; continuity of custody |
Two practical points follow. First, a catalog that is open to the public should carry summaries, not sample-level data, so that review obligations attach only to what is requested. Second, a Korean partner exchanging data with a foreign one must check each jurisdiction's rules. The standard itself notes that long-term retention across jurisdictions is an open alignment question, and the sites involved have to settle retention terms in their own agreement.
The following scenarios are hypothetical. They are built from the formulas of Phase 2 and are not reports of actual studies.
Scenario A: three sites, two assays. Sites report viability for the same cell line after thawing. Two count membrane-excluding dye uptake manually, and one uses an automated counter. The pooled estimate looks reasonable, but I² exceeds 75%. Instead of discarding the third site, the analysts stratify by assay, report each stratum, and record the assay in the catalog entry so later users can reproduce the split.
Scenario B: a small site near the ceiling. A site with few samples reports viability very close to 100%. The normal-approximation interval extends above 100%, which is not physically meaningful. Switching to the Wilson interval from Section 2.2 keeps the range inside 0–100, and the shared workflow is revised to report Wilson intervals by default.
Scenario C: mismatched time units. One site records decline rate per day and another per hour. Both submit "decline rate" with the same field label, and the pooled slope is nonsense. The remedy is the discipline of Section 5.2. Each value travels with its unit, and V(t) = V₀ × e^(−λt) is fitted only after units are harmonized.
Scenario D: planning a joint comparison. Two protocols are expected to yield success proportions of 0.8 and 0.7. With α = 0.05 (Zα = 1.96) and power 0.80 (Zβ = 0.84), the formula in Section 2.3 gives n = (2.8)² × (0.16 + 0.21) / 0.01, or about 290.08, which rounds up to 291 per group. The R function calculate_power in the reference material returns the same rounding. Splitting 291 per group across sites is a scheduling question. Whether sites differ enough to need a site term in the model is a statistical one, and I² from a pilot round can inform it.
Sharing cryopreservation data between institutions fails most often because of unstated definitions, not because of transport. A viability percentage without its denominator, a quality index without its weights, or a decline rate without its time unit cannot be compared. The remedy is to ship the inputs of each Phase 2 formula along with its output. DCAT 3.0 records say where data are and which definitions they follow. CWL-style workflows let each site run the same computation. Federated analysis keeps sample-level records at home, since meta-analysis needs only an estimate and a standard error per site. The review list in Section 5.8 and the mapping in Section 5.9 keep each exchange within its consent and legal limits.
The chapter also leaves caveats. Federation does not guarantee privacy, I² above 75% means pooled numbers need explanation, and the DQ average can hide a weak component. The calculations here assume that numbers arrive intact and in a known form. How exchanges are actually carried out, with messages, sequencing and error handling, is the subject of the standard's protocol phase (Phase 3), and the book turns to it next.