Leakage-safe primary test
Held-out family or pathway
Compare frozen UniPert features with native ridge, ESM2-only, nearest-neighbour and random-embedding controls. Report verified-unseen and unresolved-exposure results separately.
NWLE-10 · partial feasibility audit
NMD-VCell can evaluate UniPert now as a frozen gene-and-compound identity representation. The first release separates verified source and identifier readiness from every encoder, embedding, transfer and DMD-response claim.
Source and implementation audit
| Audit item | Verified state | Operational consequence |
|---|---|---|
| Formal paper | Cell 2026 · DOI 10.1016/j.cell.2026.06.005 | Citable method and reported benchmarks; not DMD validation. |
| Public repository | Audited commit 2f5d46930dcbdeb92073a13e898abe6e363e679a | Frozen model and encoding demos are present. |
| Reproduction surface | Full G2CP training and paper-split reproduction pipeline not found in the audited public tree | Do not label a local encoder run as paper reproduction. |
| Condition semantics | Identity inputs are available; direction, dose, time, modality and DMD state are not direct tokens | Use as an identity feature, not a condition-response simulator. |
| Rights | Article CC BY · code GPL-3.0 · article reports related patent filing | Retain licence provenance and perform separate deployment/FTO review. |
U0–U4 execution ladder
| Stage | Task | State | Current result |
|---|---|---|---|
| U0 | Paper, code, licence, patent-notice and implementation-surface audit | EXECUTED | Source and implementation boundaries frozen in this contract. |
| U1 | Candidate identifier and sequence-input preflight | PARTIAL_IDENTIFIER_PREFLIGHT_EXECUTED_ENCODER_NOT_RUN | 21/21 candidates map to reviewed human UniProt primary accessions. |
| U2 | Cause-space sanity checks and known-relation retrieval | REGISTERED_NOT_EXECUTED | No result |
| U3 | Leakage-safe G2 incremental benchmark | REGISTERED_NOT_EXECUTED | No result |
| U4 | Public gene-to-compound cross-domain learning curve | REGISTERED_NOT_EXECUTED | No result |
Candidate input preflight
| Gene | UniProt | Sequence length | State |
|---|---|---|---|
| ADAM10 | O14672 | 748 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| CALR | P27797 | 417 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| CPEB1 | Q9BZB8 | 566 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| DDX19B | Q9UMR2 | 479 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| DNAAF3 | Q8N9W5 | 541 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| DNM1 | Q05193 | 864 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| EHMT2 | Q96KQ7 | 1210 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| EPS8L1 | Q8TE68 | 723 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| GFOD2 | Q3B7J2 | 385 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| INTS13 | Q9NVM9 | 706 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| LMO2 | P25791 | 158 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| MON1A | Q86VX9 | 652 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| MPHOSPH6 | Q99547 | 160 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| NAGLU | P54802 | 743 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| RAC3 | P60763 | 192 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| RNASEH2C | Q8TDP1 | 164 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| RNF8 | O76064 | 485 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| WDR4 | P57081 | 412 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| ZFP69B | Q9UJL9 | 534 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| ZNF133 | P52736 | 654 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
| ZNF236 | Q9UL36 | 1845 | IDENTIFIER_PREFLIGHT_PASS_ENCODER_NOT_RUN |
Leakage-safe primary test
Compare frozen UniPert features with native ridge, ESM2-only, nearest-neighbour and random-embedding controls. Report verified-unseen and unresolved-exposure results separately.
Adoption gate
Adopt UniPert as a feature only if it improves a predeclared held-out task over simple baselines with uncertainty, survives exposure and leakage audit, and does not degrade worst-group performance.
Hard boundary
It does not specify CRISPRi, CRISPRa, KO, overexpression, dose, time, cell state, disease background, mechanism, function or safety. Those variables need explicit conditioning and independently evaluated outcomes.