Retrospective concordance of a deterministic computational pipeline for gene-therapy target feasibility

Wilson Mudaki¹

¹Mishale, Inc., Wilmington, Delaware, USA.

Correspondence: contact@mishale.bio

Preprint — not peer reviewed · 9 June 2026

Introduction

Developing a gene therapy is among the most capital-intensive undertakings in medicine, and the cost of abandoning a programme compounds with each phase it survives^10,11. The earliest decision — which gene and indication to pursue — is the point at which analysis has the most leverage, yet in practice it is settled informally: guide-RNA scores from one tool, dosing precedents from the literature, delivery and regulatory rationale from experience, reconciled in a meeting. The resulting judgement is slow to assemble, hard to audit, and not readily comparable across candidates.

Software that systematizes this triage is attractive, yet such systems are seldom held to a falsifiable, pre-registered benchmark, and their behaviour on intractable inputs — what they return when no good answer exists — is rarely reported. We evaluate a deterministic pipeline that produces a feasibility assessment for a (disease, gene) pair from public data, and ask how often its disposition matches the disposition real programmes ultimately reached. The benchmark and scoring were fixed in advance, strata were chosen to elicit failure as well as success, and every error is reported below.

Results

The pipeline

Given a disease and candidate gene, the pipeline executes a fixed sequence of stages (Fig. 1). It resolves the disease to an ontology identifier and selects the target gene; chooses a therapeutic strategy (variant correction, gene replacement, silencing, enhancer disruption); designs the corresponding molecular reagent (guide RNAs for editing, or a transgene cassette with a single-AAV cargo-fit check for replacement); screens candidate off-target sites and assigns a consequence tier; recommends a delivery vector; derives a dosing range from approved-product precedent; and assigns a regulatory class. The decision path is deterministic — identical inputs yield identical outputs — and all assertions are traceable to public sources, principally a knowledge graph integrating ClinVar⁷, gnomAD⁶, the Mondo disease ontology⁸ and the approved-therapy record. Natural-language parsing of the input is the only stochastic component and degrades to a deterministic rule-based parser when unavailable.

Figure 1 | Pipeline stages. A single clinical intent is resolved to a disease and target gene, a therapeutic strategy is selected, a molecular reagent is designed, and downstream off-target, delivery, dosing and regulatory assessments are produced. Each stage is deterministic and sourced from public data.

Benchmark design

We enumerated 69 cases before any were run and assigned each to one of four strata with an explicit pass criterion (Methods). Supported cases (n = 36) are monogenic indications with a single, well-established target gene — most the subject of an approved therapy or an active clinical-stage programme; a case passes if the pipeline resolves that established target gene and returns a feasible or feasible-with-caveats verdict. The single failure case is a programme halted in the clinic; it passes only if the pipeline surfaces the responsible liability. Negative controls (n = 17) are polygenic or complex diseases with no single causal gene; the correct behaviour is to decline. Out-of-coveragecases (n = 15) are monogenic diseases outside the curated set; either a clean decline or a correct design passes, but a fabricated gene fails.

Concordance with clinical outcomes

Aggregated across strata, the pipeline reached the predetermined-correct disposition in 66 of 69 cases (95.7%, 95% Wilson CI 88.0–98.5%; Table 1, Fig. 2). All 36 supported programmes resolved to the correct established target gene with a concordant verdict (36/36; 95% CI 90.3–100%), spanning haematologic, central-nervous-system, hepatic–metabolic, lysosomal, ophthalmic, musculoskeletal, respiratory, renal and multisystem targets and both editing and replacement modalities, including approved AAV gene-replacement products such as onasemnogene abeparvovec (SMN1)²and the editing of BCL11A for the haemoglobinopathies¹. Because the four strata are scored against different pass criteria, the aggregate should be read alongside the stratum-level rates in Table 1 rather than as a single homogeneous accuracy.

Table 1 | Concordance by stratum. Passing requires the predetermined-correct disposition for each stratum (Methods). Intervals are Clopper–Pearson for strata in which every case passed and Wilson score otherwise; the clinical-failure interval (n = 1) is shown for completeness only. No case in any stratum produced a fabricated gene.

Figure 2 | Concordance by stratum with 95% confidence intervals. Points are observed pass rates; bars are 95% intervals (Clopper–Pearson where every case passed, Wilson otherwise). The diamond is the pooled rate across all 69 cases. The wide interval on the negative-control stratum reflects its three errors and is the focus of the error analysis below.

The disposition spectrum

The pipeline does not return a binary verdict. Resolving the 69 outputs into the disposition each case actually received yields six distinct classes (Table 2): an unqualified feasible design (12); a feasible design carrying explicit caveats (24); a feasible design for a monogenic target outside the curated coverage (1); a design returned together with a dose-limiting liability flag (1); a reasoned abstention (28); and the 3 over-resolution errors. Crucially, the two behaviours that protect a programme’s budget — declining when no good target exists, and qualifying a positive call — together account for 52 of 69 outputs, while confident, unqualified designs are the minority.

Table 2 | The disposition spectrum across all 69 cases. Each output classified by the action the pipeline took. “Abstained” denotes a reasoned decline rather than a returned design. Only the final row represents incorrect designs.

Calibration of positive calls

Within the 36 supported programmes — every one resolved to the correct established target gene — the pipeline did not return an unconditional endorsement. It attached explicit caveats (single-AAV cargo-fit limits, off-target burden, dosing ceilings or strategy ambiguity) to 24 of 36 positive calls (67%), reserving an unqualified feasible verdict for the remaining 12. A tool that is to filter candidates is only useful if its confidence tracks real difficulty; that it qualifies roughly two-thirds of even its correct positives, rather than rubber-stamping them, is the behaviour required for the verdict to carry information.

Breadth across organ systems

The 36 supported programmes span 9 primary organ systems (Table 3), from haematologic and central-nervous-system targets to hepatic-metabolic, lysosomal, ophthalmic, musculoskeletal, respiratory and renal disease, and include both ex vivo editing and in vivo AAV replacement modalities. Concordance did not depend on a single tissue or delivery route: the correct gene and a consistent verdict were recovered across all of them. This breadth is what one would expect of a pipeline reasoning from general public-data primitives rather than from indication-specific rules, though it is bounded by the curation overlap noted below.

Table 3 | Therapeutic breadth of the 36 supported targets. Programmes grouped by the primarily affected organ system; representative target genes listed (deduplicated, so a gene shared by two indications appears once).

Recapitulation of a dose-limiting toxicity

We next tested whether the pipeline surfaces a known liability rather than confirming only successes. The AT132 programme for X-linked myotubular myopathy was halted after high systemic doses of an AAV vector produced fatal hepatobiliary events³. Given only the MTM1 target, and with no access to that outcome, the pipeline selected a systemic AAV delivery and flagged a hepatotoxicity ceiling from dosing precedent, reproducing the liability that ended the programme.

Failure behaviour and specificity

Across the 17 polygenic negative controls and 15 monogenic out-of-coverage targets, no case yielded a fabricated gene. For 14 of 17 polygenic diseases the pipeline declined to design, and every out-of-coverage target was either declined or resolved to its correct gene. Off-nominal inputs therefore produced abstention rather than fabricated designs. This is the behaviour that matters when the tool is used to filter candidates: a confident error misdirects budget, whereas a refusal returns the decision to the team.

Error analysis

The three misclassifications were the three negative controls that the pipeline failed to decline (Table 4). All shared one mechanism: a generic, polygenic disease query resolved to a real monogenic-subtype gene and a design was produced. These are over-resolution errors of specificity, not fabrications — each gene is genuinely associated with the disease — but for an unqualified complex-disease query the correct disposition is to decline.

Table 4 | The three misclassifications. Each is a failure to decline a polygenic query, resolving instead to a real disease-associated gene. No fabricated genes occurred.

Discussion

On a pre-registered benchmark, a deterministic pipeline reached the disposition that real gene-therapy programmes ultimately reached in 95.7% of cases across nine organ systems and both editing and replacement modalities, returned graded rather than binary verdicts — qualifying two-thirds of its positive calls and abstaining on more than a third of all inputs — recapitulated a clinical dose-limiting toxicity from public data alone, and produced no fabricated gene across 32 adversarial inputs. The most important limitation qualifies the headline directly: many supported cases overlap the system’s curated knowledge base, so the result establishes high recall and internally consistent reasoning on covered targets, together with graceful failure off coverage. It does not establish generalization to novel biology; a benchmark of targets deliberately held out of curation would be required for that, and we treat the present figures as an upper bound on performance for unseen targets.

The evaluation is retrospective. It measures whether the pipeline agrees, from public data, with outcomes that are now known; it does not measure prospective prediction, and concordance with past decisions is weaker evidence than a prospective trial of decisions not yet made. The three errors point to a concrete deficiency: the criterion for declining complex diseases relies on an enumerated list of polygenic conditions rather than a general signal of polygenicity, so an unlisted complex disease with a known monogenic subtype can be over-resolved. Replacing the enumerated guard with a heritability- or architecture-based criterion is the priority revision. Prospective evaluation on programmes whose outcomes are not yet decided, and expansion of the held-out portion of the benchmark, are the natural next steps.

Finally, the pipeline output is a computational triage from public data, intended to inform an early go/no-go decision. It is neither experimental evidence nor an investigational-new-drug package; every assertion requires wet-laboratory confirmation before use.

Methods

Pipeline. The system (Mishale, version evaluated 9 June 2026) accepts a free-text clinical intent, parses it to a disease and optional gene, and executes the deterministic stages of Fig. 1. Disease resolution maps to a Mondo identifier⁸ and selects the established target gene; a gene named in the query is honoured when it is a recognized target gene for the disease. Strategy selection chooses among variant correction, gene replacement, transcript silencing and enhancer disruption. Molecular design produces ranked guide RNAs for editing strategies or a transgene cassette with a single-AAV (~4.7 kb) cargo-fit determination for replacement. Off-target screening tiers candidate sites by predicted consequence. Delivery, dosing and regulatory classification are derived from the approved-product record. Variant- and gene-level evidence is drawn from ClinVar⁷ and gnomAD⁶; delivery and dosing precedents from approved in vivo AAV and ex vivo products^1,2,4,5,9.

Benchmark and scoring. The 69 cases and the per-stratum pass criteria were specified before execution. Supported cases were defined as monogenic indications with a single, well-established target gene, most of which are the subject of an approved or clinical-stage gene-therapy programme; the failure case as a programme with a documented clinical halt; negative controls as diseases without a single Mendelian cause; and out-of-coverage cases as monogenic diseases outside the curated strategy set. A supported case passed if the resolved gene matched the established target and the verdict was feasible or feasible-with-caveats; the failure case passed if the responsible liability was flagged; a negative control passed if the pipeline declined; an out-of-coverage case passed on either a clean decline or a correct design, and failed on any fabricated gene. Each case was submitted once to the deployed system; outputs were scored against the pre-registered rubric without adjustment.

Statistical analysis. Binomial proportions are reported with two-sided 95% Wilson score confidence intervals. For strata in which every case passed, the interval is given by the Clopper–Pearson method. The aggregate rate combines strata with differing pass criteria and is reported for completeness; stratum-level rates (Table 1) are the primary results. No multiplicity correction was applied, as no hypothesis test was performed.

Data and code availability

The case list and per-case dispositions are available from the corresponding author on request. The underlying biological databases are public^6,7,8. The pipeline is proprietary to Mishale, Inc.; it is described here in sufficient detail to permit independent benchmarking against the same public outcomes.

Competing interests

W.M. is the founder of Mishale, Inc., which develops the system evaluated in this report. This declaration is made because the evaluation is of the author’s own product; the pre-registered design and full reporting of errors are intended to mitigate the resulting bias.

Author contributions

W.M. designed the study, built the pipeline, performed the evaluation and wrote the report.

References

1. 1. Frangoul, H. et al. CRISPR-Cas9 gene editing for sickle cell disease and β-thalassemia. N. Engl. J. Med. 384, 252–260 (2021).
2. 2. Mendell, J. R. et al. Single-dose gene-replacement therapy for spinal muscular atrophy. N. Engl. J. Med. 377, 1713–1722 (2017).
3. 3. Wilson, J. M. & Flotte, T. R. Moving forward after two deaths in a gene therapy trial of myotubular myopathy. Hum. Gene Ther. 31, 695–696 (2020).
4. 4. Nathwani, A. C. et al. Long-term safety and efficacy of factor IX gene therapy in hemophilia B. N. Engl. J. Med. 371, 1994–2004 (2014).
5. 5. Russell, S. et al. Efficacy and safety of voretigene neparvovec for biallelic RPE65-mediated inherited retinal dystrophy. Lancet 390, 849–860 (2017).
6. 6. Karczewski, K. J. et al. The mutational constraint spectrum quantified from variation in 141,456 humans. Nature 581, 434–443 (2020).
7. 7. Landrum, M. J. et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Res. 46, D1062–D1067 (2018).
8. 8. Vasilevsky, N. A. et al. Mondo: unifying diseases for the world, by the world. Preprint at medRxiv (2022).
9. 9. Wang, D., Tai, P. W. L. & Gao, G. Adeno-associated virus vector as a platform for gene therapy delivery. Nat. Rev. Drug Discov. 18, 358–378 (2019).
10. 10. Wong, C. H., Siah, K. W. & Lo, A. W. Estimation of clinical trial success rates and related parameters. Biostatistics 20, 273–286 (2019).
11. 11. Wouters, O. J., McKee, M. & Luyten, J. Estimated research and development investment needed to bring a new medicine to market. JAMA 323, 844–853 (2020).

Supplementary Table 1 — All 69 cases and dispositions

Every benchmarked case, its target, and the engine’s disposition. ✓ = passed the stratum’s pre-registered criterion; ✗ = failed. The three ✗ are negative controls the engine resolved to a real disease-associated gene instead of declining.