Structure-Resolved Conditional Knockout Prediction in an Audit-Grade Harness

When biologists want to know what a gene does in one specific tissue, they build a "conditional knockout": they flank a chosen exon with recombinase sites so that, in a chosen cell type, that exon is cut out of the gene. The unstated assumption in almost every such experiment is that the result is a clean null — no protein, no function. In reality, the excision often leaves a shortened but folding, partially working protein, and the difference between "no protein" and "a mangled protein missing one domain" can be the difference between a mild phenotype and a disastrous one, or between embryonic death and a survivable adult defect. No computational tool today predicts which outcome a given floxed design produces, because sequence-level knockout predictors ignore the residual protein's three-dimensional structure, and structure modelers focus on point mutations rather than whole-exon amputations. Meanwhile, the pipeline needed to make this prediction — splicing models, structure predictors, and phenotype predictors from three separate software ecosystems — is exactly the kind of fragile, multi-tool chain that silently rots and produces confidently wrong answers.

This proposal builds a predictor that takes a conditional knockout design (gene, floxed exons, target tissue) and outputs the residual protein's structure-aware fate and a tissue-conditioned phenotype prediction, all executed inside a version-pinned, audit-logged container harness. Protein structure modeling is the core mechanism; conditional knockout is the application regime; reproducible execution is the methodological substrate without which the predictions cannot be trusted or even reproduced six months later.

Core Claim

Cre-mediated excision of floxed exons produces a predictable, internally truncated protein isoform whose residual structure — not merely its sequence — determines whether the allele behaves as a null, a hypomorph, or something neomorphic. The frame arithmetic is simple: with $\ell_e$ the length in nucleotides of floxed exon $e$,

$$\Delta = \Big(\textstyle\sum_{e \in E_{\text{flox}}} \ell_e\Big) \bmod 3,$$

where $\Delta = 0$ yields an in-frame internal truncation that keeps every domain outside the excised span, and $\Delta \neq 0$ forces a frameshift, a premature stop, and probable nonsense-mediated decay. The claim is that for the $\Delta = 0$ case — the case every knockout designer is currently flying blind on — a model that sees the truncated isoform's retained catalytic pockets and binding interfaces, conditioned on which interaction partners the target tissue actually expresses, predicts the observed phenotype class better than any sequence-only model that treats the knockout as a binary gene deletion.

Research Program

The build is a three-stage cascade, provisionally called FloxStruct. Stage one computes the transcript fate: given floxed exon coordinates, apply the frame calculus and splice-outcome logic in the spirit of SkipGuide, which chains DNA-repair and splice predictors to estimate exon skipping frequencies and reached a Pearson correlation of $r = 0.67$ and an AUC of 0.84 against an in vitro reporter library Machine learning based CRISPR gRNA design for therapeutic exon skipping. We reuse that machinery's shape — nucleotide-level predictors deciding what the mRNA looks like — but extend it from splice-disruption to Cre-recombination geometry. SkipGuide stops at the transcript; we do not.

Stage two is the new science: structural modeling of the in-frame truncated isoform. Model the engineered sequence with structure prediction, align it to the wild-type fold, and extract residual-structure features — which domains survive, whether the catalytic pocket is intact or narrowed (the kind of geometric change documented in the HPRT1 case, where a single valine-to-alanine substitution physically narrows the enzyme's concave catalytic pocket Case report: Whole exome sequencing identifies a novel variant in the HPRT1 gene in a male with developmental delay), and which binding surfaces are preserved. This stage exists because protein language models are known to be weak precisely here: ProteinGym's benchmark of more than 2.7 million mutated sequences shows alignment-based models and standard protein language models consistently fail on insertion and deletion assays, of which ProteinGym contains 66, because their architectures cannot handle length changes ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. No reported result found on Lacuna for any model's Spearman correlation on those indel assays specifically — that gap is the opening this stage attacks. A second motivation comes from the multi-modal variant-effect work of Kim & Ha, which showed that a frozen protein language model (ESM3) alone detects only 15.4% of severe loss-of-function variants, rising to 75.6% only when DNA-level context is fused in Mechanistic Synergy in Multi-Modal VEP: DNA Context Complements PLMs under Biophysical Constraints. Their finding constrains our design: protein-only representations are optimistically biased, so our residual-structure features must be combined with genomic context, not used alone.

Stage three conditions the prediction on tissue. Sequence-to-phenotype models today predict organism-wide phenotypes from gene sequences; GenePheno, the strongest of these, maps raw DNA and Gene Ontology terms to Mammalian Phenotype Ontology labels via an interpretable functional bottleneck, beating the next best method by +1.48 Fmax and +7.97 AUC on the MPO benchmark GenePheno: Interpretable Gene Knockout-Induced Phenotype Abnormality Prediction from Gene Sequences. But GenePheno models a knockout as a binary whole-gene deletion, while conditional biology shows the same gene ablated in different tissues produces wildly different outcomes: endothelial deletion of miR-21 causes thoracic aortic remodeling with diastolic blood pressure falling to roughly 72 mmHg versus 87 mmHg in wild type and opening angles rising to 161.3° versus 119.1° Induction of Thoracic Aortic Remodeling by Endothelial-Specific Deletion of MicroRNA-21 in Mice, whereas muscle-specific ablation of $\gamma_{cyto}$-actin isolates a local dystrophy without systemic collapse Prins et al.. FloxStruct's stage three weights the residual-structure features by tissue transcriptomes — which paralogous partners and interaction surfaces the target cell type expresses — using the kind of tissue-resolved RNA-seq quantification that SnakeAltPromoter demonstrates is reproducible and concordant with gold-standard CAGE data (ProActiv at 78.6% precision, Pearson $r = 0.83$) SnakeAltPromoter Facilitates Differential Alternative Promoter Analysis.

The execution substrate matters because this cascade chains tools with incompatible dependencies across three ecosystems. We package FloxStruct as version-pinned containers with append-only JSONL provenance logs and a fresh-context verifier, following SciAgent, which showed that an agent can map 1,837 BRCA1 variants to structural coordinates at a 1.000 mapping rate while an audit layer catches execution deviations a plain run misses, at a verifier overhead of only 1.7% to 9.9% of total compute Audit-Grade Harness for Agent-driven Scientific Computation Workflows. Our harness differs from SciAgent's in scope, not mechanism: SciAgent audits a static coordinate-mapping task on one gene, whereas we audit a prediction cascade whose stages can each silently degrade — the exact failure mode (silent numerical degradation, tool hallucination) that audit harnesses were built to catch.

Baselines

The naive reference point is the "assume null" practice that experimentalists use today: every conditional knockout is treated as a complete gene deletion, and the expected phenotype is read off the global knockout literature. Concretely this means running GenePheno on the unmodified wild-type sequence and taking its MPO output as the prediction. It is the strongest competitor on the question as currently asked, because it is what actually informs floxed-line decisions; beating it means demonstrating that residual-structure information changes the recommended interpretation for in-frame designs, since on frameshift ($\Delta \neq 0$) designs it should be right by construction.

The strongest existing method is GenePheno given its best possible input: the post-excision truncated coding sequence, so it sees the edit but still reasons from sequence alone GenePheno. It holds the top reported results on the MPO benchmark, the exact ontology our phenotype output targets, which makes it the competitor to beat rather than a strawman. It will be run with the same training data, the same MPO evaluation split, and the same hyperparameter search budget as FloxStruct, on identical hardware. Beating it means higher Fmax and AUC on a conditional-knockout evaluation set, and specifically a large margin on the in-frame hypomorph cases where sequence alone cannot distinguish "domain gone" from "domain present but pocket collapsed."

The key ablation strips out the structure stage: an ESM3 embedding of the truncated isoform sequence, fused with the same tissue context, predicting the same phenotypes. This isolates whether the geometry — pockets, interfaces, retained domains — adds anything beyond simply knowing the truncation. If this ablation matches FloxStruct, the structural machinery is decoration and the proposal's core claim is wrong; the experiment is designed so that this outcome would be visible and reportable.

For the harness component, the baseline is the same cascade run unaudited in a plain Conda environment, SciAgent-style Audit-Grade Harness. We will seed known failures — a swapped exon coordinate, a silently downgraded tool version, a fabricated intermediate file — and measure detection rates against identical compute.

Why not just…?

Why not just run ESM3 or AlphaFold on the truncated sequence? This is the most serious alternative, and it is exactly our ablation, so the proposal is built to answer it rather than argue around it. The evidence against it is concrete: protein language models fail on length-changing edits across ProteinGym's 66 indel assays ProteinGym, and ESM3 alone reaches only 15.4% sensitivity on loss-of-function variants, needing DNA context to reach 75.6% Mechanistic Synergy. AlphaFold-style predictors give a static fold but no calibrated severity ranking, and they struggle outside well-ordered globular regions AlphaFold prediction of structural ensembles of disordered proteins. Experiment 1 below settles the question empirically, and if the ablation wins we will say so.

Why not just use GenePheno and skip structure entirely? GenePheno is genuinely strong — it is the MPO state of the art and our hardest baseline — but it cannot separate a true null from an in-frame truncation that preserves scaffolding domains while ablating a catalytic pocket, because it never sees either. The biology says this distinction carries real phenotype consequences: domain-targeted conditional ablation of Bnc1 bypassed embryonic lethality and revealed an adult corneal-epithelium role that whole-gene assumptions would have missed Hung Tseng. Still, GenePheno fed the truncated sequence is a real threat, which is why it is the primary comparison rather than a footnote.

Why not just use Snakemake and Conda instead of containers plus an audit layer? For the deterministic, non-agentic pipeline, honest answer: Snakemake with pinned environments largely suffices for reproducibility, and we will ship the cascade that way too. The audit layer earns its cost only where execution is agentic or where silent degradation is plausible; SciAgent's photonics task returned an INSUFFICIENT audit verdict until child-process logs were recursively exposed, showing that surface-level logging misses failures Audit-Grade Harness, and Conda environments break under side-by-side evaluation of tools with conflicting system libraries while containerized evaluation does not Bencher: Simple and Reproducible Benchmarking for Black-Box Optimization. Hermetic isolation also prevents residual host state from leaking between runs A Qualitative Study on the Sources, Impacts, and Mitigation Strategies of Flaky Tests. We scope the audit component to the seeded-failure detection experiment rather than claiming it is always necessary.

Why not just run a CRISPR exon-skipping screen in vitro? Because the screens run in uniform culture — SkipGuide's own authors note that splicing regulation and functional phenotypes are severely tissue-specific, so mouse embryonic stem cell results do not transfer to adult tissues Machine learning based CRISPR gRNA design for therapeutic exon skipping — and because the phenotypes that matter (blood pressure, aortic mechanics) exist only in vivo, where each conditional line costs 12–24 months. The point of FloxStruct is to prioritize which of those expensive lines to build; it complements, not replaces, the screens.

Proposed Experiments

Experiment 1 tests whether residual structure beats sequence on length-changing edits, using ProteinGym's 66 indel assays as the evaluation set ProteinGym. FloxStruct's structure-conditioned scoring is compared against ESM3-on-truncated-sequence and against the supervised ProteinNPT, which reported a Spearman correlation of 0.613 on ProteinGym supervised fitness prediction. Because no reported result found on Lacuna for any model's indel-assay-specific Spearman, establishing that number for ESM3 is itself a contribution, and the decisive check is whether structure features raise it materially.

Experiment 2 tests tissue-conditioned phenotype prediction on a benchmark we construct from published conditional knockout studies — FloxBench — with one entry per published gene, floxed-exon, Cre-driver, observed-outcome triple drawn from the conditional-knockout literature represented in the file, including the endothelial miR-21 aortic remodeling case Induction of Thoracic Aortic Remodeling, endothelial Dicer1 loss driving angiosarcoma Canonical microRNA loss drives tumor development, macrophage MED1 deletion Zhang et al., megakaryocyte Slc35a1 deletion Ma et al., and neuronal MyD88 knockout Liu et al.. Two metrics: allele-class accuracy (null versus hypomorph, where the ground truth is whether residual protein function was observed) and phenotype-term prediction scored with the MPO Fmax and AUC used by GenePheno. We are explicit that FloxBench will start small — on the order of a dozen entries — so we will report per-entry reasoning and calibration rather than claiming statistical significance, and design the benchmark to grow as new conditional studies appear.

Experiment 3 tests the harness. Seeded failures are injected into agentic runs of the full cascade across containerized environments, and detection rate and verifier overhead are compared against SciAgent's reported figures — a 1.000 mapping rate on its genetics task and verifier overhead of 1.7% to 9.9% of total compute Audit-Grade Harness. A secondary check measures execution drift in long cascades, the failure mode observed in containerized proteomics agents SP-Mind: An Autonomous Reasoning Agent for Spatial Proteomics Analysis. The decisive criterion is that the audited harness catches every seeded failure the unaudited Conda baseline misses, at overhead within SciAgent's reported band.

References

  1. Machine learning based CRISPR gRNA design for therapeutic exon skipping
  2. Case report: Whole exome sequencing identifies a novel variant in the HPRT1 gene in a male with developmental delay
  3. ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design
  4. Mechanistic Synergy in Multi-Modal VEP: DNA Context Complements PLMs under Biophysical Constraints
  5. GenePheno: Interpretable Gene Knockout-Induced Phenotype Abnormality Prediction from Gene Sequences
  6. Induction of Thoracic Aortic Remodeling by Endothelial-Specific Deletion of MicroRNA-21 in Mice
  7. Skeletal Muscle-Specific Ablation of γcyto-Actin Does Not Exacerbate the mdx Phenotype
  8. SnakeAltPromoter Facilitates Differential Alternative Promoter Analysis
  9. Audit-Grade Harness for Agent-driven Scientific Computation Workflows
  10. AlphaFold prediction of structural ensembles of disordered proteins
  11. Bencher: Simple and Reproducible Benchmarking for Black-Box Optimization
  12. A Qualitative Study on the Sources, Impacts, and Mitigation Strategies of Flaky Tests
  13. Canonical microRNA loss drives tumor development implicating therapeutic efficacy of enoxacin in angiosarcoma
  14. MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation
  15. Slc35a1 deficiency causes thrombocytopenia due to impaired megakaryocytopoiesis and excessive platelet clearance in the liver
  16. TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation
  17. SP-Mind: An Autonomous Reasoning Agent for Spatial Proteomics Analysis