Autonomous discovery systems in functional genomics now design their own experiments: an agent picks which gene to perturb, a screen is run, the result updates the agent's model, and the cycle repeats. But these loops share a quiet flaw that bench biologists would never tolerate. They perturb in only one direction — knocking genes out, or turning them up — and they score effects at a single time point. In molecular biology, the accepted standard for calling a gene causal is a two-step test: delete the gene and watch the function disappear, then put a working copy back and watch the function return (complementation). Without that second step, a phenotype can be an artifact of off-target edits or disruption of neighboring genes. And without following the perturbation over time, a dramatic early effect can turn out to be a transient stress response that decays — the same "honeymoon period" trap documented across clinical and agricultural interventions, where 6-month cognitive gains partially decay by 9 months (ADL+ Study) and only a 3-year field trial could prove transgenic disease resistance was durable (Dale et al.).
This proposal builds a closed-loop discovery system that makes the rescue test and the time horizon first-class parts of the search itself, running on an execution substrate that makes every step of the loop verifiable. The agent does not screen knockouts and then, separately, confirm the winners. It plans paired experiments — knock down gene $g$, observe the trajectory, reintroduce $g$, observe again — and a gene only counts as a discovered hit if the model predicts, and the experiment confirms, both a loss of function that survives to a long horizon and a clean return to the wild-type state upon rescue. No existing system does this: PerTurboAgent runs a sequential Perturb-seq loop but only with unidirectional knockdowns scored at one time point; Pertura gates agents on correct interpretation of existing CRISPR screens but never designs new experiments or rescue constructs; FLeCS models temporal gene-network dynamics with learnable ordinary differential equations but is a passive, offline model with no loop; and SciAgent provides audited, containerized agent execution but only for reproducing static computational papers. Each supplies a component; none closes the combined loop.
Core Claim
The central claim is testable and specific: an acquisition function that requires predicted rescue consistency and temporal durability will discover more true causal genes per experiment than endpoint-scored, unidirectional loops, because it refuses to spend budget on two failure classes that current loops cannot distinguish from real hits — polar/off-target artifacts (which complementation exposes) and transient stress spikes (which long-horizon trajectories expose). Formally, for each candidate gene $g$ the surrogate produces a predicted knockout trajectory and a predicted complementation trajectory; the acquisition score is
$$a(g) = \sigma_H(g)\cdot \rho(g)\cdot u(g),$$
where $\sigma_H(g)$ is the predicted magnitude of the knockout effect that persists at horizon $H$ rather than at the assay's first time point, $\rho(g)$ is the predicted fraction of the effect removed when the gene is reintroduced (rescue consistency), and $u(g)$ is model uncertainty about the pair, so the loop still explores. A gene is promoted to "hit" only when the executed experiment confirms all three parts: durable loss, clean rescue, and agreement with prediction. The horizon $H$ is deliberately set beyond the window where standard screens stop, motivated by FLeCS's finding that compensatory regulatory ripples in K562 and A549 cells unfold over 72 hours and confound static differential expression.
Research Program
Three things get built. First, a temporal causal oracle to run the loop against, since no wet lab is available. We will fit FLeCS-style learnable ODE models — the technique from FLeCS, which models each gene's expression rate as regulated production minus degradation — to the Replogle CRISPRi essential-genome screens (1,075 perturbations in K562, 1,516 in RPE1, as used by LLMPERT and eVAE) and to the Norman/Weissman CRISPRa dataset (GSE133344) that Pertura used. Complementation is simulated by resetting the knocked-out gene's production term to wild-type at a chosen time. Crucially, we will seed the oracle with known decoys: genes whose knockout produces a large acute expression shift that the learned dynamics show decaying by horizon $H$, mimicking the transient stress responses FLeCS identified. Ground truth for a durable causal hit is defined by the oracle's long-horizon behavior, not by any early readout — so the benchmark itself encodes the limited-follow-up critique.
Second, the rescue-gated agent: an LLM-driven planner in the spirit of PerTurboAgent (which interleaves literature reasoning, statistical analysis, and surrogate fitting), extended so that every proposed experiment is a knockout–complementation pair with a declared horizon, and so that promotion of a hit requires the paired confirmation. Design transitions are gated by declarative preconditions in the style of Pertura, whose no-gate ablation showed agents otherwise commit unresolved boundary crossings such as confusing activation with inhibition.
Third, the execution substrate: every tool call, analysis script, and surrogate fit runs in a version-pinned OCI container with append-only JSONL provenance logs and a fresh-context LLM verifier, directly reusing the SciAgent architecture from Audit-Grade Harness for Agent-driven Scientific Computation Workflows, which catches silent fidelity degradation and fabricated outputs at an inference overhead of only 1.7% to 9.9% of run cost. This is not hygiene theater: EXP-Bench — the correct ID is EXP-Bench: Can AI Conduct AI Research Experiments? — found that nearly 30% of autonomous research failures come from environment and dependency errors, and end-to-end success was just 0.5%. A multi-round loop that silently rots halfway through is not a discovery system. We also adopt state-diff outcome checks from Agent-Diff, which showed pre/post environment snapshots catch unintended side effects that trace matching misses.
Baselines
The naive reference point is random paired selection: pick knockout–complementation pairs uniformly at random under the same budget. This is a genuinely dangerous baseline, not a strawman — LLMPERT reports that random selection is strong in pathway-specific Perturb-seq design and that uncertainty-based active learning often fails to beat it, and our Experiment 1 treats this as a real threat rather than assuming it away. Random pairs still get the rescue test and the long horizon, so this baseline isolates the value of the acquisition function specifically.
The strongest existing competitor is a PerTurboAgent-style loop: an agent with the same LLM planner, literature access, and surrogate machinery, selecting unidirectional knockdowns and scoring them at the assay endpoint. PerTurboAgent reports an average of 51.1 confirmed hits across eleven phenotypes on the genome-scale K562 Perturb-seq dataset, versus 30.7 for the strongest data-driven active-learning baseline and 24.1 for BioDiscoveryAgent, with a hit ratio of 0.44 — the best published number on sequential Perturb-seq hit discovery that Lacuna lists. We will re-implement this pipeline faithfully on our temporal oracle, give it the identical perturbation budget, compute, and tuning allowance, and score it two ways: with its native endpoint criterion, and generously re-scored against the oracle's long-horizon ground truth. Our system beats it only if it attains higher durable-hit precision (verified durable causal hits per perturbation budget) at matched budget — not merely more raw endpoint hits.
A third baseline is static-surrogate active learning in the LLMPERT style, an ensemble with a BALD-type acquisition on endpoint predictions, run under the same budget; LLMPERT achieved the lowest MSE AUC on all five tested K562 pathways under a 200-perturbation budget, making it the strongest pure-acquisition reference. Finally, two ablations isolate our key ingredients: the full system with the rescue gate removed (knockout-only acquisition, long horizon kept), and the full system with the horizon collapsed to the first time point (rescue gate kept). If the durability term does most of the work and the rescue term little — or vice versa — the ablations will say so, and the claim gets weakened accordingly.
Why not just…?
Why not just screen knockouts first and complement the top hits afterwards? This is the strongest objection, because post-hoc complementation is how biology labs actually work. The problem is selection bias, not verification: by the time complementation runs, the budget is already spent, and the loop's ranked list is contaminated with polar-effect artifacts and transient spikes. The knockout-only literature itself documents this — Ming Li et al. and Kaifeng Chen et al. needed reciprocal rescue precisely because phenotypic loss alone cannot exclude secondary mutations and polar disruption of adjacent transcription — and Pertura showed agents misattribute direction-of-effect without enforced gating. Experiment 1 settles it empirically: we will measure the fraction of the budget that a screen-then-complement pipeline spends on oracle decoys before any rescue test occurs, versus the rescue-gated loop. If the post-hoc pipeline matches our durable-hit precision at equal budget, the objection wins and the paired design is unnecessary overhead.
Why not just use a static endpoint surrogate? Because static snapshots cannot distinguish direct effects from compensatory ripples: FLeCS matched known biological mechanisms in 11 of 12 K562 knockdowns only once dynamics were modeled, and static differential expression was confounded by secondary responses. Our benchmark makes this quantitative — the decoys are constructed to look excellent at the endpoint and decay by horizon $H$ — so the endpoint-scored baselines will either avoid the decoys by luck or waste budget on them. The same trap is documented outside genomics: acute cardiac remodeling gains fail to predict survival (Sohal et al.), and short follow-ups missed atrial fibrillation recurrences beyond 12 months (Yang et al.). Honest caveat: if our decoy construction is unrepresentative of real transient perturbations, the durability advantage would be an artifact of the benchmark, which is why Experiment 2 reports results with decoy severity swept over a range.
Why not just use Conda environments and try/except logging? Because the failure mode is silent, not crashed. SciAgent documented agents fabricating plausible data when tool calls fail and silently substituting lower-fidelity models to dodge timeouts — failures standard logging does not catch — while the fresh-context verifier catches them for 1.7% to 9.9% of compute cost. EXP-Bench attributes roughly 30% of autonomous-research failures to environment and dependency misconfiguration, and Bencher exists precisely because uncontainerized benchmark stacks rot. That said, the container layer is infrastructure, not the scientific claim; we will report the loop-integrity comparison (undetected execution deviations per 100 rounds, containerized versus uncontainerized) as a separate contribution so it can be evaluated on its own.
Why not just use random selection, since it is a strong baseline? It may well be, on endpoint metrics; LLMPERT found BALD often loses to random. Our claim survives in weakened form if random pairs match our durable-hit count — it would then be the pairing and horizon, not the acquisition function, that carries the value, and the ablations are designed to detect exactly that decomposition. We commit in advance to reporting random-paired performance with the same prominence as our full system.
Proposed Experiments
Experiment 1 runs all systems — random pairs, static-surrogate active learning, the PerTurboAgent-style re-implementation, our full system, and both ablations — on the K562 oracle at budgets of 100, 200, and 400 paired experiments, with horizon $H$ at 72 hours per the FLeCS-validated regime. Primary metric: durable-hit precision, the fraction of promoted hits that the oracle confirms as both durable and cleanly rescued. Secondary metrics: decoy expenditure rate, rescue-prediction accuracy, and sample-efficiency curves. The decisive check is whether the full system's durable-hit precision exceeds the PerTurboAgent-style loop's at matched budget with non-overlapping confidence bands over seeds, and whether both ablations sit strictly between random pairs and the full system.
Experiment 2 sweeps decoy severity and horizon length to test robustness of the durability term, and cross-validates the oracle by holding out Replogle RPE1 perturbations, reporting whether surrogate dynamics learned on K562 transfer or fail — a transfer failure bounds how general any conclusion can be. Experiment 3 tests the execution substrate: run the full loop for 100 rounds containerized and uncontainerized with induced dependency drift, counting undetected fidelity deviations caught by the fresh-context verifier, benchmarked against the 1.7%–9.9% overhead SciAgent reports. As a realism anchor for the agent layer, we will also run the planner on the Norman/Weissman dataset under Pertura-style gating; no reported quantitative scalar exists for Pertura itself on Lacuna (its evaluation is qualitative gating versus a no-gate ablation), so this serves as a sanity reproduction rather than a scored comparison. If Experiment 1 shows rescue-gated acquisition does not beat random pairs on durable-hit precision, the honest conclusion will be that pairing and horizons — not intelligent acquisition — are what closed-loop genomics is missing, which would itself redirect where the field spends its engineering effort.
References
- ADL+: A Digital Toolkit for Multidomain Cognitive, Physical, and Nutritional Interventions to Prevent Cognitive Decline in Community-Dwelling Older Adults
- Transgenic Cavendish bananas with resistance to Fusarium wilt tropical race 4
- PerTurboAgent: A Self-Planning Agent for Boosting Sequential Perturb-seq Experiments
- Knowing When to Stop: Pertura for Graph-Enforced PI Gating in Perturb-seq Agents
- A scalable gene network model of regulatory dynamics in single cells
- Audit-Grade Harness for Agent-driven Scientific Computation Workflows
- LLM-guided acquisition for pathway-specific Perturb-seq design under experimental budgets
- Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction
- EXP-Bench: Can AI Conduct AI Research Experiments?
- Agent-Diff: Benchmarking LLM Agents on Enterprise API Tasks via Code Execution with State-Diff-Based Evaluation
- SalK/SalR, a Two-Component Signal Transduction System, Is Essential for Full Virulence of Highly Invasive Streptococcus suis Serotype 2
- Newly identified c-di-GMP pathway putative EAL domain gene STM0343 regulates stress resistance and virulence in Salmonella enterica serovar Typhimurium
- A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy
- A predictive model using left atrial function and B‐type natriuretic peptide level in predicting the recurrence of early persistent atrial fibrillation after radiofrequency ablation
- Bencher: Simple and Reproducible Benchmarking for Black-Box Optimization