When biologists want to know what a gene does in one specific cell type, they use a conditional knockout: flanking the gene with recombination sites so it is deleted only where a lineage-specific promoter is active, as when MyD88 is removed only from pain-sensing neurons TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation or MED1 only from macrophages MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation. Doing this at scale means Perturb-seq — pooled CRISPR screens read out by single-cell sequencing — and the field now has active-learning methods that decide which knockouts to test next. But every one of these methods treats the laboratory as having a single scalar budget: "pick 200 perturbations." A real screen does not work that way. It consumes several physically independent resources at once: viable cells of each lineage (acutely scarce in primary tissue and conditional systems, where each Cre driver gives access to a different, limited pool), sequencing reads, microfluidic droplet capacity, and guide-RNA library slots. Cloud computing hit exactly this problem a decade ago and solved it by treating jobs and machines as multi-dimensional vectors rather than scalars A Multi-Dimensional Resource Scheduling Strategy Based on Multilateral Complementarity.
The stakes are not just efficiency but interpretability. The harmonized scPerturb collection shows that a perturbation is statistically distinguishable from noise only if two independent physical minimums are met jointly: at least 300 cells per perturbation and at least 1,000 UMI counts per cell scPerturb: Harmonized Single-Cell Perturbation Data. A scalar-budget design can legally spend everything on reads and deliver batches whose measurements fall below one of these thresholds — data that is uninterpretable no matter how cleverly the genes were chosen. My proposal is to import the cloud scheduler's view wholesale: model each candidate conditional knockout as a job with a predicted multi-dimensional demand vector, and design each experimental batch as a vector-packing problem in which perturbations with opposing resource biases are deliberately paired so that no dimension is left stranded.
Core Claim
I claim that an active design policy which (i) predicts a per-perturbation demand vector from prior screening data, (ii) selects candidates by expected information gain, and (iii) composes batches by multilateral complementarity — pairing, say, high-read-depth/low-cell-yield knockouts (large-effect master regulators) with low-depth/high-yield knockouts (subtle metabolic effectors) in the same sequencing lane — will resolve strictly more perturbations per unit of physical resource than the strongest scalar-budget active learners, and that most of the gain survives even when the selection of candidates is held fixed and only the packing changes.
Formally, for a batch $B$ and a capacity vector $\vec{C}$ spanning $d$ dimensions (per-lineage cells, total reads, droplet channels, guide-pool slots), we require
$$\sum_{i \in B} \vec{d}_i \preceq \vec{C}, \qquad \vec{d}_i \succeq \vec{d}^{\min},$$
where $\vec{d}_i$ is the predicted demand of perturbation $i$ and $\vec{d}^{\min}$ encodes the scPerturb detectability floor (300 cells, 1,000 UMIs per cell, and analogues). The objective is the expected reduction in the surrogate model's prediction error on unscreened perturbations, which is approximately submodular in the set of measured perturbations — the property Ghassami et al. exploited to justify greedy selection with approximation guarantees Near-Optimal Multi-Perturbation Experimental Design for Causal Structure Learning. Their work is the closest methodological ancestor: it formalizes knockout selection as budget-constrained combinatorial optimization with submodular objectives, but its budget is a single scalar count of interventions, with no per-dimension feasibility, no detectability floors, and no notion of packing complementary demands. That gap is precisely what this proposal fills.
Research Program
Three components. First, demand-vector estimation: using the scPerturb harmonized collections scPerturb: Harmonized Single-Cell Perturbation Data, including the Replogle genome-scale subset of roughly 10,000 perturbations and up to 2.5 million cells, I will fit light predictive models mapping a perturbation's identity and prior annotations to its expected cell yield and read-depth demand. This mirrors how ATLAS predicts job runtimes from submit-time features to guide cluster scheduling ATLAS: Alibaba Dataset and Benchmark for Learning-Augmented Scheduling, and inherits its known fragility: ATLAS documents that tail objectives degrade sharply under prediction error, so the packing layer must be robust to mispredicted demands, not just accurate on average.
Second, the policy itself: a two-level scheduler. The outer level selects candidates by an information-gain acquisition function over a GEARS-style surrogate Predicting transcriptional outcomes of novel multigene perturbations with GEARS, which predicts transcriptional outcomes of unseen perturbations and can therefore score untested candidates — I reuse it as the predictor, not as a design policy, since GEARS itself contains no resource-aware experiment-selection mechanism. The inner level solves the vector-packing problem with a complementarity-aware greedy in the spirit of the MQP family, which pairs workloads with opposing resource biases and improved node utilization by 2–5% over scalar-style heuristics on the Google Cluster Dataset A Multi-Dimensional Resource Scheduling Strategy Based on Multilateral Complementarity. The key adaptation is that in a sequencing lane the "capacity" is reads and droplets per run, and the payoff of a packed batch is the count of perturbations clearing the $\vec{d}^{\min}$ floor, not raw utilization.
Third, an evaluation infrastructure that no existing paper provides: a capacity-calibrated replay simulator built from scPerturb datasets, in which each perturbation's empirical cell count and UMI distribution define its true demand vector, a chosen $\vec{C}$ emulates a real instrument and tissue supply, and the E-test from scPerturb decides whether a simulated measurement counts as resolved. A heterogeneous-lineage mode stratifies cells by type to emulate conditional knockout settings, where each lineage's availability is its own capacity dimension — the regime that IterPert explicitly leaves open, since it was evaluated only on uniform K562 leukemia cultures Sequential Optimal Experimental Design of Perturbation Screens Guided by Multi-modal Priors.
Baselines
The naive reference point is uniform random selection of perturbations at matched budget. This must be taken seriously rather than dismissed: the LLM-guided acquisition study found that uncertainty-based methods frequently lose to random selection in high-dimensional biological spaces, calling random a resilient baseline that active learners struggle to defeat LLM-guided acquisition for pathway-specific Perturb-seq design under experimental budgets. Random will be run inside the same simulator with the same capacity vectors, and any method failing to beat it under multi-resource constraints is disqualified.
The strongest existing competitor is IterPert, which fuses six biological prior modalities into a single kernel to drive sequential acquisition, and which reached target predictive accuracy using only one-third of the perturbations required by the next-best active learner on a CRISPRi screen of 2,058 essential genes in leukemia cells Sequential Optimal Experimental Design of Perturbation Screens Guided by Multi-modal Priors. It is the strongest because it is the only method demonstrating large, published sample-efficiency gains in exactly this sequential-budget regime. To be fair, I will run IterPert with its own priors and tuning budget inside my simulator; when it selects a set that violates the capacity vector, I will give it the most favorable repair (greedy substitution of feasible candidates) rather than counting violations against it. Beating IterPert means resolving more perturbations per unit of total resource spend and reaching a lower downstream surrogate error at matched spend — not merely at matched perturbation counts, since resource cost per perturbation differs and that difference is the whole point.
The LLM re-ranking pipeline of LLM-guided acquisition for pathway-specific Perturb-seq design under experimental budgets, which achieved the lowest MSE AUC across all five MSigDB Hallmark pathways in K562 under a 200-perturbation cap, is the second comparator, run identically. The most important baseline, however, is an ablation of my own method: the same acquisition function and the same surrogate, but with the capacity vector collapsed into a single scalar (a normalized weighted sum of dimensions) so the inner layer degenerates to a one-dimensional knapsack. Any gap between this ablation and the full method isolates the multi-dimensional packing ingredient from everything else, and if that gap is negligible the proposal's central claim is refuted. Finally, a within-round mixed-integer programming solver serves as an oracle: VMR²L's authors found MIP needed 50.55 minutes where their learned scheduler needed 1.1 seconds Towards VM Rescheduling Optimization Through Deep Reinforcement Learning, so the oracle bounds how much of the achievable packing gain greedy leaves on the table.
Why not just…?
Why not just convert everything into a single dollar budget? This is the most natural objection, and the evidence against it is direct. Scalarizing multi-resource allocation provably strands capacity: the placement-optimization literature shows that optimizing a single dominant cost exhausts one resource while leaving others idle Automating Enterprise Application Placement in Resource Utilities, and the MQP work quantifies the recoverable waste at 2–5% utilization even in commodity clusters A Multi-Dimensional Resource Scheduling Strategy Based on Multilateral Complementarity. In biology the failure is worse than waste: because detectability requires jointly meeting 300 cells and 1,000 UMIs per cell scPerturb: Harmonized Single-Cell Perturbation Data, a scalar budget can purchase measurements that are individually uninterpretable. My scalar-collapse ablation is designed precisely to test this objection empirically rather than argue it away.
Why not just run IterPert or an LLM agent with a smaller budget? Because neither method checks per-dimension feasibility at all. Under real vector constraints their selected sets can be physically unmeasurable — a perturbation whose cell demand exceeds the lineage supply is not a "smaller experiment," it is a nonexistent one. There is also documented brittleness: LLM guidance degraded sharply on RPE1 cells in the Apoptosis pathway, and stripping Gene Ontology annotations actually improved performance in some settings, showing that richer generic priors can bias selection toward globally famous genes rather than cell-type-specific regulators LLM-guided acquisition for pathway-specific Perturb-seq design under experimental budgets. That said, this is a genuine threat: if IterPert's repaired selections match my resolved-perturbation count at matched spend, the multi-resource view adds nothing, and the experiments are constructed so that this outcome would be visible and reportable.
Why not just solve the packing exactly with a mixed-integer solver each round? Honestly, at a few hundred candidates per round this might work, and that is why MIP appears as an oracle rather than being dismissed. Two things nonetheless favor greedy. The candidate space is large — the genome-scale Replogle screens cover roughly 10,000 perturbations — and Ghassami et al. show exact search over composite interventions grows doubly exponentially Near-Optimal Multi-Perturbation Experimental Design for Causal Structure Learning. More importantly, there is a moving-target parallel: VMR²L identified a 5-second elbow beyond which an exact solution describes an obsolete system state Towards VM Rescheduling Optimization Through Deep Reinforcement Learning, and biological samples drift across rounds through passage number and batch effects, so a policy that spends hours optimizing each round targets a stale cell-state distribution. If the MIP oracle closes the gap cheaply at realistic scale, I will say so and recommend it.
Why not just screen everything at once in one giant batch? Sequencing economics forbid it in primary and conditional settings, and the efficiency cost is documented: sequential design reached the same accuracy with one-third the perturbations of non-sequential baselines Sequential Optimal Experimental Design of Perturbation Screens Guided by Multi-modal Priors, while combinatorial knockouts grow as $O(N^2)$ pairs and $O(N^3)$ triplets, making exhaustive profiling hopeless Predicting transcriptional outcomes of novel multigene perturbations with GEARS.
Proposed Experiments and Decisive Checks
The first experiment measures the headline quantity: perturbations resolved per fixed multi-dimensional budget, in the capacity-calibrated replay of the Replogle-scale scPerturb data, against random, BALD-style uncertainty sampling, IterPert, the LLM pipeline, and the scalar-collapse ablation, all at matched spend. Downstream value is measured by retraining a GEARS-style surrogate on the resolved data and evaluating mean squared error on the top-20 differentially expressed genes — the metric on which GEARS itself reports a 30–50% improvement over prior predictors Predicting transcriptional outcomes of novel multigene perturbations with GEARS — plus zero-shot accuracy on the 131 two-gene perturbations of the Norman et al. dataset, where GEARS reports a 53% improvement over prior models. The second experiment isolates packing from selection: fix the selected set and vary only its composition into lanes; if complementarity-aware assignment alone raises the resolved count, Gap 2 is confirmed independent of any acquisition function. The third experiment runs the heterogeneous-lineage mode with per-lineage cell caps, emulating conditional knockout across multiple Cre drivers, since that is the regime where per-lineage scarcity binds hardest and where existing active learners were never evaluated.
The most important decisive check comes first, as a gate: measuring the actual dispersion of demand vectors across perturbations in scPerturb. If per-perturbation cell yields and read demands are nearly uniform in existing datasets — which is plausible, since genome-scale screens were designed for uniform loading — then complementarity has little to exploit and the gains will concentrate in the constrained primary-tissue simulations rather than the uniform-cell-line replays. I would report that boundary honestly: the proposal's value is then a claim about where screening is going (conditional, primary, multi-lineage) rather than where it has been.
References
- TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation
- MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation
- A Multi-Dimensional Resource Scheduling Strategy Based on Multilateral Complementarity
- scPerturb: Harmonized Single-Cell Perturbation Data
- Near-Optimal Multi-Perturbation Experimental Design for Causal Structure Learning
- ATLAS: Alibaba Dataset and Benchmark for Learning-Augmented Scheduling
- Predicting transcriptional outcomes of novel multigene perturbations with GEARS
- Sequential Optimal Experimental Design of Perturbation Screens Guided by Multi-modal Priors
- LLM-guided acquisition for pathway-specific Perturb-seq design under experimental budgets
- Towards VM Rescheduling Optimization Through Deep Reinforcement Learning
- Automating Enterprise Application Placement in Resource Utilities