Automated planners get fast by ignoring things. When a planner restricts itself to the "primary effect" of each action — using a drug only for what it is prescribed to do, and discarding its collateral effects — the number of possibilities it must consider shrinks so much that search time drops exponentially (Fink and Yang). Biomedical translation fails for exactly the mirror-image reason: preclinical screens isolate a compound's primary target engagement and discard its collateral effects on the immune system, vasculature, and host physiology, and it is precisely those discarded effects that sink candidates when they reach living organisms (Zou et al., Scott, Xu et al.). This is not a loose analogy; it is the same abstraction operator producing success in one field and attrition in the other.
The proposal here takes that tension seriously instead of resolving it in advance. I will build a search algorithm for sequential, timed cancer drug regimens — staggered multi-compound schedules, not static cocktails — in which the primary-effect abstraction drives a fast satisficing search over the in-vitro surrogate, but the pruned side-effect branches are not thrown away. They are stored lazily, and re-expanded only when a completed candidate regimen is rejected by an in-vivo viability model, in direct analogy to the FindAnotherMember routine that lets temporal planners recover pruned path histories on demand (Coles et al.). The claim is that this design gets the exponential speedup of primary-effect search while paying the accuracy cost only where the in-vivo layer says it matters, and that the result is a drug-regimen planner whose candidates fail the preclinical-to-clinical translation check far less often than candidates ranked or predicted by existing static models.
Core Claim
No existing system searches over sequential therapeutic regimens at all, let alone with a principled optimality-efficiency mechanism. The closest works are static: DeepSADR transfers single-drug response prediction from in-vitro cell lines to patient tumors but performs no search; CellVS-Net maps compounds to cellular gene-network rewiring but explicitly disclaims in-vivo or clinical prediction; the sRGES framework of Chen et al. ranks compounds by exhaustive correlation against precomputed profiles and validates one prediction in mouse xenografts, but evaluates single-point interventions through brute force; and the combination-prediction work of Jeon et al. and Xu et al. handles only simultaneous pairs, in vitro, via supervised models or static network diffusion. My proposal differs on all three axes at once: it searches (rather than ranks or classifies), over sequences (rather than single drugs or pairs), under an explicit satisficing budget with a lazy recovery mechanism (rather than exhaustive or static evaluation). The side-effect paradox identified above — prune for speed, but pruning is what kills translation — becomes the design principle rather than a caveat.
The Planner: Lazy Side-Effect Expansion
The planner operates on a state space of cellular gene-dependency networks. Following CellVS-Net, an amortized surrogate maps a compound's molecular structure and the current cellular context to the precision parameters of the rewired dependency network, so applying a drug is a state transition in network space (Ellington et al.). Each action carries a primary effect — its contribution to reversing the disease gene-expression signature, in the spirit of sRGES (Chen et al.) — and side effects: off-target dependency rewiring and host-context risk terms drawn from a toxicity/efficacy model fine-tuned on patient data in the manner of DeepSADR's frozen-encoder transfer (Zhang et al.). To suppress the combinatorial explosion of moving several compounds at once, the search uses operator decomposition — one compound per step, one step per decision epoch — borrowing from simultaneous sampling-and-search in multi-robot planning (Okumura and Défago).
The efficiency mechanism is a biological metastate. During search, two partial regimens that achieve the same primary-effect profile (the same cumulative disease-signature reversal at the same step) are treated as equivalent, and only one member is expanded, exactly as the temporal planner of Coles et al. collapses weakly equal states differing only in path history. The full regimen score is
$$\text{score}(r) \;=\; D(r) \;-\; \lambda\, R_{\text{vivo}}(r), \qquad |r| \;\le\; C\,|r^{*}|,$$
where $r$ is a regimen, $D(r)$ its disease-signature reversal, $R_{\text{vivo}}(r)$ the in-vivo viability penalty from the patient-fine-tuned model, $\lambda$ a risk-aversion parameter, and the length bound comes from primary-effect planning theory, which guarantees plans at most a factor $C$ longer than optimal (Fink and Yang). When a completed candidate fails because $R_{\text{vivo}}$ is large — the planner's analog of a downstream temporal inconsistency — the lazy layer identifies which collapsed metastate the failure traces back to and re-expands its pruned siblings under the full primary-plus-side-effect model. The hypothesis, supported by the metastate results where lazy expansion raised coverage from 164 to 189 solved problems with orders-of-magnitude fewer node expansions (Coles et al.), is that most regimens never need re-expansion, so the average cost stays near the satisficing floor while translational failures are caught. A secondary ingredient, held to one ablation, is treating pathway granularity — how coarsely the 1,000-odd landmark genes are clustered into functional groups — as a budget-dependent abstraction parameter rather than the fixed 13-pathway compression of DeepSADR, refined lazily only when a coarse-level plan stalls, in the manner of hierarchical pathfinding abstractions (Sturtevant).
Baselines
The do-nothing reference point is brute-force in-vitro ranking: score every compound by disease-signature reversal against precomputed LINCS L1000 profiles and take the top one, which is exactly how Chen et al. generated and xenograft-validated pyrvinium pamoate (Chen et al.). This baseline costs almost nothing and has produced an in-vivo-validated hit, so any planner must beat it on more than speed: the criterion is higher recall of clinically validated drugs and regimens at a matched count of surrogate-model evaluations.
The strongest existing competitor is a decoupled pipeline assembled from the best static components: an Extremely Randomized Trees synergy model, which on the O'Neil dataset achieved a Pearson correlation of 0.738 for synergy prediction and an F1 of 0.954 for binary synergy classification, placing the most synergistic combination in the top ten for 28 of 31 cell lines (Jeon et al.); a Random-Walk-with-Restart diffusion screen over signaling networks, which recovered 76% of validated combinations within its top 30% of predictions (Xu et al.); and DeepSADR for single-drug clinical response, which achieved an AUC of 0.957 for Sorafenib against 0.811 for the next best method (Zhang et al.). It will be run as a shortlist-then-optimize pipeline: rank compounds and pairs with these models, then greedily sequence the top-ranked items. It gets the same data (DepMap, LINCS L1000, TCGA), the same surrogate-evaluation budget, and the same tuning effort as the planner. Beating it means two things: regimens with higher TCGA-grounded clinical fidelity at equal budget, and performance that does not collapse when regimen length grows past two or three steps, where the static models have no defined output at all.
Two ablations isolate the key ingredient. The first removes lazy re-expansion, leaving a pure primary-effect satisficing search; this should retain the speed but, by the core claim, produce regimens that the in-vivo model rejects at a high rate. The second replaces lazy expansion with always-on side-effect tracking, an effectively exact search; this should match the fidelity of the full planner while blowing past the evaluation budget, reproducing in this domain the gap between A*'s nearly 100,000 node expansions and the abstraction's roughly 1,000 on long paths (Sturtevant). The proposal succeeds only if the full planner sits near the exact search's fidelity at near the satisficing search's cost.
Why not just…?
Why not just train a supervised model on combination data and extend it to sequences? This is the most reasonable objection, and DeepSADR shows how much structure such models can capture — but there is no supervised signal to extend. The O'Neil dataset that powers the strongest synergy models contains 583 drug pairs and 16,575 samples, all simultaneous, all in vitro (Jeon et al.); no dataset of timed sequential regimens with patient outcomes exists in the file, and DeepSADR itself transfers only 555 patient samples across 5 drugs. Supervised sequential prediction has no labels; a model-based search over surrogate transitions is the only route to sequences. If sequential-schedule data does emerge, this becomes a real threat, and experiment E3 below is designed to detect it by comparing the planner against the best achievable static model on the subset of tasks where both are defined.
Why not just brute-force rank everything by in-vitro score, as sRGES already does? sRGES earned a Spearman correlation of 0.61 with preclinical efficacy in liver cancer and produced an in-vivo-validated compound (Chen et al.), so this is not a strawman. But exhaustive ranking is single-point: regimens of length three drawn from 12,442 LINCS compounds outnumber the 16,575 physical assays needed for 583 pairs by many orders of magnitude, and pure in-vitro maximization is precisely the local-optimum failure mode — the evidence that cheap structure-blind scoring fails is that molecular fingerprints achieved a Hits@1 of 0.0357 on drug-disease retrieval, barely above the random baseline of 0.0179, where a system-level model reached 0.1250 (Ellington et al.). Brute force cannot reach sequences and cannot see the host context; the planner addresses both.
Why not just decouple: pre-screen a shortlist, then run exact optimization on it? This is the genuinely dangerous alternative, and I will not argue around it. Okumura and Défago document the "density dilemma" for decoupled pipelines — a sparse roadmap contains no solution, a dense one makes the subsequent search intractable — with decoupled baselines capped near 80% success where integrated sampling-and-search reached near 100% (Okumura and Défago); that is evidence for integration, but a well-chosen 50-compound shortlist from DeepSADR-grade models might sidestep the dilemma entirely. The decisive check in E4 therefore runs this exact pipeline at matched budget as a first-class competitor rather than an afterthought. If the shortlist-plus-exhaustive pipeline matches the planner's translational fidelity at lower cost, the honest conclusion is that lazy expansion is unnecessary at current library scales, and the planner's value would rest on scaling beyond them — the experiments are built to settle this either way.
A fourth skeptic will demand exact optimality outright. The planning literature is unambiguous that this is infeasible at scale — unrestricted search needed 14,450 ms against 6,550 ms for a size-7 plan and escalates exponentially (Fink and Yang), exact temporal search solved 164 problems against 189 for lazy satisficing search (Coles et al.), and exact SMT solving reached 262 problems against 315 for heuristic portfolios on numeric tasks (Chen and Thiébaux) — so exactness is folded into the always-on ablation and the shortlist competitor rather than dismissed.
Proposed Experiments
E1 validates the transition surrogate on held-out LINCS L1000 perturbations: how well the amortized network-rewiring model predicts the next state after a compound. The reference point is CellVS-Net's reported 12–47% MSE reduction across perturbation types (Ellington et al.); the planner needs comparable fidelity, since every search step depends on it.
E2 measures the efficiency-quality frontier on tasks built from DepMap and the O'Neil data: the number of surrogate evaluations needed to reach a target disease-reversal threshold, the in-vivo violation rate of the resulting regimens, and the suboptimality factor, which should land in a band comparable to the 5–12% inflation of abstraction-based pathfinding (Sturtevant). The prediction is that pure primary-effect search and always-on search define the two ends of the frontier and the lazy planner dominates the middle.
E3 tests translational fidelity retrospectively on TCGA: among regimens the planner proposes per cancer type, what fraction of component drugs are clinically validated in OpenTargets, and what is the Spearman correlation between predicted regimen efficacy and patient response, with sRGES's rho of 0.61 in liver cancer as the single-intervention reference (Chen et al.) and DeepSADR's Sorafenib AUC of 0.957 as the single-drug ceiling (Zhang et al.).
E4 is the decisive check: the budget-matched head-to-head against the decoupled shortlist-plus-exhaustive pipeline and both ablations, swept over regimen lengths from one to five steps and over shortlist sizes, plotting fidelity against evaluation count. The proposal stands or falls on whether lazy re-expansion holds near-exact fidelity at satisficing cost as length grows. As a final grounding step, the top predicted sequential regimen for liver hepatocellular carcinoma — the setting where sRGES validated pyrvinium pamoate in mouse xenografts — would be carried into the same subcutaneous-xenograft protocol, comparing staggered versus simultaneous administration of the same compounds, the cleanest possible test of whether sequencing changes what survives in vivo.
References
- Planning with Primary Effects: Experiments and Analysis
- Aldehyde dehydrogenase activity is a cancer stem cell marker of tongue squamous cell carcinoma
- Mast cells exert pro-inflammatory effects of relevance to the pathophyisology of tendinopathy
- FOXC1 Contributes To Microvascular Invasion In Primary Hepatocellular Carcinoma Via Regulating Epithelial-Mesenchymal Transition
- Efficient Temporal Planning Using Metastates
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Cell-Level Virtual Screening
- Reversal of cancer gene expression correlates with drug efficacy and reveals therapeutic targets
- In silico drug combination discovery for personalized cancer therapy
- Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies
- Quick Multi-Robot Motion Planning by Combining Sampling and Search
- Memory-Efficient Abstractions for Pathfinding
- Novelty Heuristics, Multi-Queue Search, and Portfolios for Numeric Planning