Causal Safety as a Search Objective in Multi-Step Synthesis Planning

Multi-step synthesis planning asks a computer to work a complex target molecule backwards, step by step, into cheap building blocks that can actually be bought — a search over a tree of chemical reactions. Mendelian randomization (MR) asks a different question entirely: because we each inherit genetic variants essentially at random at conception, variants that shift the level of a specific protein act like a lifelong natural experiment, letting epidemiologists estimate whether perturbing that protein truly changes disease risk, free of the confounding and reverse causation that plague ordinary observational data. Today these two tools meet only in the loosest way: MR picks a causal protein target, and then — at best — someone docks a known compound against it. The compound's synthetic accessibility is simply assumed.

The proposal here is to close that loop in the opposite direction from the obvious one. Rather than using MR to pick a target and handing a static hit structure to a planner afterwards, genetic causal evidence should enter the retrosynthetic search itself as an objective and a pruning signal. Population-scale MR and phenome-wide association studies (PheWAS) tell us, for thousands of druggable genes, what perturbing each gene does to hundreds of human phenotypes — including adverse ones. That is exactly the kind of global, non-additive constraint that current planners lack, and exactly the kind of human-relevant evidence that no docking score can substitute for.

Core Claim

A multi-step retrosynthesis planner can be made target-aware and safety-aware by adding one objective axis: a genetically grounded causal score for the molecule being planned, estimated by mapping the molecule onto a predicted protein-binding profile and aggregating Mendelian-randomization and PheWAS evidence for those proteins. Searching the Pareto front over (synthetic cost, chemical feasibility, causal risk) will yield routes and scaffold choices that survive downstream genetic safety screening at a much higher rate than the prevailing pipeline of MR-then-dock-then-plan, without sacrificing solvability.

The closest existing work stops short of this in three distinct ways. The galangin study for primary biliary cholangitis runs a full druggable-genome MR across 6,888 genes, confirms signals by colocalization, screens on-target side effects by PheWAS in FinnGen, and then queries the static DSigDB database for existing molecules, docking galangin to ADORA2A at −8.8 kcal/mol — but its "lead" is a purchasable natural product, and the pipeline has no mechanism whatsoever to plan the synthesis of new chemical matter if no adequate ligand exists in stock Identification of galangin as a therapeutic candidate for primary biliary cholangitis.... The traditional-medicine study repeats this pattern for NSCLC, prioritizing TNF and PIK3CA and docking resveratrol at −8.2 and −8.5 kcal/mol, while explicitly admitting it cannot model pharmacokinetics or metabolism and never touches synthesis Integrating Network Pharmacology And Mendelian Randomization.... And Synthon Contrastive Learning co-designs ligand conformers with synthetic pathways for a fixed pocket, but the pocket is a static crystal structure with no genetic validation of whether perturbing that protein is causally beneficial or harmful in humans Synthon Contrastive Learning for Synthesizable 3D Molecule Generation. Nobody, in this corpus at least, has put genetic causal evidence inside the search.

Research Program

Three components, each reusing established machinery wherever possible.

The first component builds a gene-level causal liability map. For each gene $g$ in the druggable genome, drug-target MR with cis-eQTL and cis-pQTL instruments (eQTLGen blood data, GTEx v10 tissue data) plus Bayesian colocalization yields an effect estimate $\hat\beta_{gp}$ on each phenotype $p$ in a PheWAS resource such as FinnGen, exactly the inputs demonstrated by the PBC study. From these we define a liability magnitude $L_g$, which is large and positive when genetically increased activity of $g$ causally raises risk of adverse phenotypes, and negative when the evidence marks $g$ as protective for the disease of interest. Standard pleiotropy-robust MR sensitivity estimators and ancestry-aware checks guard the map, though the well-documented European-ancestry bias of GWAS resources remains an honest limitation Identification of galangin as a therapeutic candidate....

The second component bridges genes to molecules, which is the representational gap a skeptical reviewer will attack first, and we address it explicitly rather than by hand-waving. A molecule $m$ does not have genes; it has a predicted protein-binding profile. We define

$$S_{\text{causal}}(m) \;=\; \sum_{g \in \mathcal{G}} \pi_g(m)\, L_g,$$

where $\pi_g(m) \in [0,1]$ is a binding prior for molecule $m$ against the protein product of gene $g$, estimated by docking $m$ across a panel of structures for the MR-prioritized targets and their known off-target families. The intended efficacy target contributes a negative (rewarding) term; off-targets with adverse PheWAS signatures contribute penalties. This score is admittedly a modeled surrogate — it is only as good as the binding priors — so the design treats it as uncertain: we propagate docking-score uncertainty into $S_{\text{causal}}$ and report results under optimistic and pessimistic binding assumptions. The corpus itself validates that this structural bridge is how MR findings get operationalized: both boundary papers land on docking scores as their quantitative endpoint, and both treat scores below −5.0 kcal/mol as strong binding Identification of galangin...; Integrating Network Pharmacology....

The third component puts the score into the search. We extend multi-objective retrosynthetic planning, following MORetro's formulation of planning as Pareto optimization over competing objectives with Bayesian-optimization-informed weight sampling and vector-valued lower-bound pruning From Feasible to Practical: Pareto-Optimal Synthesis Planning, by appending $S_{\text{causal}}$ as a third axis and — crucially — by enforcing it as a global, non-additive constraint rather than a sum of per-step penalties. That matters because MORetro itself documents that current planners assume step-wise additivity and cannot enforce route-wide criteria; a causal liability attached to a scaffold family is precisely such a criterion. At the root of the search, the score also conditions scaffold choice: among analogue series that dock comparably to the causal target, the planner preferentially expands those whose off-target profiles are genetically clean, in the spirit of DESP's goal-constrained search but with the goal being a causally validated pharmacophore rather than a fixed starting material Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search. For the single-step model underneath, we follow the retrosynthetic-crosstalk finding that template-based models are more reliable inside tree search and hallucinate less chemistry than template-free generators Retrosynthetic crosstalk between single-step reaction and multi-step planning, evaluated under the Syntheseus standardized benchmarking library to avoid the deduplication and metric-inflation artifacts that benchmark re-evaluation exposed Re-evaluating Retrosynthesis Algorithms with Syntheseus.

Baselines

The do-nothing reference point is the prevailing static pipeline: drug-target MR with colocalization and PheWAS, a database lookup (DSigDB), and docking — exactly the galangin study's method, whose only quantitative computational output is the −8.8 kcal/mol galangin–ADORA2A docking energy, and for which no reported result exists on Lacuna for the route solvability or synthetic accessibility of its prioritized hits Identification of galangin as a therapeutic candidate.... Beating it means, concretely, producing executable routes for molecules it can only assume are purchasable, and measuring how often its database hits actually fall into synthesis dead ends.

The strongest existing competitor is MORetro, the multi-objective planner whose objectives already include material cost, energy, and solvent toxicity, and which reports a 5–8% hypervolume improvement, vector-valued pruning that reduced the search space by an average of 54%, success rates within roughly 2% of single-objective A, and route-finding in 6–12 minutes on a standard workstation; it also showed that 18% to 36% of routes from single-objective searches are strictly dominated From Feasible to Practical: Pareto-Optimal Synthesis Planning. We will run MORetro with its native objectives on the identical target sets, identical single-step model, identical eMolecules building-block library, identical expansion budget, and the same tuning budget for its weight sampling, so the only difference is the causal axis. Beating it means holding solve rate within about 2% while raising the fraction of Pareto-optimal routes that pass a held-out causal-safety filter, and improving hypervolume on the extended objective vector. As secondary search controls we run Retro and DESP under the same budgets, referencing the reported USPTO-190 solve rates of roughly 36% for Retro* and approximately 40% for DESP-F2E within 500 expansions, and the 38 versus 73 average expansions to solution on Pistachio Reachable Double-Ended Synthesis Planning....

The decisive ablation replaces $S_{\text{causal}}$ with MORetro*-style additive per-reagent local toxicity, keeping everything else fixed. This isolates whether the gain comes from genuinely genetic, global, human-population evidence or merely from having any additional safety-flavored term. A second ablation drops the docking uncertainty propagation and treats $\pi_g(m)$ as a point estimate, testing whether the causal score's noise dominates its signal.

Why not just…?

Why not just run MR to pick a target, dock a hit, and hand it to an off-the-shelf planner? This is the most serious alternative and we treat it as a genuine threat rather than a strawman. The evidence against pure sequential pipelines is that error compounds when independently trained components are chained — the STEP benchmark shows a chemical model at 45.03% single-step accuracy collapsing to 22.07% on a two-step composite task Two Steps from Hell: Compositionality on Chemical LMs — and that molecules designed only for binding, without synthesis awareness, fall into planner dead ends with feasibility decaying as route depth grows, with all evaluated planner combinations struggling on the complex medicinal-chemistry targets in Literature30 and Patent200 Retrosynthetic crosstalk.... Still, docking-then-planning might work often enough in practice. Our third experiment is designed to settle this directly: it measures the hit-to-route survival rate of the sequential pipeline against the joint planner on the same candidate pools, and we regard a comparable survival rate for the sequential pipeline as a negative result for the fusion.

Why not just look up known compounds in DSigDB or buy close analogues from eMolecules? For some targets this plainly suffices, and if the causal target has an adequate purchasable ligand, lookup is cheaper and should win. The corpus shows where it fails: the galangin pipeline's discovery space is capped at pre-existing natural products with whatever potency and intellectual-property position they happen to carry Identification of galangin..., and the Synthon Contrastive work makes the complementary point that without synthesis awareness built into design, newly designed high-affinity molecules simply stay in silico Synthon Contrastive Learning.... Our experiments include a measured breakdown of how often database hits are synthesis-solvable and causally clean versus how often the joint planner's de novo routes are needed.

Why not just use docking scores as the safety proxy and skip MR entirely? Docking scores compress into a narrow band of plausibly strong binding — −8.8, −8.5, and −8.2 kcal/mol across two papers and four entirely different protein–ligand pairs — and say nothing about the sign or human consequence of perturbing the target Identification of galangin...; Integrating Network Pharmacology.... A molecule can dock tightly to a protein whose genetic perturbation raises disease risk; only MR and PheWAS carry that population-level causal sign. We are equally honest about the reverse limitation: MR gives gene-level effects, not atomic detail, which is why the molecule-level score is a modeled surrogate with explicit uncertainty and sensitivity analysis rather than a claimed identity.

Proposed Experiments and Decisive Checks

The first experiment reconstructs the PBC case study on its documented inputs — 6,888 druggable genes, eQTLGen blood data from 31,684 samples, GTEx v10 liver tissue, the 8,021-patient PBC GWAS, and FinnGen PheWAS — and verifies that the causal liability map recovers the known protective ADORA2A signal before any chemistry is run Identification of galangin...; a parallel map is built for the NSCLC TNF/PIK3CA setting from the traditional-medicine study's MR results Integrating Network Pharmacology....

The second experiment measures planning quality on the standardized multi-step benchmarks: USPTO-190, ChEMBL, and Pistachio Reachable, with eMolecules as the termination library, all evaluated through Syntheseus to avoid the deduplication artifacts that inflated single-step claims by as much as 5.8% top-5 accuracy for one model Re-evaluating Retrosynthesis Algorithms with Syntheseus. Metrics are solve rate under a fixed expansion budget, expansions to first solution, hypervolume on the extended objective vector, and wall-clock time, against MORetro, Retro, and DESP under matched budgets. Self-improved-style search-feedback training Self-Improved Retrosynthetic Planning, which raised success from 86.84% to 96.32% while cutting model calls from 208.09 to 96.22, is applied identically to all planners so no method gains an unfair search-efficiency advantage.

The third and decisive experiment measures hit-to-route survival. Around the flavonoid and stilbene cores of the two boundary papers, we generate analogue series, pool the top-docked candidates, and compare two conditions: the sequential pipeline (dock first, then plan the winners) versus the joint planner conditioning scaffold expansion on the causal score. The headline number is the fraction of top candidates that are simultaneously solvable within the building-block library and pass the held-out causal off-target filter. A fourth experiment stress-tests on Literature30 and Patent200, where the corpus reports all evaluated methods struggle. A final check quantifies the price of the causal axis: if enforcing it costs solve rate the way fixed-weight baselines in MORetro* failed on up to 25% more molecules, we report the full trade-off curve rather than a single operating point From Feasible to Practical....

The decisive read on the whole proposal is simple. If, at matched solve rate and matched compute, routes chosen under the causal objective pass a held-out genetic safety filter at a substantially higher rate than MORetro* routes and sequential-pipeline routes — and if the additive-toxicity ablation does not reproduce that gain — then genetic causal evidence belongs inside retrosynthetic search, not just upstream of it.