Generative models now design molecules by learning a continuous "latent space" — a smooth map in which nearby points correspond to similar molecules — and then searching that map for structures with desired properties, instead of screening the roughly $10^{60}$ possible drug-like compounds one at a time. The most advanced of these models design molecules to fit a specific binding pocket on a disease protein: TransDiffSBDD generates ligands conditioned on 3D pocket coordinates, and GraphVF conditions a flow-based generator on both pocket geometry and 2D substructure motifs. But every one of these models conditions on a single, frozen, crystallographic snapshot of the target, and that snapshot comes from databases whose structures do not represent human genetic diversity: the same protein can carry different amino acids in different ancestral populations, reshaping the very pocket the molecule must bind. The clinical analogue of this failure is already documented — diagnostic criteria for monogenic diabetes developed on White Europeans misclassify non-White patients Precision diagnosis for monogenic diabetes — yet no generative model in the literature accounts for it.
We propose to build molecule generators whose latent space is explicitly split into two parts: a sub-vector that must behave identically across population-specific variants of the target pocket, and a sub-vector free to specialize to particular variants. This decomposition is a direct transplant of a technique proven in genetics: an autoencoder that separates an ancestry-invariant genetic instrument from ancestry-specific confounding reduced Mendelian randomization bias from 0.3 to near zero on All of Us cohorts Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning. We pair it with target selection by cross-ancestry genetic causal analysis rather than observational exposure data, because the exposures feeding observational studies are retrospective self-reports that human memory systematically rewrites — a problem documented from depression-distorted childhood recall Wu et al. to post-injury baseline inflation Lyons and decaying recall of infant feeding and past injections Liu et al., Zhang et al.. The reader's two-sentence version: today's designers fit one frozen version of a target protein; we make them fit all population-specific versions at once by splitting the molecule's latent code into an ancestry-invariant part and a variant-tuned part, and we choose which protein to target using genetics corrected for ancestry bias.
Core Claim
A pocket-conditioned generator whose latent code $z$ is partitioned into an invariant component $w$ and a variant component $v$, trained to keep $w$ distributionally identical across an ensemble of ancestry-specific pocket variants, will produce molecules that retain binding affinity across those variants — including variants held out from training — at rates that single-pocket or per-variant-conditioned generators cannot match, because the invariant code captures the binding pharmacophore shared by all versions of the protein rather than memorizing individual pockets. The same invariance machinery also yields something no per-variant pipeline gives: a readable list of which substructural features are ancestry-robust, which is exactly the information a medicinal chemist needs to avoid building a drug that works well in the population whose crystal structure happened to be deposited and poorly elsewhere.
Research Program
The program has three linked parts, each grounded in an existing method that we extend rather than reinvent.
Part 1: cross-ancestry target validation. We select targets by drug-target Mendelian randomization — using genetic variants as naturally randomized proxies for lifelong perturbation of a protein, which sidesteps both recall distortion and reverse causation. The starting candidates come from the literature: PRKAG1, whose genetically proxied down-regulation reduces colorectal cancer risk (OR 0.74, 95% CI 0.63–0.87, adjusted $p = 0.001$) with valid instruments found for only PRKAG1 and GPD1 among eleven metformin targets Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes, and GLP1R, the diabetes target used in the All of Us representation-learning study Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning. Crucially, we re-run the instrument construction per ancestry in the All of Us Biobank using the invariant/variant autoencoder, keeping only targets whose causal effect is sign- and magnitude-stable across East Asian, African, and European cohorts. This matters because standard estimators (2SLS, MR-Egger) showed confounding bias up to 0.3 in exactly this setting while the invariant representation reduced it to near zero Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning, and because summary-data frameworks like GSMR are bound to European reference panels — Zhu et al. mapped 45 causal associations, including a 3.29-fold type-2 diabetes risk increase per standard deviation of BMI, but explicitly flagged their European-only constraint Causal associations between risk factors and common diseases inferred from GWAS summary data.
Part 2: the variant pocket panel. From All of Us genotype data we catalog non-synonymous variants in the selected target genes, note their ancestry-specific frequencies, and computationally introduce each amino-acid substitution into the corresponding crystallographic pocket from CrossDocked2020 (the ~22.5 million protein–ligand pair benchmark used by TransDiffSBDD, GraphVF, and FuncBind), relaxing each mutated pocket to a plausible conformation. This produces a panel in which the "same" target exists in many population-specific versions — the thing current structure-based generators, which treat receptors as static and uniform, structurally cannot represent.
Part 3: the invariant chemical latent. We take a pocket-conditioned generator — GraphVF's architecture is the natural starting point because its 2D substructure latent already acts as an uncoupled control knob for property optimization without retraining GraphVF — and split its molecule latent into $z = (w, v)$, trained with
$$\mathcal{L} = \mathcal{L}_{\text{rec}} + \lambda \, \text{MMD}\big(\hat{w}_{\text{variant } a},\, \hat{w}_{\text{variant } b}\big) + \mu \, \text{HSIC}(\hat{w}, \hat{v}),$$
where $\mathcal{L}_{\text{rec}}$ reconstructs the molecule, MMD (maximum mean discrepancy) forces the distribution of the invariant code $\hat{w}$ to match across pocket-variant groups, and HSIC (an independence criterion) keeps $\hat{w}$ and $\hat{v}$ statistically separate. Generating from $\hat{w}$ alone, with $\hat{v}$ marginalized, should yield molecules that bind robustly across variants; pinning $\hat{v}$ gives deliberate allele-specific design. We deliberately keep the dimension of $\hat{w}$ small: the genetics work showed that over-specifying the invariant latent lets ancestry noise leak back in Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning, and we will audit for the same leakage (see experiments). This is a conscious inversion of the field's scaling instinct — FuncBind's gains came from a 5-billion-parameter latent Unified all-atom molecule generation with neural fields — and the tension between capacity and leakage is precisely what the ablations will measure.
As a secondary analysis, we apply the recall-bias lens to the training data itself. CrossDocked2020 is a retrospective record of complexes that happened to bind and crystallize — a survivorship-filtered "memory" of chemical interaction, structurally analogous to the reconstructive memory bias that distorts human exposure reports. Just as Guo et al. audited self-reported acetaminophen use against liquid-chromatography mass spectrometry and found detectable exposure in 100% of samples from self-reported non-users Acetaminophen Use in Pregnancy, we audit the database's "recall" against objective oracles (docking and, for a subset, quantum-chemical refinement in the spirit of cG-SchNet, whose generated structures required DFT refinement Inverse design of 3d molecular structures with conditional generative neural networks) and measure how the retrospective composition of the pocket ensemble skews cross-variant generalization.
Baselines
The "do nothing" reference point is current practice: run an unmodified pocket-conditioned generator on the single reference crystal pocket and then test its molecules against the variant pockets. Concretely, TransDiffSBDD TransDiffSBDD is the strongest published system on this benchmark and reports an 83.9% success rate (a combined affinity, drug-likeness, and synthesizability threshold) versus 52.5% for DecompOpt, with an average AutoDock Vina binding score of −9.37 on CrossDocked2020; GraphVF outperformed GraphBP and Pocket2Mol on the same benchmark, but no reported result found on Lacuna for GraphVF's raw metric values. The interesting quantity is not its reference-pocket number, which we expect to roughly reproduce, but the size of the drop when its molecules face variant pockets — no reported result found on Lacuna for any generator's performance on ancestry-variant pockets, which is the gap itself.
The strongest competitor is per-variant conditioning: the same TransDiffSBDD model, retrained or re-prompted separately on each variant pocket, with molecules then ranked or intersected by worst-case affinity across variants. This is the honest "you don't need new machinery" baseline and it shares our data, our oracle budget, and our tuning budget — we give every method the same CrossDocked2020-derived training corpus, the same number of pocket-conditioning instances, and an equal search budget over hyperparameters, so the comparison isolates the invariance ingredient rather than compute. A docking-based virtual screening baseline over existing chemical libraries against each variant pocket completes the non-generative comparison. The key ablation removes exactly our ingredient: the identical architecture conditioned on variant pockets but with a single undivided latent — no MMD invariance, no HSIC split — which distinguishes "conditioning on variants helps" from "enforcing invariance in the latent helps." On the target-selection side, the baseline is standard 2SLS and MR-Egger per ancestry, which exhibited bias up to 0.3 on All of Us data before the representation-learning correction Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning. Beating the per-variant baseline means a higher worst-case-per-variant success rate and better Vina scores on held-out variants at matched validity, novelty, and synthesizability; beating the ablation means the invariant latent specifically, not variant conditioning generally, drives the gain.
Why not just…?
Why not just run the existing generator once per variant pocket? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. It will likely work well for common variants with many carriers and representative structures. Its weakness is extrapolation: the variants that most differentiate ancestries are often rare, with no crystal structure and too few carriers to train on, and a per-variant model has no mechanism for generalizing to them — whereas a latent forced to be invariant across observed variants has a defined inductive bias toward the shared pharmacophore. The decisive experiment is the held-out-variant test in E3 below; if per-variant conditioning extrapolates as well as the invariant latent, the machinery is unnecessary and we will report that.
Why not just screen existing chemical libraries against the variant pockets? Screening is cheaper and needs no training, but it is confined to known chemotypes and inherits human heuristic bounds on chemical space Machine-Learning-Assisted De Novo Design of Organic Molecules and Polymers; the case for generation is that conditional latent models find structure beyond the catalog — cG-SchNet discovered 32% more unique low-energy isomers of $\text{C}_7\text{O}_2\text{H}_{10}$ than the QM9 dataset itself contained Inverse design of 3d molecular structures with conditional generative neural networks. We still run screening as a baseline and concede that for a target with a deep ligand history it may be competitive on per-variant affinity; our claim is about worst-case and held-out-variant behavior, which is where the comparison will be settled.
Why not just use standard Mendelian randomization for target selection? Because the instruments themselves break across ancestries: linkage-disequilibrium patterns shift, tagging SNPs stop tracking the causal allele, and standard estimators carried bias up to 0.3 in the All of Us setting where the invariant representation reduced it to near zero Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning. Summary-data methods like GSMR with HEIDI-outlier pleiotropy filtering are powerful but anchored to European reference panels Causal associations between risk factors and common diseases inferred from GWAS summary data, and the diagnostic precedent shows what happens when European-derived criteria transfer unmodified Precision diagnosis for monogenic diabetes.
Why not just fix the exposure data instead of routing through genetics? Neither available fix is adequate. Better questionnaires cannot undo reconstructive memory: current depressive states amplify recalled negative childhood events and suppress positive ones Wu et al., and recall of even concrete events like breastfeeding duration or injection frequency decays badly over years Liu et al., Zhang et al.. Objective spot biomarkers are temporally blind: acetaminophen metabolites clear in under 4 hours, so mid-pregnancy urine could not recover first-trimester use, even though the same study showed self-report does correlate with biomarkers for proximal windows ($r > 0.70$ for third-trimester use versus cord serum) Acetaminophen Use in Pregnancy. Genetic instruments proxy lifelong exposure and are immune to both failure modes — though we concede they capture average lifelong exposure, not timing, which no method in this literature resolves.
Proposed Experiments and Decisive Checks
E1 (target validation): replicate the drug-target analysis for PRKAG1 and GLP1R per ancestry in All of Us, with standard 2SLS/MR-Egger versus the invariant-instrument autoencoder as in Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning. Decisive outcome: at least one target whose causal estimate is stable across ancestries under the invariant method but unstable under standard estimators, establishing a population-robust target list.
E2 (cross-variant generation): train the invariant-latent generator on the variant pocket panel; evaluate worst-case and mean Vina score and TransDiffSBDD-style success rate across all variants, against the single-pocket baseline, the per-variant conditioning baseline, the screening baseline, and the no-split ablation, all under matched budgets. Success must be reported alongside synthesizability and novelty, since unconstrained latent exploration is known to produce plausible-but-unmakeable structures Generative Models as an Emerging Paradigm in the Chemical Sciences. No reported result found on Lacuna for any variant-pocket baseline, so this comparison is new data, not re-scoring.
E3 (held-out extrapolation): exclude variants above a carrier-frequency threshold (and entire ancestry-stratified variant groups) from training and test binding on them. This is the check that adjudicates the per-variant-conditioning alternative; cG-SchNet's demonstrated ability to generate into "data deserts" — property regimes sparser than its 55,000-molecule training set — is the precedent that conditional latent generation can extrapolate Inverse design of 3d molecular structures with conditional generative neural networks.
E4 (leakage audit): test whether $\hat{w}$ predicts pocket variant or ancestry label above chance; it should not, and if it does, the latent dimension of $\hat{w}$ is over-specified — the failure mode documented for the genomic version of this decomposition Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning.
E5 (oracle audit and retrospective-composition analysis): refine a subset of generated structures with higher-fidelity quantum calculations, following the DFT-refinement precedent for cG-SchNet outputs Inverse design of 3d molecular structures with conditional generative neural networks, and retrain with the pocket ensemble reweighted away from its retrospective (crystallized-and-bound) composition to quantify how much of the cross-variant gap is database survivorship rather than model limitation. We will use the existing CrossDocked preprocessing pipeline from prescient-design/funcbind rather than rebuilding it.
The program is falsifiable at a specific point: if per-variant conditioning matches the invariant latent on held-out variants in E3, then invariance in the chemical latent is unnecessary and the contribution collapses to a benchmark and a negative result. We consider that a publishable outcome either way, but the mechanism, the genetics-proven loss functions, and the ancestry-diverse pocket panel give the invariant route a principled reason to win where memorization cannot.
References
- TransDiffSBDD: Causality-Aware Multi-Modal Structure-Based Drug Design
- GraphVF: Controllable Protein-Specific 3D Molecule Generation with Variational Flow
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
- Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning
- Clinical Features of Patients with Dysthymia in a Large Cohort of Han Chinese Women with Recurrent Major Depression
- Measuring the Population Burden of Injuries—Implications for Global and National Estimates: A Multi-centre Prospective UK Longitudinal Study
- Breastfeeding and myopia: A cross-sectional study of children aged 6–12 years in Tianjin, China
- Alternative risk factors of HCV infection in a rural community in China
- Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug‐Target Mendelian Randomization Study
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- Unified all-atom molecule generation with neural fields
- Acetaminophen Use in Pregnancy: A Comparison of Self-Reported Intake with Maternal and Newborn Biomarker Measures
- Inverse design of 3d molecular structures with conditional generative neural networks
- Machine-Learning-Assisted De Novo Design of Organic Molecules and Polymers: Opportunities and Challenges
- Generative Models as an Emerging Paradigm in the Chemical Sciences