A Standard Ruler for Activity Cliffs

Cosmologists once faced a measurement problem that, stripped of its physics, sounds like a machine-learning problem. They had a signal of known size — a preferred separation between galaxies, about 150 megaparsecs, left over from sound waves in the early universe — but gravitational motion had smeared it: galaxies drifted by roughly 10 megaparsecs each, blurring the sharp feature. Their fix was not a bigger model but a correction with the right shape: estimate the smooth displacement field that moved the galaxies, shift everything back along it, and the sharp scale reappears. That single step, density-field reconstruction, cut the error on the distance measurement and effectively gave the BOSS survey nearly 2.3 times its real volume The clustering of galaxies in the SDSS-III Baryon Oscillation Spectroscopic Survey: baryon acoustic oscillations in the Data Release 9 spectroscopic galaxy sample, Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys.

Chemistry has the mirror-image problem, and it has not been solved. Molecular representations place structurally near-identical molecules next to each other, yet a one-atom edit can change biological potency a hundredfold — an "activity cliff" — while completely different scaffolds can converge on the same function. Benchmarks show that every current model, from fingerprints to graph neural networks, collapses when moved to a protein target it has not seen: the best performer on unseen targets in the ACNet benchmark, a graph convolutional network, reaches only AUC 0.579, barely above chance Activity Cliff Prediction: Dataset and Benchmark. This proposal borrows the cosmologist's fix: treat measured pairs of analogues with known potency differences as standard rulers — fixed distances in functional space — and estimate a smooth, low-order displacement field that warps a structural embedding space until those rulers hold. The calibrated geometry then answers a question drug delivery currently cannot: how fast must a pH-triggered carrier release its cargo so that the concentration inside a cell stays on the therapeutic side of a cliff?

The third direction enters as the application regime and the honesty test. Nanocarriers that are stable at blood pH (7.4) and fall apart in acidic tissue (pH 6.5–6.8) or endosomes (pH 4.5–5.5) are designed under an implicit assumption of monotonicity: more released drug means more efficacy. But large-scale phenotypic profiling shows this assumption is sometimes false — compounds hitting the same pathway can produce diametrically opposite cellular phenotypes, a validated "bimodal effect" across 30 pathways Large scale compound selection guided by cell painting reveals activity cliffs and functional relationships. If release rate and cliff thresholds interact, the carrier itself becomes a phenotype selector, and nobody currently has the calibrated geometry to predict which way it selects.

Core Claim

The claim is twofold. First, a first-order, ruler-anchored reconstruction operator — a smooth warp estimated from a few dozen measured analogue pairs per target — will substantially outperform all existing molecular representations in the regime where they fail: transfer to unseen protein targets and prediction across cliff boundaries, because the fixed functional distances constrain the warp in a way that unconstrained embeddings and fine-tuning are not constrained. Second, the cliff thresholds this geometry produces are actionable: they determine the release-rate windows within which a pH-responsive carrier produces the intended phenotype rather than its opposite, and this can be tested in vitro before any in vivo work.

One honest caveat up front. In cosmology the displacement field obeys known physics (gravity, to first order the Zel'dovich approximation). Molecular space has no such governing equations. What transfers is not the physics but the design pattern: anchor to an invariant scale, keep the correction smooth and first-order, and characterize the data-density thresholds below which the correction does more harm than good. The last point matters because cosmology already quantified its own limits — reconstruction removes scale bias only above a tracer density of roughly $4 \times 10^{-4}\,h^3\,\mathrm{Mpc}^{-3}$ and an effective volume of $0.5\,h^3\,\mathrm{Gpc}^3$ Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys — and our proposal predicts, and plans to measure, an analogous anchor-density threshold in molecular space.

Research Program

Stage one defines the molecular standard ruler and the reconstruction operator. Let $g_i$ be a structural embedding of molecule $i$ (we will use both ECFP-derived and graph-network embeddings, since the benchmark evidence says simple fingerprints are the stronger substrate Activity Cliff Prediction: Dataset and Benchmark). For a measured anchor pair $(i, j)$ — two analogues with assayed potencies on the same target — the ruler is the known functional distance $s_{ij}$, taken as the absolute log-potency ratio (the $\ge 10$-fold criterion that defines a cliff in Exposing the Limitations of Molecular Machine Learning with Activity Cliffs gives a natural unit). The warp is $W(x) = x + u(x)$ with $u$ a smooth displacement field, fitted per target:

$$\min_{u} \; \sum_{(i,j)\in A}\left(\|W(g_i) - W(g_j)\| - s_{ij}\right)^2 \; + \; \lambda \int \|\nabla u(x)\|^2\, dx$$

where $A$ is the anchor set and $\lambda$ plays the role of the smoothing length $R$ that cosmology tuned to 15 and 10 $h^{-1}$ Mpc for its two survey samples Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys. Cliff prediction for an untested pair is then a statement about their warped separation: if $W$ places them across a cliff-distance contour, the warp predicts the sign and size of the potency change.

Stage two evaluates transfer. The bet is that a warp anchored to fixed distances estimated from a small anchor set on a new target generalizes better than any embedding trained to memorize target-specific patterns, because the ruler constraint leaves far fewer degrees of freedom to overfit. The key comparison is against the documented collapse at AUC 0.579 on unseen targets Activity Cliff Prediction: Dataset and Benchmark, and against the cliff-specific error metric that Exposing the Limitations of Molecular Machine Learning with Activity Cliffs showed is only loosely tied to overall error (Pearson $r = 0.81$ between global RMSE and cliff RMSE, with large cliff-error variation at equal global error).

Stage three couples the calibrated cliffs to release kinetics. For a payload sitting near a cliff or in a bimodal pathway, the carrier's pH-dependent release rate $r(\mathrm{pH})$ sets an intracellular concentration trajectory $C(t)$, and the therapeutic condition is simply that it clear the cliff threshold $C^{*}$:

$$\max_{t \in [0,T]} \; C(t;\, r(\mathrm{pH})) \;\ge\; C^{*}_{\mathrm{cliff}}$$

We will build pH-degradable carriers of at least two rate classes — a fast acid-cleaved metal-organic-framework type, following the ZIF-8 chemistry that releases its cargo when imidazole protonation breaks the zinc-nitrogen bonds below pH 6.5 Rapid synthesis of a Bi@ZIF-8 composite nanomaterial as a near-infrared-II (NIR-II) photothermal agent for the low-temperature photothermal therapy of hepatocellular carcinoma, and a slower mixed-micelle type modeled on the 221.3 ± 5.3 nm platform with roughly 80.6% doxorubicin encapsulation Author Correction: pH-sensitive micelles for the intracellular co-delivery of curcumin and Pluronic L61 unimers for synergistic reversal effect of multidrug resistance, with peptide-nanoporation carriers as a third rate knob where acid-triggered structural transitions govern membrane escape pH-Responsive Peptide Nanoparticles Deliver Macromolecules to Cells via Endosomal Membrane Nanoporation — and test whether the predicted phenotype flips match observed ones.

Proposed Experiments

The computational experiments use two public benchmarks. On the ACNet matched-molecular-pair benchmark (over 400,000 pairs across 190 human targets), we reproduce the target-split out-of-distribution protocol on the "Mix" subset and measure mean AUC with anchor sets capped at 50 pairs per held-out target. On MoleculeACE (30 targets from ChEMBL v29, between 615 molecules for JAK1 and 3,657 for DRD3), we hold overall RMSE fixed across methods and compare cliff-specific RMSE, following the protocol of Exposing the Limitations of Molecular Machine Learning with Activity Cliffs. Alongside these we run the anchor-density scaling study: error as a function of anchor-pair count per target, from 5 to 500, expecting a cosmology-like operating curve where the warp helps above some density and actively hurts below it, mirroring the tracer-density and effective-volume thresholds documented in Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys.

The wet-lab experiment is designed to be decisive and small. From the 81 chemical series (2,277 compounds) that the JUMP-CP analysis flagged as phenotypic activity cliffs Large scale compound selection guided by cell painting reveals activity cliffs and functional relationships, we select cliff pairs in pathways with a validated bimodal phenotypic split. The warp, built only from anchor measurements, pre-registers a prediction for each (payload, carrier, release-rate) combination: which side of the bimodal split the Cell Painting profile lands on, using the same DINOv2-feature, sphered, cosine-similarity profiling pipeline. We then run the release experiment across a pH gradient stepping from 7.4 to 6.8 to 5.0 and score prediction accuracy against chance. The micelle platform's own result — 87.89% late apoptosis in resistant MCF-7/ADR cells only when curcumin and Pluronic L61 were co-encapsulated at working ratios Author Correction: pH-sensitive micelles for the intracellular co-delivery of curcumin and Pluronic L61 unimers for synergistic reversal effect of multidrug resistance — is the standing evidence that formulation details, not just chemistry, decide the biological endpoint; our experiment tests whether release rate is one of those details.

If the in vitro predictions hold, a small terminal in vivo arm maps biodistribution and tissue-level phenotypes with the SCP-Nano pipeline, whose 3D U-Net segmentation reaches an average F1 of 0.73 for lipid nanoparticle detection and detects doses as low as 0.0005 mg kg$^{-1}$, 100–1,000 times below conventional PET, CT, or MRI Nanocarrier imaging at single-cell resolution across entire mouse bodies with deep learning. This is also where the standard-ruler idea has a second, smaller life: the same anchor-based calibration logic could normalize single-cell distribution maps across organ beds with different permeability and pH, though we treat that as an extension, not a claim.

Baselines

The do-nothing reference is the raw structural-similarity heuristic that defines a cliff nominally: flag any pair with Tanimoto similarity at or above 0.9, the criterion used in Exposing the Limitations of Molecular Machine Learning with Activity Cliffs. This baseline detects that a cliff might exist but predicts neither its direction nor its size; any calibrated method must beat it on signed potency-change prediction, and it calibrates what chance looks like on the phenotype-flip task.

The strongest existing competitors are the ones the benchmarks already crowned. On ACNet, an ECFP fingerprint plus a multi-layer perceptron achieved AUC 0.984 on the in-distribution "Large" subset, beating all sixteen deep encoders Activity Cliff Prediction: Dataset and Benchmark; on MoleculeACE, SVM plus ECFP gave the lowest average error on cliff compounds across 30 targets Exposing the Limitations of Molecular Machine Learning with Activity Cliffs; and a graph convolutional network was the best of a weak field out-of-distribution at AUC 0.579 Activity Cliff Prediction: Dataset and Benchmark. We will run all three under identical conditions: the same target-split folds, the same anchor pairs made available to every method as ordinary training data, the same random-search hyperparameter budget, and the same compute. Beating them means a mean out-of-distribution AUC of at least 0.70 across held-out targets with no more than 50 anchor pairs, and a lower cliff-specific RMSE on MoleculeACE at matched overall RMSE. We consider a gain smaller than 0.10 AUC over 0.579 a failure, not a success.

The critical ablation is the same warp machinery with the ruler removed: fine-tune the identical embedding on the identical anchor pairs with an ordinary regression loss, no fixed-distance constraint. This isolates the proposal's actual ingredient — the standard-scale anchoring, not per-target adaptation per se. If unconstrained fine-tuning matches the ruler-anchored warp, the cosmological principle contributed nothing and the claim reduces to "fine-tuning helps," which we would report as a negative result. For the wet-lab arm, the empirical ceiling is Cell Painting cosine-similarity selection itself, which doubled hit rates when picking the top 5% most phenotypically similar compounds across 65 assays Large scale compound selection guided by cell painting reveals activity cliffs and functional relationships; it is not a predictor of untested compounds, but it bounds what measurement without prediction can achieve.

Why not just…?

Why not just use fingerprints and classical QSAR? In-distribution, this wins, full stop: ECFP plus an MLP hit AUC 0.984 where deep models did worse Activity Cliff Prediction: Dataset and Benchmark, and SVM plus ECFP was the best cliff performer on MoleculeACE Exposing the Limitations of Molecular Machine Learning with Activity Cliffs. Our proposal does not compete on that turf; it targets the regime where the same benchmarks show everything collapsing to roughly 0.579 AUC on unseen targets. If fingerprints match our warp out-of-distribution under the matched-anchor protocol, the idea is falsified, and we have designed the experiment so that this outcome is visible rather than arguable.

Why not just run Cell Painting and pick by cosine similarity? Because profiling is a measurement, not a predictor: it requires imaging every candidate, it inherits batch effects across the 12 laboratories of JUMP-CP that demand sphering and inverse-normal correction, and the bimodal effect across 30 pathways means similarity rankings cannot tell you which side of a phenotypic split an untested analogue will land on Large scale compound selection guided by cell painting reveals activity cliffs and functional relationships. We use Cell Painting as ground truth, not as the method.

Why not just tune the pH chemistry empirically? This is the most serious objection, and we treat it as such. Empirical tuning demonstrably works for single agents: the Bi@ZIF-8 platform achieved 24.4% photothermal conversion efficiency and complete tumor disappearance over 14 days with no systemic toxicity Rapid synthesis of a Bi@ZIF-8 composite nanomaterial as a near-infrared-II (NIR-II) photothermal agent for the low-temperature photothermal therapy of hepatocellular carcinoma. But the micelle study needed exact co-encapsulation of multiple adjuvants to reach 87.89% apoptosis Author Correction: pH-sensitive micelles for the intracellular co-delivery of curcumin and Pluronic L61 unimers for synergistic reversal effect of multidrug resistance, and bimodal pathways break the monotonicity assumption outright. Our release-rate experiment is built precisely to settle this: if varying release rate alone never flips a phenotype for cliff-adjacent payloads, then the coupling claim — and with it the main reason to need the calibrated geometry — fails, and we will say so.

Why not just fine-tune a model on each new target? Because fine-tuning is exactly what the ablation tests. The literature's transfer collapse Activity Cliff Prediction: Dataset and Benchmark is evidence that unconstrained adaptation memorizes target-specific patterns from sparse data; the cosmological precedent is that a correction constrained to an invariant scale keeps working where a free-form model would overfit, and that its benefit has a measurable data-density floor Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys. The ablation, not rhetoric, decides whether that constraint is what makes the difference.

Decisive Checks

Four outcomes decide this proposal. First, out-of-distribution AUC of at least 0.70 on ACNet target splits with at most 50 anchors per target, against the documented 0.579 ceiling Activity Cliff Prediction: Dataset and Benchmark. Second, an anchor-density operating curve showing a threshold below which the warp hurts — the direct molecular analog of the reconstruction density limits cosmology measured Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys — because a method that claims to work at any data density has not been characterized. Third, pre-registered phenotype-flip predictions in the pH-release experiment that beat chance, which alone would establish that release kinetics and cliff geometry interact measurably. Fourth, the ablation: ruler-anchored warp must beat unconstrained fine-tuning, or the cosmology was decoration. Two of the four failing would kill the idea honestly; all four passing would give medicinal chemistry something it currently lacks — a calibrated, transferable geometry of the structure-to-function map — and give drug delivery a principled way to choose a release rate.