Allosteric drugs bind proteins somewhere other than the active site and change the protein's behavior indirectly — through a structural ripple rather than by blocking the catalytic cleft. This indirect action gives them two properties that matter here: they are state-dependent (they preferentially hit proteins in particular conformations or cell states) and they tune rather than abolish function. A use-dependent calcium-channel blocker that inhibits only hyper-activated neurons — showing roughly eight-fold higher potency at inactivated channels ($IC_{50} = 1.3\ \mu$M) than resting ones ($10.2\ \mu$M) — while leaving normal signaling intact, is the archetype C2230, a preferential use- and state-dependent CaV2.2 channel blocker. That kind of partial, state-selective pharmacology is exactly what modern single-cell and high-content screens should be able to see — and, I will argue, exactly what current screening pipelines are built to destroy.
The reason is an unfortunate collision. High-throughput screens are riddled with batch effects: systematic non-biological differences between plates, reagent lots, and labs that can overwhelm biological signal, so that samples cluster by laboratory rather than phenotype Cross-platform analysis of cancer microarray data improves gene expression based classification of phenotypes. The field's answer has been to make representations batch-invariant by averaging: pairing replicate wells and pooling cells. But averaging is also the classic cause of heterogeneity masking — the loss of rare, biologically decisive subpopulations, such as the stem-like cells that drive tamoxifen resistance in breast cancer Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties or the high-plasticity subpopulations invisible to population-level profiling A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers. A state-dependent allosteric modulator produces precisely a subpopulation-specific signal. Current batch-correction therefore risks classifying the therapeutic signature of allostery as noise and filtering it out.
Core Claim
I propose that allosteric and orthosteric inhibitors of the same target protein leave distinguishable signatures in single-cell perturbation data — not in the direction of the population mean, but in the shape of the within-treatment cell-state distribution — and that these signatures can be recovered only by a batch-correction scheme that aligns distributions rather than averages them. Concretely: an orthosteric competitive inhibitor shifts the whole population; an allosteric modulator redistributes it, concentrating its effect in the cell states where the target is active or in the conformation the drug stabilizes. The research program has three parts: a representation-learning objective that enforces cross-batch replicate consistency at the distribution level while preserving per-cell structure; a quantitative "state-selectivity" readout that summarizes where a compound's effect mass lands across baseline cell states; and a set of experiments testing whether this readout separates allosteric from orthosteric compounds on existing benchmarks, and whether current pipelines miss it.
What Prior Work Does and Does Not Do
No method I found on Lacuna attempts this. The closest neighbors each solve one corner of the problem. Weakly Supervised Set-Consistency Learning Improves Morphological Profiling of Single-Cell Images (Set-DINO) achieves batch robustness by averaging sets of 16 cells — the authors explicitly describe the mean as a low-pass filter that washes out single-cell variance, which is the very signal I want to keep; it reaches 46.9% biological recall on the CORUM database at a 10% cutoff versus 35.6% for engineered features, but its design goal is the opposite of mine. Asymmetric Contrastive Objectives for Efficient Phenotypic Screening (SPC) learns batch-invariant morphology and clusters compounds by mechanism of action, but its prototypes receive only attractive updates, so an intermediate or subpopulation-specific phenotype risks being pulled into the wrong cluster; it operates on population-level features and never asks whether allostery is visible. Efficient Cell Painting Image Representation Learning via Cross-Well Aligned Masked Siamese Network (CWA-MSN) pairs masked anchors with same-treatment cells from other wells, beating CellCLIP by 9% on gene-gene retrieval with 22M parameters — a mechanism I reuse, but its positives are individual cells, not distributions. On the transcriptomic side, BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes shows that global alignment erases biologically real subtypes when batch compositions are unbalanced — the central cautionary result for my design — and scPerturb: Harmonized Single-Cell Perturbation Data supplies both 44 harmonized perturbation datasets (including small molecules) and the energy-distance statistic I repurpose as a training objective. Finally, two results constrain the design: Deep Learning for BioImaging: What Are We Really Learning? shows that on RxRx3-core an untrained Vision Transformer reaches Recall@5% of ~0.19 versus ~0.20 for pretrained models, meaning benchmarks can be passed on low-level shortcuts; and PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction shows zero-shot single-cell foundation models barely beat a mean baseline (scGPT at AUSPC ~0.045 versus GEARS at ~0.008 on Norman), so I will not lean on foundation-model embeddings.
Research Program
The method — call it distribution-level replicate alignment — trains a per-cell encoder (a small vision or transcriptomic encoder, in the spirit of SPC's 11M-parameter ResNet beating a 307M-parameter model) with one objective. Let $z_{t,b}$ denote the set of per-cell embeddings for treatment $t$ in batch $b$, and let $E(\cdot,\cdot)$ be the empirical energy distance between two point sets (twice the mean cross-set distance minus the mean within-set distance — zero only when the full distributions match, not just their means). The loss aligns replicate distributions of the same treatment across batches while separating different treatments:
$$\mathcal{L} \;=\; \sum_{t}\;\sum_{b \neq b'} E\!\left(z_{t,b},\, z_{t,b'}\right) \;-\; \lambda \sum_{t \neq t'} E\!\left(z_{t},\, z_{t'}\right)$$
The crucial contrast with Set-DINO is that matching means is cheap and destroys subpopulation structure, whereas matching distributions forces the encoder to preserve the shape of the cell-state distribution — including rare states — while still discarding whatever does not replicate across batches. This yields an identification strategy for the hardest confound: a subpopulation-specific drug effect and a batch-composition artifact look identical within one plate, but only the drug effect reproduces in the same subpopulation across independent batches. Replication at the subpopulation level, not global invariance, is what separates the two — the lesson BERMUDA taught for cluster-matched alignment, applied to perturbation screens.
On top of the encoder sits the readout. Partition unperturbed cells into $K$ baseline cell-state clusters; for a compound, distribute its effect mass (per-cell effect magnitudes, measured as distance to the matched unperturbed cluster) across those clusters, giving proportions $p_t$. The state-selectivity index is
$$S_t \;=\; 1 - H(p_t)/\log K,$$
where $H$ is Shannon entropy. A pan-popularion inhibitor scores near zero; a state-dependent modulator scores high. The biological prior comes from allosteric pharmacology itself: non-competitive inhibition that persists under substrate accumulation Cytoplasmic isoleucyl tRNA synthetase as an attractive multistage antimalarial drug target and modulation of only hyperactive receptor states C2230, a preferential use- and state-dependent CaV2.2 channel blocker both predict concentrated, state-biased effect mass; the structural basis for such specificity — remote pockets under weaker evolutionary constraint than catalytic clefts — is documented for allosteric inhibitors of UBE2T/RACK1 signaling Correction: A novel UBE2T inhibitor suppresses Wnt/β-catenin signaling hyperactivation and gastric cancer progression by blocking RACK1 ubiquitination. A key enabling component is an annotation layer: map screened compounds to binding-site class (orthosteric, allosteric, non-canonical) using ChEMbl and UniProt, which appear as knowledge resources in AI approaches for the discovery and validation of drug targets, and define same-target compound pairs where both classes exist.
Proposed Experiments
Four experiments, all on data in the file. First, a diagnostic that requires no new model: within the drug-to-knockdown retrieval framing of Cell-Level Virtual Screening — which reports Hits@1 of 0.1250 on DDR-Bench against a random baseline of 0.0179 — ask whether allosteric compounds are disproportionately the misses when a drug's effect is treated as a synonym for its target's knockdown. The premise of that benchmark, that matched chemical agents act as inhibitory synonyms of knockdowns, should fail most for allosteric modulators, which tune without eliminating the target. A measurable retrieval gap between allosteric and orthosteric compounds would be the first quantitative evidence for Gap 3 and the motivating result for the whole program.
Second, the imaging test on BBBC021 (113 compounds, multiple batches, mechanism-of-action labels): train the distribution-aligned encoder, annotate compounds by binding-site class, and evaluate AUROC for allosteric-versus-orthosteric discrimination within same-target pairs, while reporting Not-Same-Compound-or-Batch to confirm batch invariance has not been sacrificed. Third, the transcriptomic test on the small-molecule datasets within scPerturb, respecting its own statistical floor of at least 300 cells per perturbation and 1,000 UMIs per cell, and exploiting the multi-study structure (25 publications) as a natural cross-batch setting — including cross-dataset transfer of the state-selectivity index. Fourth, a synthetic ground-truth test: plant a known subpopulation-specific effect in one batch alongside a batch artifact that mimics it compositionally, and measure recovery rates for distribution-level alignment versus mean pooling. As a sanity check on representation quality, evaluate gene-gene retrieval on RxRx3-core, where CWA-MSN's reported margins (+9% over CellCLIP, +29% over OpenPhenom) define the competitive band, and include the untrained-ViT control so that any allosteric discrimination we claim cannot be explained by pixel or cell-density shortcuts.
Baselines
The naive reference point is an untrained Vision Transformer with random weights plus the energy-distance statistic on raw features. This baseline matters because Deep Learning for BioImaging: What Are We Really Learning? reports that such a model achieves Recall@5% of ~0.19 on RxRx3-core, nearly matching pretrained models at ~0.20; if random features separate allosteric from orthosteric compounds as well as my encoder, the task is measuring low-level artifacts, not pharmacology, and the proposal fails its own control.
The strongest existing competitor is SPC, which reports a Not-Same-Compound-or-Batch score of 96.7% on BBBC021, a 10.2% improvement over weakly supervised baselines Asymmetric Contrastive Objectives for Efficient Phenotypic Screening. It is the strongest because it already solves batch invariance with small models and mechanism-of-action structure — the environment in which an allosteric readout would plausibly live. I will run it on the same BBBC021 and scPerturb splits with the same compute and tuning budget, freeze its embeddings, and fit a linear probe for allosteric-versus-orthosteric discrimination on top. Beating it means a higher same-target discrimination AUROC while staying within about two points of its NSCB, so the gain cannot be attributed to trading away batch robustness.
The ablation isolating my key ingredient replaces distribution-level alignment with Set-DINO-style mean pooling of $n=16$ cells on the same encoder and data. Set-DINO's own numbers — 46.9% CORUM recall versus 35.6% for engineered features Weakly Supervised Set-Consistency Learning Improves Morphological Profiling of Single-Cell Images — establish it as a strong batch-robust method; the prediction specific to my claim is that mean pooling collapses the planted and real subpopulation-specific effects that distribution alignment recovers, which the synthetic-confound experiment measures directly. A classical reference completes the set: hand-engineered CellProfiler features (90–94% Not-Same-Compound accuracy on BBBC021, versus 97% for the weakly supervised CNN of Weakly Supervised Learning of Single-Cell Feature Embeddings) with per-plate negative-control z-scoring and scPerturb's E-distance.
Why not just…?
Why not just take existing batch-invariant embeddings (SPC, CWA-MSN) and post-hoc probe for allostery? This is the most serious alternative and I will run it as the headline baseline rather than argue around it. The evidence that it falls short is structural, not empirical hand-waving: SPC's prototypes receive only attractive updates, so an allosteric compound whose phenotype partially overlaps several orthosteric classes gets pulled into the wrong cluster rather than standing apart; and CWA-MSN's cross-well positives are single cells, which stabilizes training but provides no mechanism for distinguishing a replicating subpopulation effect from a batch artifact. Still, if the frozen-SPC probe matches my encoder on allosteric discrimination, the representation contribution is dead — and I would report the finding (allostery is readable in existing screens) as the result, since that alone changes how these screens are used.
Why not just normalize per plate and compute energy distances on engineered features? This needs no deep learning at all: z-score each plate to its negative controls and use scPerturb's E-distance, which was designed precisely to quantify distribution-level effect sizes. It is a genuine threat and my second baseline. Where it falls short is identification: E-distance within a plate cannot tell a subpopulation-specific drug effect from a batch-composition shift, and BERMUDA showed that when compositions are unbalanced, global correction actively erases the affected subtypes BERMUDA, while Cross-platform analysis of cancer microarray data showed residual lab-level clustering surviving standard normalization. My cross-batch replicate structure at the subpopulation level is the proposed fix; if the simple pipeline matches on BBBC021, the synthetic-confound experiment is the deciding ground, because only there the ground truth distinguishes the two.
Why not just fix the knockdown-equals-inhibitor assumption inside CellVS-Net? That would patch Gap 3 where it was identified, but it operates on LINCS L1000 data, which is population-averaged by construction — so a state-dependent effect concentrated in 10% of cells is diluted ten-fold before any model sees it, the same masking that hides rare resistant clones in bulk sequencing Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance. The fix belongs at single-cell resolution; my first experiment nonetheless uses the CellVS-Net benchmark to quantify how much the assumption costs, which is the evidence a reviewer would need before accepting the more expensive single-cell route.
Decisive Checks
Three outcomes would settle the proposal. If the state-selectivity index separates allosteric from orthosteric same-target compounds across BBBC021 and the scPerturb small-molecule datasets, at effect sizes that survive the untrained-ViT control, the core claim holds. If the synthetic-confound test shows distribution-level alignment recovering planted subpopulation effects that mean pooling removes — and per-plate E-distance cannot distinguish from artifacts — the mechanism is validated. And if either the frozen-SPC probe or the per-plate E-distance pipeline matches my full method, I will report that honestly: the finding that allosteric pharmacology is legible in existing single-cell screens would then be the contribution, and the simpler pipeline its recommended instrument.
References
- C2230, a preferential use- and state-dependent CaV2.2 channel blocker, mitigates pain behaviors across multiple pain models
- Cross-platform analysis of cancer microarray data improves gene expression based classification of phenotypes
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers
- Weakly Supervised Set-Consistency Learning Improves Morphological Profiling of Single-Cell Images
- Asymmetric Contrastive Objectives for Efficient Phenotypic Screening
- Efficient Cell Painting Image Representation Learning via Cross-Well Aligned Masked Siamese Network
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes
- scPerturb: Harmonized Single-Cell Perturbation Data
- Deep Learning for BioImaging: What Are We Really Learning?
- PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction
- Cytoplasmic isoleucyl tRNA synthetase as an attractive multistage antimalarial drug target
- Correction: A novel UBE2T inhibitor suppresses Wnt/β-catenin signaling hyperactivation and gastric cancer progression by blocking RACK1 ubiquitination
- AI approaches for the discovery and validation of drug targets
- Cell-Level Virtual Screening
- Weakly Supervised Learning of Single-Cell Feature Embeddings