Sideband Background Modeling for Virtual Screening

Screening a chemical library for molecules that bind a disease target is, structurally, the same problem a dark-matter experiment faces: a handful of genuine signals buried under a flood of background, with a classifier whose apparent skill may be an artifact of how the background was assembled. That worry is not hypothetical in drug discovery. A CNN trained on the standard DUD-E screening benchmark scores an AUC of 0.98 even when the receptor is deleted and replaced by a dummy atom, and collapses to 0.50 — pure chance — when evaluated on the "Actives as Decoys" (AD) benchmark, where the decoys are real actives of other proteins rather than synthetically matched fillers (Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening). In other words, today's supervised scorers largely learned what a manufactured decoy looks like, not what binding looks like — and that paper explicitly left open how to build models that cannot learn such artifacts in the first place.

Rare-event physics has a mature answer to precisely this failure mode. In a CMS analysis of collider open data, the CATHODE procedure trains a density model on "sidebands" — data regions guaranteed to contain almost no signal — interpolates that learned background into the signal region, and then trains a classifier to distinguish real data from synthetic background; because no signal or decoy labels are ever used, the classifier cannot memorize label artifacts, and the significance of a buried $\Upsilon \to \mu^+\mu^-$ resonance rose from $1.6\sigma$ with a simple invariant-mass cut to $6.4\sigma$ (Isolating Unisolized Upsilons with Anomaly Detection in CMS Open Data). This proposal transplants and adapts that sideband idea to structure-based virtual screening, where the analogue of the signal-free sideband is the vast low-scoring bulk of the library itself.

Core Claim

A virtual-screening ranker can be made robust to decoy and dataset bias by learning an explicit, label-free model of the background library and scoring compounds by how improbable they are under that model — without ever training on actives-versus-decoys labels. Concretely, we will build a conditional normalizing flow — a generative model that learns a probability density over feature vectors — trained on unlabeled library compounds in low-docking-score "sideband" bands, stratified by chemical class, and extrapolated into the high-score tail where true binders must live. The output is a per-compound likelihood ratio and a local false discovery rate (LFDR): the posterior probability that a top-ranked compound is background. The decisive claim is empirical: this ranker maintains enrichment on the AD benchmark, where supervised models trained on DUD-E collapse to 0.50 AUC, while remaining competitive on standard benchmarks.

How This Differs From Existing Work at the Intersection

The closest existing work is SubDyve, which already imports false-discovery thinking into screening: it mines discriminative 2D subgraphs from labeled actives, propagates activity over a molecular network, and uses an LFDR as a statistical brake on seed expansion, gaining up to $+34.0$ BEDROC and $+24.6$ $\text{EF}_{1\%}$ over baselines on DUD-E (SubDyve: Subgraph-Driven Dynamic Propagation for Virtual Screening Enhancement Controlling False Positive). But SubDyve is supervised at its core: it needs labeled actives, its subgraph mining costs 17–18 minutes of target-specific preprocessing, and it is confined to 2D ligand topology with no receptor structure. Our proposal inverts the logic: instead of guarding a supervised signal-chaser with statistics, we learn the background directly from the unlabeled library, use no decoy labels and no per-target labeled actives, and wrap any 3D scorer. Three contrastive-learning systems — Uni-Clip, which aligns four protein and ligand encoders and even handles "false negative collisions" in its loss (Enhancing Challenging Target Screening via Multimodal Protein-Ligand Contrastive Learning); SaBAN, which maps residue–fragment interactions with bilinear attention (Structure-Aware Contrastive Learning with Fine-Grained Binding Representations for Drug Discovery); and S-MolSearch, which distills a supervised encoder into a student over 19M unlabeled compounds via inverse optimal transport (S-MolSearch: 3D Semi-supervised Contrastive Learning for Bioactive Molecule Search) — all produce uncalibrated similarity scores with no generative background model and no false-discovery control in the ranked tails, which is exactly the gap we fill. The hidden-bias paper motivates the problem and supplies the AD benchmark but offers no method (Hidden bias in the DUD-E dataset...); the CMS paper supplies the method but in a domain where sidebands are defined by a physically meaningful observable (Isolating Unisolized Upsilons...). The transplant is non-trivial, and we say below where it can fail.

Research Program

The pipeline has four components. First, the "mass-like" observable that defines signal and sideband regions must itself be free of learned decoy bias, so we use a physics docking score (AutoDock Vina) rather than any model trained on DUD-E: the signal region is the top scoring band of the library, and sidebands are the adjacent lower bands. Second, each docked compound is described by a feature vector of secondary observables: ligand physicochemical descriptors, pose geometry (contact counts, buried surface, a lightweight clash descriptor — relevant because AI-generated poses are frequently unphysical, with SurfDock's raw stereochemical validity at only 41.7% before relaxation on the PoseX self-docking set (PoseX: AI Defeats Physics-based Methods on Protein Ligand Cross-Docking)), and pocket-residue interaction profiles. Third — the key methodological adaptation — because chemical space is far more heterogeneous than a collider feature space, a single flow trained on low-score compounds would face a severe extrapolation problem in the high-score tail. We therefore stratify the library into chemotype/property bins and train a flow conditioned on score and stratum, $p_\phi(x \mid s, c)$, on sideband bands within each stratum, then evaluate it in the signal region. Fourth, we either use the flow density directly or follow CATHODE and train a discriminator between real signal-region compounds and flow-sampled synthetic background; both routes yield a density ratio. The tail-level statistic is the local false discovery rate

$$\mathrm{LFDR}(x) \;=\; \frac{\pi_0 \, p_\phi(x \mid s > s_{\mathrm{hi}}, c)}{p_{\mathrm{data}}(x \mid s > s_{\mathrm{hi}}, c)},$$

where $\pi_0$ is the (near-one) prior fraction of background compounds, $p_\phi$ is the learned background density, and $p_{\mathrm{data}}$ is the observed density of real library compounds in the signal region. Compounds with small LFDR are those whose features are poorly explained as background. A secondary ingredient, borrowed back from SubDyve's logic but recontextualized (SubDyve), is to report LFDR-calibrated top-$N$ lists with a target false-positive budget, validated empirically.

The honest research risk is tail extrapolation: unlike CATHODE, which interpolates between sidebands flanking a mass peak, we extrapolate a conditional density into a tail. We mitigate this with a "sideband ladder" validation — the flow's predicted background counts are checked against held-out data in each successively higher sideband band before its tail predictions are trusted — and with the stratified conditioning above. If the ladder check fails, that is itself a finding about how far background-learned screening can be pushed.

Proposed Experiments

The headline evaluation is the AD benchmark of Hidden bias..., where ligand-only supervised models drop from 0.98 AUC on DUD-E to 0.50: our method should hold AUC substantially above chance (we pre-register $\geq 0.65$) precisely because it never saw decoy labels to memorize. Three controls guard against self-deception, mirroring the "Same-Sign" control that the CMS analysis used to detect artificial background sculpting (Isolating Unisolized Upsilons...): a dummy-receptor control (does our enrichment vanish when the receptor is replaced, proving we use binding information rather than ligand artifacts?), a shuffled-receptor control (does the pipeline produce null enrichment when receptor assignments are scrambled?), and the sideband-ladder fidelity check. Standard-benchmark competitiveness is measured on LIT-PCBA and DUD-E, and scalability on the PU-style CDK7 task of 1,468 actives hidden in 10M ZINC20 compounds (SubDyve); on the CDK7 library the retrieval metric for SubDyve itself has no reported number in the Lacuna summary, so we will reproduce SubDyve under matched compute for the head-to-head.

Baselines

The "do nothing" reference is chance ranking, which fixes EF$_{1\%}=1$ and AUC $=0.5$; on the AD benchmark this is not a strawman but the empirical fate of supervised models, which fall to 0.50 AUC there (Hidden bias...). A second naive reference is cut-and-count on the Vina docking score itself — the analogue of the $1.6\sigma$ mass-cut in the CMS study, which the multifeature likelihood ratio raised to $6.4\sigma$ (Isolating Unisolized Upsilons...); our method must beat Vina's own ranking to justify its existence. The strongest published competitors on our benchmarks are SaBAN, which reports LIT-PCBA AUROC of 68.16% and BEDROC of 13.35% (Structure-Aware Contrastive Learning...), and S-MolSearch, which reports DUD-E AUROC of 92.56% against DrugCLIP's 80.93% and LIT-PCBA AUROC of 61.78% (S-MolSearch); we will rerun both under our identical feature budget and tuning protocol, and where reruns are infeasible use their published numbers with matched split definitions. SubDyve is the strongest competitor specifically on false-positive control (+34.0 BEDROC, +24.6 EF$_{1\%}$ over baselines on DUD-E) and is run head-to-head on the CDK7 10M task. The decisive ablation isolates our key ingredient: an architecturally identical discriminator trained with conventional supervision (DUD-E actives versus DUD-E decoys) instead of flow-sampled background. If that ablation matches us on the AD benchmark, the sideband ingredient is doing nothing and the proposal fails; we predict it collapses to chance, as the hidden-bias diagnostics imply. A second ablation removes the stratified conditioning, testing whether chemistry-space stratification is what saves the tail extrapolation.

Why not just…?

Why not just retrain supervised models on unbiased decoys like the AD set? This is the most serious alternative. The AD benchmark shows the bias is fixable in principle, and supervised learning is stronger than density estimation when good labels exist. But the AD set is a diagnostic, not a training resource: real screening libraries of $10^7$–$10^{10}$ compounds come with no labels at all, and any decoy set one constructs reintroduces a construction signature — the exact mechanism behind the 0.98-to-0.50 collapse (Hidden bias...). Our method needs zero decoy labels by construction. The experiment that settles it: train the supervised ablation on AD-style decoys and test transfer to held-out targets; if it matches our AD performance and transfers, supervised-with-honest-decoys wins and we will say so.

Why not just dock everything with physics (Vina, OpenMM relaxation)? Physics force fields carry no learned bias, which is exactly why we use a Vina score as our region-defining observable — the objection partly dissolves. But pure physics ranking is both too slow at scale (SubDyve ran ~12.3× faster than AutoDock Vina on a 10M library; Uni-Clip reports a 300-fold speedup over AutoDock GPU SubDyve, Uni-Clip) and demonstrably not the ceiling for pose quality (PoseX). Our contribution is a cheap statistical layer on top of physics scores, not a replacement for them.

Why not just use 2D fingerprint similarity to known actives? It is the cheapest baseline and it fails in the regimes that matter: fingerprint models are among those that collapse on AD, they cannot cross activity cliffs (S-MolSearch), and they degrade in few-seed, novel-target settings where SubDyve's seeds are scarce (SubDyve). We include it as a reference anyway, because a method that cannot beat similarity search on LIT-PCBA should not be deployed.

Why not just recalibrate the existing scorer's outputs? Isotonic or Platt scaling fixes the marginal distribution of scores, not their ranking; a scorer whose ranking is driven by decoy artifacts is confidently wrong, and no monotone rescaling changes the order. The ligand-only CNN's 0.98 AUC on DUD-E was likely well calibrated on its own training data and still meaningless on AD (Hidden bias...). Calibration is complementary — we need it for the LFDR readout — but it cannot substitute for learning the background.

Decisive Checks

The proposal succeeds if four checks pass together: the sideband ranker holds AUC $\geq 0.65$ on the AD benchmark while the supervised ablation sits near 0.50; the dummy-receptor and shuffled-receptor controls show null enrichment, proving the signal comes from receptor-aware features and not ligand artifacts or sculpting; the sideband ladder shows the conditional flow reproduces held-out background counts band by band, licensing the tail extrapolation; and LFDR calibration is honest — the empirical false-positive fraction among compounds flagged below a 10% LFDR threshold actually lands near 10% on the CDK7 10M task. If the ladder check fails while AD performance holds, the result would paradoxically suggest the method is exploiting residual library structure rather than binding physics, and we would report it as such.