Pulse-Shape Discrimination for Efflux Pump Screens

Efflux pumps are membrane transporter proteins — most famously P-glycoprotein — that bacteria and cancer cells use to actively eject therapeutic drugs before those drugs can act, which is a main route of multidrug resistance. A promising countermeasure is small-molecule adjuvants that disable the pump: for example, ribociclib physically jams the substrate-binding pocket of P-glycoprotein, and the BTK inhibitor RN486 locks the transporter in an inactive conformation by binding two sites at once (Ribociclib Inhibits P-gp-Mediated Multidrug Resistance in Human Epidermoid Carcinoma Cells; Bruton's Tyrosine Kinase (BTK) Inhibitor RN486 Overcomes ABCB1-Mediated Multidrug Resistance in Cancer Cells). But finding such adjuvants by high-throughput screening is notoriously unreliable, because many "hits" never touch the pump at all: they quench the fluorescent reporter, scatter light, or otherwise trick the instrument, and the screen cannot tell these optical artifacts from real biology. Meanwhile, an adjacent community has spent decades solving a structurally identical problem: rare-event physicists pull a handful of genuine neutrino or dark-matter events out of overwhelming instrumental noise, and they learned that the trick is to read the shape of the detector waveform, not just its total size.

This proposal ports that machinery. The core observation is a clean physical analogy: genuine efflux inhibition is slow biology — intracellular dye accumulates over minutes, with a rate that depends on inhibitor dose and on how many pumps the cells express — whereas optical interference is essentially instantaneous and dose-independent in shape. This is the same distinction liquid-argon dark matter detectors exploit, where nuclear recoils (real signal) produce a sharp burst of scintillation photons while electron recoils (background) produce slow, dispersed emission, separable by the ratio of prompt to total light (First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB). We propose to discriminate real efflux modulation from assay artifacts at the level of the time-resolved optical readout itself, using the two-stage architecture proven in germanium detectors by Deep learning based pulse shape discrimination for germanium detectors: an unsupervised convolutional autoencoder — a network that learns to compress and reconstruct each raw time trace, filtering noise in the process — followed by a classifier of only ~100–200 parameters.

Core Claim

The strongest existing defense against screening artifacts predicts interference from the chemical structure of the compound before it is ever tested: in Identification of Compounds That Interfere with High-Throughput Screening Assay Technologies, Laurianne David et al. train Random Forest Classifiers on molecular fingerprints to flag "compounds that interfere with assay technologies" (CIATs). That method reaches ROC AUC 0.99 on compounds previously screened across multiple assays, but only 0.64 on novel single-assay compounds, and on the public PubChem counter-screen data it caught just 35% of interferents. Our claim is that discrimination should instead happen after the measurement, on the readout itself, because the artifact manifests in the waveform's temporal morphology, not in the molecular graph — and that a physics-style autoencoder pipeline can push novel-compound interference rejection well beyond what structure-based prediction achieves, while also catching run-specific instrumental problems (baseline drift, photobleaching) that no structure-based model can even in principle see.

Research Program

The program has three components: a labeled time-resolved benchmark (which does not yet exist anywhere), the discrimination architecture, and physics-derived scalar features that serve both as baselines and as sanity checks.

The data problem must be faced honestly. No public corpus of time-resolved efflux assay readouts was located: no reported result found on Lacuna for public time-resolved efflux assay waveform datasets, and the artifact-assay data of David et al. was used only through compound fingerprints, not raw instrument traces. We will therefore build the benchmark in two tiers. First, a forward-model simulator: a transporter-kinetics layer (dye influx against concentration-dependent efflux, producing slow accumulation curves for true inhibitors) composed with an optics layer that models known artifact physics — fluorescence quenching, inner-filter effects, light scattering, photobleaching, and detector baseline drift — to generate large labeled trace datasets with known ground truth. Second, a wet-lab pilot: Calcein-AM and rhodamine-123 accumulation assays in P-glycoprotein-expressing cells, read at fine time resolution (on the order of one point every 30 seconds for an hour), with a compound panel assembled from three sources — established inhibitors, with verapamil as the internal reference standard as recommended by Annotating Human P‐Glycoprotein Bioassay Data; known CIATs drawn from the AstraZeneca and PubChem artifact-assay sets of David et al.; and matched presumptively inactive compounds, held out until evaluation.

Training follows the calibration-by-proxy strategy that makes the physics pipeline work with almost no labeled signal. Holl et al. could not obtain true neutrinoless double-beta decay events, so they trained on calibration spectral peaks — the double-escape peak as a clean proxy for signal-like pulses, the full-energy peak as a proxy for background — and reached 90% background rejection at 90% signal acceptance. Our proxies are the assay's own controls: acellular or pump-free wells give pure interference curves (optics with no biology), and verapamil time courses give canonical true-inhibition curves (biology with no artifact). These play exactly the role of the physics calibration peaks, letting the small supervised classifier be trained without any assumption about novel compounds. The autoencoder itself is trained unsupervised on all traces, as in the germanium work, so it learns the manifold of plausible readouts rather than the identities of specific compounds. We also build the physics-inspired scalar summaries as explicit comparators: most importantly a prompt-fraction analog of the liquid-argon discriminator,

$$F_{\text{prompt}} = \frac{I(t_1) - I(t_0)}{I(t_{\infty}) - I(t_0)},$$

where $I(t)$ is the measured reporter intensity, $t_0$ the first read, $t_1$ an early read, and $t_{\infty}$ the final read — the fraction of total optical change that happens immediately, which should be large for interferents and small for genuine efflux modulation.

As a secondary thread, we will test whether the same compact-bottleneck discipline helps on the existing structure-only benchmarks, where models are known to collapse under realistic splits: the 822-compound substrate/non-substrate dataset of [ADMET Evaluation in Drug Discovery.

  1. Development of in Silico Prediction Models for P-Glycoprotein Substrates](/work/admet-evaluation-in-drug-discovery-13-development) (independent-test accuracy 83.5%, Matthews correlation coefficient 0.667) and the 1273-compound multidrug-resistance-reversal dataset catalogued in Computational models for predicting substrates or inhibitors of P-glycoprotein, evaluated under scaffold splits in the spirit of Therapeutics Data Commons, which showed Pearson correlations falling from roughly 0.7 in-distribution to roughly 0.4 out-of-distribution. This thread is secondary because the primary claim is about readouts, not fingerprints.

Baselines

The naive reference point is doing nothing: accept the raw hit list from the screen, in which false positives occur at the natural rate of optical interferents in the compound library. This establishes the floor that any filtering must beat, and no quantitative prevalence figure for interferents in efflux-specific libraries was found on Lacuna, so the pilot itself will measure it. The rule-based competitors are the PAINS substructure filter and the Binomial Survivor Function; David et al. report that PAINS achieves ROC AUC of only 0.50 on multi-assay compounds and 0.51 on novel compounds, and correctly flags just 9% of interferents in AlphaScreen assays and 1.5% in FRET and TR-FRET assays, while the Binomial Survivor Function reaches 0.63 and 0.57 respectively and requires a compound to have been screened roughly 10–30 times before its statistics become informative. These will be run with their standard settings on the identical held-out compound panel.

The strongest competitor is the Random Forest Classifier on ECFP4 fingerprints from Identification of Compounds That Interfere with High-Throughput Screening Assay Technologies, the best method in the literature we have. It is strongest because it is the only existing approach with demonstrated, quantified performance on exactly this failure mode. To make the comparison fair, we will retrain it on our pilot data with the same hyperparameter search budget given to our model, evaluate it on exactly the same novel-compound hold-out and the same labels, and report both ROC AUC and the physics-style criterion of interferent rejection at fixed 90% true-inhibitor acceptance. The published numbers to beat are ROC AUC 0.64 on novel single-assay compounds and, on external PubChem counter-screens, mean ROC AUC 0.67 with 35% of interferents detected and precision 0.63. We would count as beating it if, on novel compounds never previously screened, we reach ROC AUC of at least 0.85 and reject at least 80% of interferents while retaining 90% of true inhibitors.

Finally, the key ablation isolates the waveform-morphology ingredient: the identical tiny classifier is fed either the autoencoder's low-dimensional latent representation of the full time trace, or the scalar kinetic summaries alone (endpoint value, initial rate, and the prompt fraction defined above). If the latent representation does not outperform the scalars, the proposal's central mechanism — that subtle multi-time-constant structure in the trace carries information a single ratio misses — is wrong, and we would report that the physics-derived scalar features, not the autoencoder, carry the value.

Why not just…?

Why not just use PAINS filters or hit-rate statistics? The evidence is directly against it: on experimentally measured artifact assays, PAINS performs at chance (ROC AUC 0.50–0.51) and misses 91–98.5% of interferents, and the Binomial Survivor Function fails precisely in the regime that matters for new discovery, where a compound has no screening history (Identification of Compounds That Interfare with High-Throughput Screening Assay Technologies). Both are still included as baselines so the comparison is empirical, not rhetorical.

Why not just dock candidates into the published P-glycoprotein structures? Docking addresses a different question — binding — and even there it fails at the relevant discrimination: the review of computational P-glycoprotein models reports that substrate and non-substrate docking-score distributions overlap too severely to separate the classes (Computational models for predicting substrates or inhibitors of P-glycoprotein). More fundamentally, docking predicts a compound's interaction with the pump, not its interaction with the assay optics; a compound can bind P-glycoprotein and quench the fluorophore, and docking is silent about the second failure. Our method is orthogonal: it interrogates the measurement, not the molecule.

Why not just train a big graph neural network on ChEMBL? Three cited findings argue against it. Large deep models collapse under realistic splits — Pearson correlation dropping from about 0.7 to about 0.4 under temporal shifts in the Therapeutics Data Commons benchmarks (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Therapeutics). The pooled labels themselves are uncalibrated: of roughly 1200 ChEMBL and 352 TP-search P-glycoprotein records, only 40 compounds have comparable readouts across the two databases because assay protocols conflict (Annotating Human P‐Glycoprotein Bioassay Data) — so a network trained on the pool is partly trained on noise. And like docking, it predicts a compound-intrinsic property and cannot see run-specific instrumental drift. This said, we will benchmark a fingerprint-based model retrained on our data (the Random Forest baseline), which is the fair version of this objection.

Why not just use a simple kinetic threshold, like the prompt fraction, and skip the autoencoder? This is the most serious alternative, and it maps onto a real lesson from physics: the amplitude-to-energy ratio is a cheap, well-understood discriminator, yet Holl et al. showed it "ignores much of the information contained in the full temporal evolution of the pulse" and that specific background classes with high current peaks systematically fooled it, classes their autoencoder caught (Deep learning based pulse shape discrimination for germanium detectors). The analogous risk here is an interferent whose optical effect is itself slow (slow quenching chemistry, slow precipitation). The ablation described above is designed to settle this exactly, and if the scalar features match the latent representation on both the simulated and pilot benchmarks, we will say so and reposition the contribution around the benchmark, the proxy-labeling scheme, and the scalar features themselves.

Why not just curate the databases better, as the annotation study suggests? Curation of the kind proposed in Annotating Human P‐Glycoprotein Bioassay Data — merging only assays with identical protocols, cell lines, and substrates, with verapamil as a reference — cleans historical labels but does nothing for a live screen producing false positives today. The two are complementary, and we adopt the verapamil reference standard as part of our pilot design rather than arguing against curation.

Proposed Experiments and Decisive Checks

The first experiment runs on the simulator: train the autoencoder on unlabeled traces, train the tiny classifier on the proxy-labeled controls, and report interferent rejection at 90% true-inhibitor acceptance alongside ROC AUC, with the scalar-threshold and Random Forest baselines run on the same data. The second experiment is the wet-lab novel-compound hold-out, structured to mimic the hardest published setting (novel single-assay compounds, where the Random Forest achieved ROC AUC 0.64): our method must materially exceed that number, with the pre-registered success threshold of 0.85 AUC and 80% interferent rejection at 90% acceptance. The third experiment tests cross-assay transfer: train on Calcein-AM traces, test on rhodamine-123 traces. This is the sharpest conceptual test, because a structure-based predictor has nothing new to say when the same compound moves to a different assay — its fingerprint is unchanged — whereas the artifact's signature in the readout changes with the reporter and instrument, and a readout-level discriminator should adapt. The fourth is the ablation already described, plus one further check in the physics spirit: inject known instrumental pathologies (baseline drift, photobleaching) into real traces and verify the autoencoder's reconstructions degrade gracefully while classifier decisions remain stable, echoing the noise-filtering role the autoencoder plays in the germanium pipeline. If the scalar ablation wins or the cross-assay transfer fails, the proposal's claim narrows honestly to what the evidence supports; if both succeed, the case is that efflux screening should adopt the layered, readout-level defense that rare-event physics built for exactly this shape of problem.