Predicting Quantifiability from Primary Screens to Prioritize Dose-Response Profiling

Sean Lim

cs.LGAug 27, 2026

Predicting Quantifiability from Primary Screens to Prioritize Dose-Response Profiling

Authors: Sean Lim

Organizations: Department of Natural Science, Rice University, Houston, TX, USA

Abstract

High-throughput drug screening relies on low-cost primary assays to prioritize compounds for more expensive dose-response profiling, where potency is ultimately quantified. Current screening strategies largely focus on identifying compounds that will confirm biological activity on follow-up, implicitly assuming that confirmed activity will also yield a usable potency estimate. However, confirmed biological activity in screening does not necessarily translate into a quantifiable potency, because active compounds can still fail to produce a reportable dose-response estimate. We therefore present a framework for modeling quantifiability, whether follow-up testing will yield a usable potency estimate, as a distinct triage objective from biological activity. Quantifiability was strongly predictable from the preceding low-cost screen, with most predictive information arising from the observed screening features rather than molecular structure. Response-based predictors remained robust on previously unseen chemical scaffolds and generalized across held-out assay-mechanism families, while the probability of successful quantification varied strongly with response amplitude and assay context. These findings establish experimental measurability, distinct from biological activity, as a predictable property of screening outcomes and show that quantifiability-aware triage can improve the allocation of costly dose-response profiling capacity.

Figures & tables

Policy	AUROC	AUPRC	Recall@50%	Yield@25%
Random	$0.502\pm 0.003$	$0.449\pm 0.005$	$0.500\pm 0.003$	$0.450\pm 0.009$
Random forest (fingerprints only)	$0.688\pm 0.010$	$0.628\pm 0.010$	$0.645\pm 0.008$	$0.666\pm 0.011$
Maximum three-point activity	$0.848\pm 0.003$	$0.781\pm 0.005$	$0.807\pm 0.005$	$0.850\pm 0.007$
Random forest (three-point only)	$0.850\pm 0.005$	$0.825\pm 0.005$	$0.796\pm 0.006$	$0.858\pm 0.004$
Mean three-point activity	$0.863\pm 0.002$	$0.829\pm 0.003$	$0.814\pm 0.002$	$0.888\pm 0.003$
Logistic regression (full features)	$0.880\pm 0.006$	$0.849\pm 0.005$	$0.833\pm 0.006$	$0.900\pm 0.006$

Table 1: Five-fold random cross-validation performance for predicting whether a promoted compound–assay pair yields a reportable potency. Values are mean

\pm

standard deviation. Recall@50% is the fraction of quantifiable pairs recovered when the top-ranked 50% of pairs are profiled. Yield@25% is the fraction of profiled pairs that are quantifiable when the top-ranked 25% are selected. Fingerprint-only and three-point-only random forests use a single feature block. Logistic regression and the full random forest use the complete screening-stage feature set.

Feature set	Contents	$N$	AUROC	AUPRC	Rec@50%	$\Delta$ AUROC
Isolated blocks (five-fold)
Three-point only	3pt block	16	$0.886\pm 0.002$	$0.867\pm 0.002$	$0.835\pm 0.006$	—
Fingerprints only	Morgan bits	1,024	$0.761\pm 0.004$	$0.690\pm 0.006$	$0.711\pm 0.003$	—
Full model	all blocks	1,074	$0.939\pm 0.002$	$0.926\pm 0.003$	$0.894\pm 0.003$	—
Nested addition (three-fold)
Three-point only	3pt block	16	$0.886\pm 0.003$	—	—	—

Table 2: Feature-group ablation for a random forest trained on the quantified label (

n_{\mathrm{estimators}}=300

, class-balanced). Isolated blocks were evaluated by five-fold stratified cross-validation. Nested models added groups cumulatively and were evaluated by three-fold stratified cross-validation. The three-point block contains 15 response summaries plus the reported screening maximum (16 features). Compound and assay profiles comprise six compound-level and three assay-level historical screening statistics (9 features). The final nested step restores assay metadata, mechanism family, target class, technology, screening category, promotion reason, and a mutant flag, to recover the full 1,074-feature matrix. Share of nested gain is the fraction of the AUROC increase from the three-point block (

0.886

) to the full model (

0.938

). Isolated fingerprint-only performance in this table uses the same 300-tree forest as the nested models and is therefore not identical to the fingerprint-only row in Table 1 .

Figure 1: Predictive information is concentrated in the screening response. Top: summed Gini importance by feature group for a random forest trained on the full screening-stage matrix. The 16 three-point features account for the largest share, followed by Morgan fingerprints; historical profiles and assay metadata contribute smaller amounts. Bottom: AUROC under random versus Bemis–Murcko scaffold-disjoint splitting for fingerprint-only, three-point-only, and full models. Bars are five-fold means; error bars are one standard deviation. The three-point block is nearly unchanged on unseen scaffolds, whereas fingerprint-only performance declines substantially.

Figure 2: Schematic 11-point responses illustrating the amplitude distinction among outcomes. Quantifiable profiles reach a high plateau from which a Hill midpoint can be reported. Active but unquantifiable profiles show a real concentration-dependent response that remains too incomplete or too weak for a reportable fit. Inactive profiles remain near baseline. The dashed line marks 50% of assay maximum. Curves are schematic and are not fitted EvE Bio traces.

Mechanism family	$N$ active	Quantifiable	Unquantifiable	Failure (%)
G-protein activation	6,824	5,542	1,282	18.8
Competition binding	4,199	3,038	1,161	27.6
Barr2 recruitment	8,375	5,394	2,981	35.6
Co-factor recruitment	890	489	401	45.1
Heterodimer	1,134	209	925	81.6
All active pairs	21,422	14,672	6,750	31.5

Table 3: Hill-fitting failure among compound–assay pairs that were active in the 11-point experiment, by assay-mechanism family. Failure is the fraction of active pairs that did not yield a reportable potency. Totals match the 21,422 pairs confirmed active during profiling, of which 6,750 (31.5%) were unquantifiable.

Target class
Subgroup	$N$	Pos. (%)	AUROC	Rec@50%
Kinase (competition binding)	1,547	37.8	0.975	0.974
Nuclear receptor	704	19.9	0.944	0.964
7TM	4,344	50.9	0.920	0.828
Screening category
Compound of interest	2,710	73.5	0.923	0.665

Table 4: Subgroup ranking and residual errors on a held-out fold of 6,595 promoted pairs. Target-class and screening-category metrics are from a gradient-boosted tree trained on the remaining pairs.

N

is the number of held-out pairs in the subgroup; Pos. is the fraction that were quantifiable; Rec@50% is recall when the top-ranked half of that subgroup is profiled. Missed-pair counts and screening maxima are from a random forest evaluated at a 50% budget over the full held-out fold (2,619 of 2,935 quantifiable pairs recovered). Kinase and competition-binding rows coincide because the kinase pairs in this dataset are the competition-binding assays.

Release	Test $N$	Quantifiable	Pos. (%)	AUROC	Rec@50% (model)	Rec@50% (screening)
4	4,731	1,966	41.6	0.814	0.785	0.751
5.1	4,462	1,851	41.5	0.867	0.832	0.774
6	2,739	1,297	47.4	0.918	0.865	0.850
7	2,722	1,079	39.6	0.944	0.943	0.902
8	2,475	805	32.5	0.977	0.985	0.974
9	3,571	2,077	58.2	0.885	0.748	0.703

Table 5: Leakage-safe rolling evaluation. For each test release, the model was trained only on promoted pairs from earlier releases, and historical compound- and assay-level features were recomputed from that prior history. AUROC is from a random forest trained on the quantified label. Rec@50% is the fraction of quantifiable pairs recovered when the top-ranked 50% of that release is profiled, compared with ranking by screening activity. Release 5.1 is the first test window after release 4 (there is no release 5.0). The eight windows contain 30,120 pairs in total.

Figure 3: Leakage-safe rolling evaluation across eight incoming releases. Top: fraction of quantifiable pairs recovered when the top-ranked 50% of each release is profiled, compared with ranking by screening activity. The model recovers more potencies on every release. Bottom: fraction of each release that must be profiled to retain 90% of the potencies obtained by exhaustive profiling. The dashed line is the aggregate model budget of 55.4%. The model required a smaller budget than screening activity on seven of eight releases and tied it on release 8.

Figure 4: Precision–recall performance tracks the prevalence of quantifiable pairs across chronological test releases. AUPRC remains high while the positive rate is moderate, then falls to 0.385 on release 11, where only 11.0% of promoted pairs were quantifiable. Fixed-budget recall is therefore interpreted within each release relative to the corresponding screening policy rather than compared directly across releases.

Dose-Response Estimation

See all 8 papers

ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening
Aug 12, 2026Antoine de Mathelin, Christopher Tosh, Wesley TanseyFew-Shot Learning Drug Discovery
Computational Oncology Memorial Sloan Kettering Cancer Center New York, NY 10065
Hit Selection Using SSMD-Based Machine Learning Performance Metrics in High-Throughput Screening Assays
Aug 6, 2026Xiaohua Douglas ZhangDrug Discovery
Department of Biostatistics, College of Public Health, University of Kentucky, Lexington, KY40536, USA
Empowering Chemical Structures with Biological Insights for Scalable Phenotypic Virtual Screening
Mar 16, 2026Xiaoqing Lian, Pengsen Ma, Tengfeng Ma +9Virtual Screening Drug Discovery
State Key Laboratory of Chemo and Biosensing, College of Computer Science and Electronic Engineering; The Ministry of Education Key Laboratory of Fusion Computing of Supercomputing and Artificial Intelligence, Hunan University, 410082, Changsha, China · School of Artificial Intelligence and Computer Science, Jiangnan University, 214122, Wuxi, China · Department of Computer Science, University of Tsukuba, Tsukuba, Ibaraki, Japan +2

Aug 12, 2026J/K move · Enter open · S save

ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening

Antoine de Mathelin, Christopher Tosh, Wesley Tansey

Computational Oncology Memorial Sloan Kettering Cancer Center New York, NY 10065

Few-Shot Learning Drug Discovery

Open paper AskPDF arXiv

Treating patients with combinations of drugs reduces the risk of resistance to any individual drug. Finding effective combinations is difficult because the large search space makes combinatorial screens prohibitively expensive, time consuming, and often technically infeasible. Predictive models can fill this gap, yet existing methods typically require molecular profiling of each sample and per-cohort training, limiting their applicability when time and tissue are scarce. To address this challenge, we introduce ScreenShot, a hierarchical transformer pretrained on 40 drug screening datasets covering 3,700 drugs and 6,000 biological samples, whose architecture mirrors the nested structure of screening data. Given a few-shot context of observations from a new patient, ScreenShot predicts the response of the sample to combination therapies through in-context learning, operating directly on functional measurements with no fine-tuning and no molecular profiling. On four held-out datasets, ScreenShot outperforms all baselines in both prediction accuracy and identification of selectively effective treatments. ScreenShot's internal representations are directly useful for experimental design: we use them to drive a weighted k-means++ active learning strategy that selects which experiments to run, achieving the same hit detection as uniform screening with a third of the budget. Source code and interactive dashboard: https://github.com/tansey-lab/screenshot.

Abstract

Dose-Response Estimation

Explore similar work

ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening

Hit Selection Using SSMD-Based Machine Learning Performance Metrics in High-Throughput Screening Assays

Empowering Chemical Structures with Biological Insights for Scalable Phenotypic Virtual Screening