TopU-LBVS: A Realistic Multi Target Benchmark for Ligand Based Virtual Screening

Surbhi Kumar; Yuhe Zhou; Varun Shiralkar; Niu Huang; Baris Coskunuzer

cs.LGSep 24, 2026

TopU-LBVS: A Realistic Multi Target Benchmark for Ligand Based Virtual Screening

Authors: Surbhi Kumar, Yuhe Zhou, Varun Shiralkar, Niu Huang, Baris Coskunuzer

Organizations: UT Dallas, Mathematical Sciences Richardson, TX, USA

Abstract

Ligand-based virtual screening (LBVS) is a practical first-pass tool in early-stage drug discovery, but existing benchmarks can overestimate performance through random negatives, easy decoys, limited target coverage, and non-standardized evaluation protocols. We introduce TopU-LBVS, a multi-target benchmark for LBVS under hard-negative screening conditions. Starting from curated ChEMBL~35 bioactivity data, TopU-LBVS covers 93 protein targets across 7 protein classes and constructs target-specific screening libraries with property-matched, structurally similar decoys at a fixed 1:40 active-to-decoy ratio. Libraries contain roughly 400 to 10,000 compounds and are designed to reduce simple physicochemical and nearest-neighbor fingerprint shortcuts. TopU-LBVS provides three fixed protocols. TopU-LBVS-full evaluates ChEMBL $^\ast \rightarrow$ TopU generalization across all 93 targets. TopU-LBVS-low evaluates low-data TopU $\rightarrow$ TopU learning within the hard-negative distribution. TopU-LBVS-mini provides a compact seven-target protocol with a paired random-decoy control that changes only the test decoys, enabling low-cost development and direct measurement of the gap between random ChEMBL $^\ast$ and TopU decoys. Across ten reference baselines spanning fingerprint methods, molecular GNNs, fingerprint hybrids, and modern molecular models, performance under random-decoy evaluation degrades sharply under hard-negative screening. We release data, fixed splits, evaluation code, and baseline implementations for reproducible comparison of future LBVS and molecular representation learning methods. Code and data are available at https://github.com/topu-benchmark/topu-lbvs and https://huggingface.co/datasets/topu-benchmark/topu-lbvs.

Figures & tables

Benchmark	Targets	Classes	Hard-negative design	Act:Dec ratio	Learning protocol	Fixed splits
DUD-E ( Mysinger et al., 2012 )	102	8	✗	1:50	✗	✗
MUV ( Rohrer and Baumann, 2009 )	17	4	✓	1:500	✗	✗
DEKOIS 2.0 ( Bauer et al., 2013 )	81	5	✗	1:30	✗	✗
LIT-PCBA ( Tran-Nguyen et al., 2020 )	15	3	$\dagger$	varied	✗	✗
FS-Mol ( Stanley et al., 2021 )	varied	varied	✗	$\sim$ balanced	✓	✓
TopU-LBVS (ours)	93	7	✓	1:40	✓	✓

Table 1: Comparison with representative LBVS benchmarks. Hard-negative design denotes explicit construction intended to reduce easy active–inactive separation or artificial enrichment. Learning protocol denotes a canonical training-based setup, and fixed splits denotes released benchmark-level train/validation/test splits. TopU-LBVS additionally provides a 93-target low-data protocol, a seven-target mini benchmark, and a paired random-decoy control.

Figure 1: TopU-LBVS framework. ChEMBL 35 is curated into ChEMBL ∗ , which supplies historical SAR data and the inactive pool for constructing TopU hard-negative libraries. TopU-LBVS-full evaluates generalization from historical ChEMBL ∗ SAR to disjoint TopU screening libraries; TopU-LBVS-low evaluates learning from limited labeled data within the TopU hard-negative distribution; and TopU-LBVS-mini provides a seven-target protocol with a paired random-decoy control that changes only the test decoys.

Parameter	TopU-LBVS-full	TopU-LBVS-low, Tier 1	TopU-LBVS-low, Tier 2	TopU-LBVS-mini
	ChEMBL ${}^{\ast}\rightarrow$ TopU	TopU $\rightarrow$ TopU ( $<50$ actives)	TopU $\rightarrow$ TopU ( $\geq 50$ actives)	ChEMBL ${}^{\ast}\rightarrow$ TopU
Training source	ChEMBL ∗	TopU	TopU	ChEMBL ∗
Test source	TopU	TopU	TopU	TopU
Number of targets	93	46	47	7
Train/val/test split	85% / 15% / TopU test	6:2:2	7:1:2	85% / 15% / paired test
Train ratio (act:dec)	1:10	1:40	1:40	1:10

Table 2: Overview of TopU-LBVS evaluation protocols. TopU-LBVS-full and TopU-LBVS-low contain 93 target-specific tasks. TopU-LBVS-mini follows the full protocol on seven representative targets, one from each protein class. Ratios and splits are fixed for all baselines.

Group	Model	Cytochrome	GPCR	Ion Channel	Kinase	Nuc. Receptor	Misc.	Protease	Overall
Classical	Morgan-RF	14.18	5.21	12.41	9.15	11.12	10.01	6.66	9.18
Classical	Tanimoto-NN	9.84	5.06	9.15	8.84	9.13	7.50	4.40	7.58
	GIN	6.96	3.15	3.87	5.62	5.63	3.92	2.58	4.47
	GIN+FP	14.70	4.88	7.00	7.38	6.29	6.42	4.53	6.64
GNN &	GAT	6.03	2.82	3.68	5.13	3.98	3.42	1.10	3.76
Hybrid	GAT+FP	13.87	4.78	6.41	7.89	7.33	6.05	3.04	6.57

Table 3: TopU-LBVS-full: class-averaged EF@1%. Mean EF@1% under the ChEMBL ∗

\to

TopU protocol, averaged over targets within each protein class. The Overall column is the mean over all 93 targets. Higher is better.

Group	Model	Cytochrome	GPCR	Ion Channel	Kinase	Nuc. Receptor	Misc.	Protease	Overall
Classical	Morgan-RF	1.24	1.02	0.58	0.79	0.57	0.30	0.35	0.63
Classical	Tanimoto-NN	1.13	0.53	0.58	0.67	0.51	0.38	0.05	0.50
	GIN	1.15	1.46	0.75	0.96	0.98	0.90	1.09	1.03
	GIN+FP	0.91	0.85	0.63	1.00	0.83	0.86	0.42	0.84
GNN &	GAT	1.14	0.94	0.88	0.95	1.27	0.96	0.74	0.97
Hybrid	GAT+FP	0.96	0.75	0.56	0.70	0.60	0.89	0.73	0.75

Table 4: TopU-LBVS-low: class-averaged EF@10%. Mean EF@10% under the TopU

\to

TopU few-shot protocol, averaged over targets within each protein class. Tier 1 contains targets with fewer than 50 TopU actives; Tier 2 contains targets with at least 50 TopU actives. Higher is better.

	aa2ar		aces		cp3a4		egfr		esr1		kcnh2		thrb		Average
Model	T	R	T	R	T	R	T	R	T	R	T	R	T	R	T	R
Morgan-RF	6.2	27.2	10.0	28.8	15.9	16.2	6.4	32.1	6.2	24.4	5.4	19.2	10.1	30.6	8.6	25.5
Tanimoto-NN	3.6	18.6	8.0	18.7	10.9	8.5	5.6	17.4	8.0	17.5	4.8	12.0	8.3	22.6	7.0	16.5
GIN	2.8	18.3	2.6	16.2	10.3	9.8	2.8	22.6	2.1	18.1	1.5	10.2	2.2	19.1	3.5	16.3
GIN+FP	6.2	29.9	5.1	30.2	15.6	16.2	3.9	32.8	3.3	25.0	2.4	20.7	3.5	31.5	5.7	26.6
GAT	4.1	13.4	2.8	16.5	9.0	9.4	2.0	16.9	2.4	15.4	2.0	10.2	2.2	18.1	3.5	14.3

Table 5: TopU-LBVS-mini: TopU vs. random decoy test sets (EF@1%). Each target shows TopU hard decoys (T) and randomly sampled ChEMBL ∗ decoys (R). The gap between R and T measures how much random-decoy evaluation overestimates performance. The last two columns are averages across all seven targets. Higher is better.

Appendix figures & tables54 assets

Supplementary material from the paper’s appendix.

Appendix

Class	Target	Full name	Actives	Target	Full name	Actives
	Tier 1 ( $<$ 50 actives, 6:2:2 split)			Tier 2 ( $\geq$ 50 actives, 7:1:2 split)
Cytochrome P450	–	(no Tier 1 target)	–	cp3a4	Cytochrome P450 3A4	154
GPCR	5ht6r	5-hydroxytryptamine receptor 6	47	aa2ar	Adenosine A2a receptor	127
Ion Channel	5ht3a	Serotonin 3a (5-HT3a) receptor	49	kcnh2	hERG	356
Kinase	akt2	RAC-beta serine/threonine kinase	48	egfr	Epidermal growth factor receptor	153
Nuclear Receptor	ppara	Peroxisome proliferator-activated receptor $\alpha$	40	esr1	Estrogen receptor $\alpha$	113

Appendix

Table 6: TopU-LBVS-low-mini target collection. One Tier 1 and one Tier 2 target per protein class. Tier 2 targets match the TopU-LBVS-mini collection. TopU active counts determine tier assignment (

<

50: Tier 1;

\geq

50: Tier 2). Cytochrome P450 contributes one target only (no Tier 1 CYP exists). Splits, metrics, and evaluation code are identical to TopU-LBVS-low.

			ChEMBL (raw)		ChEMBL ∗ (cleaned)		TopU
Target	PDB	Class	active	decoy	active	decoy	active	decoy
cp2c9	1r9o	Cytochrome P450	1893	24225	1882	24052	119	4760
cp1a2	2hi4	Cytochrome P450	1029	22581	1024	22415	136	5440
cp3a4	3nxu	Cytochrome P450	3702	30526	3686	30235	154	6160
cp2d6	4wnt	Cytochrome P450	1891	25216	1884	25028	139	5560
cxcr4	3odu	GPCR	919	855	907	853	15	600

Appendix

Table 7: Per target statistics for the TopU LBVS benchmark (part a). For each target we report counts of ChEMBL compounds before and after cleaning, as well as the number of actives and decoys in the TopU library.

			ChEMBL (raw)		ChEMBL ∗ (cleaned)		TopU
Target	PDB	Class	active	decoy	active	decoy	active	decoy
mcr	2aa2	Nuclear Receptor	944	490	943	489	10	400
rxra	1mv9	Nuclear Receptor	610	1066	609	1063	20	800
thb	1q4x	Nuclear Receptor	574	5734	574	5709	20	800
ppard	2znp	Nuclear Receptor	1698	2055	1689	2038	35	1400
ppara	2p54	Nuclear Receptor	2566	2365	2559	2341	40	1600

Appendix

Table 8: Per target statistics for the TopU LBVS benchmark (part b).

Group	Model	Paradigm	Representation	Parameters
Naive	Morgan-RF ( Rogers and Hahn, 2010 )	Fingerprint + Ensemble	Morgan FP (2048-bit, $r=2$ )	up to 500 trees
Naive	Tanimoto-NN ( Willett, 2006 )	Similarity search	Morgan FP (2048-bit, $r=2$ )	–
	GIN ( Xu et al., 2019 )	Message passing (sum)	Molecular graph	5L, $d=256$
	GIN+FP	Message passing + FP	Graph $\\|$ Morgan FP	5L, $d=256$
GNN &	GAT ( Veličković et al., 2018 )	Attention weighted MP	Molecular graph	5L, $d=256$ , 4 heads
Hybrid	GAT+FP	Attention weighted MP + FP	Graph $\\|$ Morgan FP	5L, $d=256$ , 4 heads

Appendix

Table 9: Summary of all baseline models. “Group” indicates the model family (Similarity based, GNN / Hybrid, Modern pretrained). All neural models use BCE loss with positive class weighting and Adam optimisation. Validation metric is PR-AUC for all models.

Tier	Cutoff	Targets	Split	Train actives	Val actives	Test actives
Tier 1	$<50$	46	6:2:2	6 to 29	2 to 10	2 to 10
Tier 2	$\geq 50$	47	7:1:2	37 to 246	5 to 35	11 to 71

Appendix

Table 10: Tier summary for Setting 2. Active counts are shown as min to max ranges across targets within each tier.

Class	Target	Full Name	TopU actives	ChEMBL ∗ actives
Cytochrome P450	cp3a4	Cytochrome P450 3A4	154	3,686
GPCR	aa2ar	Adenosine A2a receptor	127	5,159
Ion Channel	kcnh2	hERG	352	7,067
Kinase	egfr	Epidermal growth factor receptor	153	7,509
Nuclear Receptor	esr1	Estrogen receptor $\alpha$	113	3,231
Misc. Enzymes	aces	Acetylcholinesterase	118	4,665

Appendix

Table 11: Selected targets for Setting 3, one per protein class.

Target	Class	Train actives	Val actives	Test actives
cp3a4	Cytochrome P450	3,002	530	154
aa2ar	GPCR	4,277	755	127
kcnh2	Ion Channel	5,707	1,008	352
egfr	Kinase	6,252	1,104	153
esr1	Nuclear Receptor	2,646	468	112
aces	Misc. Enzymes	3,789	669	117

Appendix

Table 12: Dataset sizes for Setting 3. All counts are number of active compounds. Decoys follow a 1:10 active-to-decoy ratio in the train and validation splits and a 1:40 ratio in the test split.

Part I — Primary metrics, all models, all targets
Content	Table(s)	Section
TopU-LBVS-full: EF@1%, all 10 models $\times$ 93 targets	20 – 39	H.1
TopU-LBVS-low: EF@10%, all 10 models $\times$ 93 targets	40 – 59	H.1
TopU-LBVS-mini: EF@1%, TopU hard decoys (T), all 10 models $\times$ 7 targets	18	H.1
TopU-LBVS-mini: EF@1%, random decoys (R), all 10 models $\times$ 7 targets	19	H.1
Part II — Full metrics per model, TopU-LBVS-full

Appendix

Table 13: Appendix table index. All per-target result tables and their locations.

Cytochrome P450
Target	Mor-RF	Tan-NN	GAT	GAT+FP	GIN	GIN-FP	GPS	GPS-FP	DMPNN	MolF
cp1a2	12.4	12.2	5.1	12.2	4.9	13.2	3.2	13.2	11.7	7.8
cp2c9	12.6	6.4	1.4	13.4	5.0	14.2	3.1	15.1	11.4	11.2
cp2d6	15.8	9.8	8.6	15.6	7.7	15.8	8.9	15.1	16.1	11.3
cp3a4	15.9	10.9	9.0	14.3	10.2	15.6	5.8	16.4	14.9	14.5
GPCR

Appendix

Table 14: TopU-LBVS-Full EF1% results - part 1. Per-target EF@1% comparison across baseline models.

Nuclear Receptor
Target	Mor-RF	Tan-NN	GAT	GAT+FP	GIN	GIN-FP	GPS	GPS-FP	DMPNN	MolF
andr	12.2	12.7	1.5	4.9	0.5	4.4	2.4	3.9	4.4	4.9
esr1	6.2	8.0	2.4	4.2	2.1	3.3	2.4	3.3	5.9	2.7
esr2	8.0	8.3	2.4	4.8	2.4	3.6	2.4	5.9	4.2	2.4
gcr	10.1	7.1	5.9	8.9	5.3	10.7	3.0	5.9	8.3	8.3
mcr	7.0	6.2	0.0	8.2	0.0	0.0	2.7	2.7	2.7	8.2

Appendix

Table 15: TopU-LBVS-Full EF1% results - part 2. Per-target EF@1% comparison across baseline models.

Cytochrome P450
Target	Tier	Morgan-RF	Tanimoto-NN	GIN	GIN+FP	GAT	GAT+FP	GPS	GPS+FP	DMPNN	MolFormer
cp1a2	2	1.5	1.2	0.9	1.1	0.7	1.1	1.6	0.5	1.6	2.2
cp2c9	2	0.7	0.6	1.5	0.8	1.7	0.7	1.1	0.7	1.1	1.2
cp2d6	2	1.5	1.7	1.0	1.0	1.2	1.1	1.7	0.8	2.3	2.0
cp3a4	2	1.2	1.1	1.3	0.7	1.0	1.0	1.2	0.6	1.6	1.4
GPCR

Appendix

Table 16: TopU-LBVS-few: EF@10% results - part 1. Rows are individual targets grouped by protein class; the Tier column indicates the split strategy (1: fewer than 50 TopU actives, 6:2:2 split; 2: at least 50 TopU actives, 7:1:2 split). Higher is better.

Nuclear Receptor
Target	Tier	Morgan-RF	Tanimoto-NN	GIN	GIN+FP	GAT	GAT+FP	GPS	GPS+FP	DMPNN	MolFormer
andr	2	0.7	0.0	0.8	0.8	0.8	0.8	0.5	0.8	0.5	1.0
esr1	2	0.8	0.1	1.6	1.2	1.0	0.3	0.9	0.0	1.7	1.2
esr2	2	0.7	0.3	1.2	0.6	1.5	0.0	1.2	0.0	0.6	0.3
gcr	2	0.1	0.3	2.7	0.9	0.9	0.9	1.2	0.6	3.3	1.5
mcr	1	0.0	0.0	0.0	0.0	1.5	0.0	0.0	3.0	0.0	1.5

Appendix

Table 17: TopU-LBVS-few: EF@10% results - part 2. Rows are individual targets grouped by protein class; the Tier column indicates the split strategy (1: fewer than 50 TopU actives, 6:2:2 split; 2: at least 50 TopU actives, 7:1:2 split). Higher is better.

EF@1%
Target	Morgan-RF	Tanimoto-NN	GAT	GAT+FP	GIN	GIN+FP	GPS	GPS+FP	DMPNN	MolFormer
aa2ar	6.2	3.6	4.1	5.2	2.8	6.2	1.0	5.7	3.1	2.8
aces	10.0	8.0	2.8	4.8	2.6	5.1	3.4	6.8	6.5	3.4
cp2d6	12.6	9.8	8.6	15.6	7.7	15.8	8.9	15.1	16.1	11.3
egfr	6.4	5.6	2.0	2.0	2.8	3.9	2.6	3.9	7.4	3.5
esr1	6.2	8.0	2.4	4.2	2.1	3.3	2.4	3.3	5.9	2.7

Appendix

Table 18: TopU-LBVS-mini: full metric results on TopU hard-decoy test sets. Each block reports one metric; rows are the 7 mini targets, columns are models. PR-AUC, ROC-AUC, BEDROC, and LogAUC are on a 0–100 scale. Higher is better for all metrics.

Target	Morgan-RF	Tanimoto-NN	GAT	GAT+FP	GIN	GIN+FP	GPS	GPS+FP	DMPNN	MolFormer
EF@1%
aa2ar	27.2	18.6	13.4	29.7	18.3	29.9	9.3	30.2	28.1	25.0
aces	28.8	18.7	16.5	32.2	16.2	30.2	13.1	30.5	30.5	27.3
cp2d6	15.5	10.1	8.2	14.6	7.2	15.1	5.5	15.1	14.1	10.3
egfr	32.1	17.4	16.9	31.7	22.6	32.8	15.8	30.8	28.0	28.4
esr1	24.4	17.5	15.4	23.5	18.1	25.0	10.1	22.3	24.7	19.6

Appendix

Table 19: TopU-LBVS-mini: full metric results on random-decoy test sets. Each block reports one metric; rows are the 7 mini targets, columns are models. PR-AUC, ROC-AUC, BEDROC, and LogAUC are on a 0–100 scale. Higher is better for all metrics.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	12.7 $\pm$ 1.2	8.0 $\pm$ 0.3	5.3 $\pm$ 0.1	17.5 $\pm$ 0.3	81.4 $\pm$ 0.4	36.7 $\pm$ 1.0	40.2 $\pm$ 0.3
cp2c9	12.0 $\pm$ 0.4	7.6 $\pm$ 0.3	5.4 $\pm$ 0.1	16.6 $\pm$ 0.7	84.3 $\pm$ 0.2	36.4 $\pm$ 0.9	41.0 $\pm$ 0.4
cp2d6	16.3 $\pm$ 0.3	9.5 $\pm$ 0.2	6.2 $\pm$ 0.2	23.4 $\pm$ 0.9	84.6 $\pm$ 0.2	44.3 $\pm$ 0.4	45.2 $\pm$ 0.4
cp3a4	15.8 $\pm$ 0.8	9.5 $\pm$ 0.1	6.3 $\pm$ 0.2	24.5 $\pm$ 0.3	85.6 $\pm$ 0.4	45.4 $\pm$ 0.4	46.1 $\pm$ 0.4
GPCR

Appendix

Table 20: TopU-LBVS-Full Morgan-RF results. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	11.7 $\pm$ 0.0	5.6 $\pm$ 0.4	4.4 $\pm$ 0.1	13.7 $\pm$ 1.0	82.7 $\pm$ 0.8	28.0 $\pm$ 1.9	37.4 $\pm$ 1.1
esr1	5.4 $\pm$ 0.0	5.4 $\pm$ 0.2	4.0 $\pm$ 0.1	9.3 $\pm$ 0.3	72.4 $\pm$ 0.3	22.5 $\pm$ 0.6	30.7 $\pm$ 0.3
esr2	7.1 $\pm$ 0.0	5.0 $\pm$ 0.6	4.2 $\pm$ 0.2	13.3 $\pm$ 1.0	75.4 $\pm$ 0.5	23.7 $\pm$ 2.1	32.8 $\pm$ 1.0
gcr	8.3 $\pm$ 0.8	8.2 $\pm$ 0.5	6.1 $\pm$ 0.2	15.0 $\pm$ 0.2	84.8 $\pm$ 0.2	36.1 $\pm$ 0.7	40.1 $\pm$ 0.1
mcr	8.2 $\pm$ 0.0	3.2 $\pm$ 1.8	3.0 $\pm$ 0.8	12.6 $\pm$ 4.2	76.7 $\pm$ 1.5	17.4 $\pm$ 5.6	29.8 $\pm$ 2.4

Appendix

Table 21: TopU-LBVS-Full Morgan-RF results. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	12.2 $\pm$ 1.5	8.5 $\pm$ 0.3	5.1 $\pm$ 0.1	15.6 $\pm$ 1.1	75.4 $\pm$ 1.0	37.0 $\pm$ 1.4	37.8 $\pm$ 1.1
cp2c9	6.4 $\pm$ 1.0	7.7 $\pm$ 0.3	6.1 $\pm$ 0.1	14.1 $\pm$ 0.5	82.8 $\pm$ 0.6	35.4 $\pm$ 1.0	39.0 $\pm$ 0.5
cp2d6	9.8 $\pm$ 0.3	8.9 $\pm$ 0.1	6.4 $\pm$ 0.1	16.2 $\pm$ 0.2	81.9 $\pm$ 0.2	39.3 $\pm$ 0.3	40.5 $\pm$ 0.2
cp3a4	10.9 $\pm$ 0.0	7.8 $\pm$ 0.2	5.8 $\pm$ 0.1	15.0 $\pm$ 0.5	82.3 $\pm$ 0.3	36.4 $\pm$ 0.9	39.7 $\pm$ 0.5
GPCR

Appendix

Table 22: TopU-LBVS-Full Tanimoto-NN results. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	12.7 $\pm$ 0.7	6.8 $\pm$ 0.3	5.0 $\pm$ 0.3	16.3 $\pm$ 0.4	86.0 $\pm$ 0.9	34.3 $\pm$ 1.5	39.9 $\pm$ 0.7
esr1	8.0 $\pm$ 0.0	4.8 $\pm$ 0.2	4.7 $\pm$ 0.1	10.2 $\pm$ 0.1	77.3 $\pm$ 0.5	25.1 $\pm$ 0.2	33.0 $\pm$ 0.3
esr2	8.3 $\pm$ 1.7	4.8 $\pm$ 0.3	4.3 $\pm$ 0.2	10.0 $\pm$ 1.0	76.1 $\pm$ 0.3	23.7 $\pm$ 2.3	31.9 $\pm$ 0.9
gcr	7.1 $\pm$ 0.0	7.8 $\pm$ 0.3	5.8 $\pm$ 0.1	14.2 $\pm$ 0.2	86.6 $\pm$ 0.2	33.9 $\pm$ 0.7	39.2 $\pm$ 0.1
mcr	5.5 $\pm$ 3.9	4.6 $\pm$ 1.8	2.3 $\pm$ 0.9	8.2 $\pm$ 1.8	75.4 $\pm$ 0.8	18.0 $\pm$ 8.1	26.8 $\pm$ 2.1

Appendix

Table 23: TopU-LBVS-Full Tanimoto-NN results. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	5.1 $\pm$ 1.0	4.4 $\pm$ 0.2	3.5 $\pm$ 0.2	7.2 $\pm$ 0.4	72.7 $\pm$ 0.4	18.4 $\pm$ 1.1	28.7 $\pm$ 0.6
cp2c9	1.4 $\pm$ 0.8	2.9 $\pm$ 0.1	2.6 $\pm$ 0.2	5.6 $\pm$ 0.2	75.2 $\pm$ 1.4	10.7 $\pm$ 0.4	26.3 $\pm$ 0.7
cp2d6	8.6 $\pm$ 0.6	6.3 $\pm$ 0.4	4.8 $\pm$ 0.3	13.3 $\pm$ 1.1	81.6 $\pm$ 0.6	30.2 $\pm$ 1.9	37.2 $\pm$ 1.1
cp3a4	9.0 $\pm$ 1.0	4.8 $\pm$ 0.1	4.2 $\pm$ 0.2	11.2 $\pm$ 0.5	78.9 $\pm$ 0.9	23.4 $\pm$ 1.0	33.7 $\pm$ 0.7
GPCR

Appendix

Table 24: TopU-LBVS-Full GAT results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	1.5 $\pm$ 0.0	2.0 $\pm$ 0.4	2.8 $\pm$ 0.3	6.2 $\pm$ 0.1	76.7 $\pm$ 0.2	8.8 $\pm$ 0.8	26.8 $\pm$ 0.2
esr1	2.4 $\pm$ 1.5	2.7 $\pm$ 0.4	2.9 $\pm$ 0.2	5.9 $\pm$ 0.4	73.1 $\pm$ 1.2	11.5 $\pm$ 1.7	26.3 $\pm$ 1.1
esr2	2.4 $\pm$ 0.8	2.3 $\pm$ 0.2	2.4 $\pm$ 0.1	6.1 $\pm$ 0.7	72.1 $\pm$ 0.1	8.0 $\pm$ 0.3	24.5 $\pm$ 0.1
gcr	5.9 $\pm$ 2.2	5.0 $\pm$ 0.5	4.9 $\pm$ 0.1	11.7 $\pm$ 0.8	85.5 $\pm$ 0.5	26.1 $\pm$ 1.8	36.7 $\pm$ 0.9
mcr	0.0 $\pm$ 0.0	2.0 $\pm$ 1.6	2.7 $\pm$ 0.9	5.9 $\pm$ 0.2	72.3 $\pm$ 1.9	7.7 $\pm$ 1.4	23.8 $\pm$ 0.7

Appendix

Table 25: TopU-LBVS-Full GAT results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	12.2 $\pm$ 1.2	7.1 $\pm$ 0.4	4.9 $\pm$ 0.2	16.0 $\pm$ 1.0	77.3 $\pm$ 1.1	33.1 $\pm$ 1.7	37.6 $\pm$ 1.3
cp2c9	13.4 $\pm$ 1.4	7.5 $\pm$ 0.2	5.3 $\pm$ 0.1	17.2 $\pm$ 1.3	82.8 $\pm$ 0.3	36.3 $\pm$ 1.1	40.5 $\pm$ 0.6
cp2d6	15.6 $\pm$ 0.9	8.7 $\pm$ 0.1	5.7 $\pm$ 0.2	21.8 $\pm$ 0.8	81.6 $\pm$ 1.0	41.1 $\pm$ 0.9	42.8 $\pm$ 0.5
cp3a4	14.3 $\pm$ 2.1	8.4 $\pm$ 0.6	5.8 $\pm$ 0.1	21.7 $\pm$ 2.3	81.9 $\pm$ 0.5	40.8 $\pm$ 2.0	42.8 $\pm$ 1.2
GPCR

Appendix

Table 26: TopU-LBVS-Full GAT+FP hybrid model results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	4.9 $\pm$ 1.8	5.3 $\pm$ 0.7	4.2 $\pm$ 0.3	10.5 $\pm$ 0.6	83.1 $\pm$ 0.8	23.8 $\pm$ 2.6	34.6 $\pm$ 1.1
esr1	4.2 $\pm$ 1.1	4.4 $\pm$ 0.8	3.8 $\pm$ 0.1	7.3 $\pm$ 0.5	72.7 $\pm$ 0.3	18.5 $\pm$ 1.9	28.9 $\pm$ 0.8
esr2	4.8 $\pm$ 1.7	4.0 $\pm$ 0.6	3.6 $\pm$ 0.5	8.9 $\pm$ 0.8	75.2 $\pm$ 1.4	18.2 $\pm$ 3.3	29.9 $\pm$ 1.7
gcr	8.9 $\pm$ 0.0	6.3 $\pm$ 0.4	5.5 $\pm$ 0.4	14.8 $\pm$ 0.4	85.4 $\pm$ 0.6	32.1 $\pm$ 1.3	39.3 $\pm$ 0.6
mcr	8.2 $\pm$ 0.0	3.9 $\pm$ 0.0	3.3 $\pm$ 0.5	11.9 $\pm$ 2.0	78.2 $\pm$ 1.5	19.7 $\pm$ 2.9	31.7 $\pm$ 1.2

Appendix

Table 27: TopU-LBVS-Full GAT-FP hybrid model results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	4.9 $\pm$ 0.7	4.7 $\pm$ 0.4	3.7 $\pm$ 0.2	8.0 $\pm$ 0.1	75.8 $\pm$ 1.1	20.2 $\pm$ 0.6	30.5 $\pm$ 0.4
cp2c9	5.0 $\pm$ 0.0	2.9 $\pm$ 0.3	2.4 $\pm$ 0.2	6.3 $\pm$ 0.3	75.5 $\pm$ 0.7	12.3 $\pm$ 1.1	27.3 $\pm$ 0.5
cp2d6	7.7 $\pm$ 1.8	6.5 $\pm$ 0.6	4.9 $\pm$ 0.2	12.5 $\pm$ 1.2	81.9 $\pm$ 0.9	28.4 $\pm$ 2.7	36.4 $\pm$ 1.5
cp3a4	10.2 $\pm$ 0.9	6.2 $\pm$ 1.0	4.7 $\pm$ 0.4	13.5 $\pm$ 1.7	80.6 $\pm$ 0.8	29.3 $\pm$ 3.5	36.8 $\pm$ 1.7
GPCR

Appendix

Table 28: TopU-LBVS-Full GIN results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	0.5 $\pm$ 0.7	2.5 $\pm$ 0.1	2.6 $\pm$ 0.2	6.1 $\pm$ 0.3	79.1 $\pm$ 0.5	9.2 $\pm$ 1.3	27.3 $\pm$ 0.7
esr1	2.1 $\pm$ 0.8	2.3 $\pm$ 0.5	2.8 $\pm$ 0.3	5.7 $\pm$ 0.6	72.9 $\pm$ 1.2	9.8 $\pm$ 1.4	25.6 $\pm$ 0.9
esr2	2.4 $\pm$ 0.8	2.1 $\pm$ 0.0	2.1 $\pm$ 0.1	5.8 $\pm$ 0.6	73.0 $\pm$ 1.0	6.9 $\pm$ 0.7	24.6 $\pm$ 0.9
gcr	5.3 $\pm$ 1.5	5.6 $\pm$ 0.3	5.2 $\pm$ 0.1	11.9 $\pm$ 0.7	85.1 $\pm$ 0.5	27.9 $\pm$ 1.6	37.0 $\pm$ 0.8
mcr	0.0 $\pm$ 0.0	2.0 $\pm$ 1.6	2.3 $\pm$ 0.5	6.2 $\pm$ 0.7	72.5 $\pm$ 0.6	7.8 $\pm$ 3.6	24.1 $\pm$ 1.4

Appendix

Table 29: TopU-LBVS-Full GIN results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	13.2 $\pm$ 0.6	7.4 $\pm$ 0.4	4.9 $\pm$ 0.4	15.7 $\pm$ 0.2	77.6 $\pm$ 1.0	33.8 $\pm$ 0.9	38.0 $\pm$ 0.5
cp2c9	14.2 $\pm$ 1.4	7.6 $\pm$ 0.3	5.3 $\pm$ 0.1	17.1 $\pm$ 0.6	83.1 $\pm$ 0.4	35.8 $\pm$ 0.5	40.6 $\pm$ 0.5
cp2d6	15.8 $\pm$ 0.0	9.2 $\pm$ 0.5	5.8 $\pm$ 0.3	22.2 $\pm$ 1.9	82.4 $\pm$ 0.4	41.8 $\pm$ 1.6	43.5 $\pm$ 0.8
cp3a4	15.6 $\pm$ 0.8	9.0 $\pm$ 0.3	6.2 $\pm$ 0.2	22.8 $\pm$ 1.6	82.6 $\pm$ 0.8	43.2 $\pm$ 1.8	44.0 $\pm$ 0.9
GPCR

Appendix

Table 30: TopU-LBVS-Full GIN+FP hybrid model results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	4.4 $\pm$ 0.0	5.3 $\pm$ 0.6	4.2 $\pm$ 0.1	9.8 $\pm$ 0.7	82.7 $\pm$ 1.4	22.9 $\pm$ 1.8	34.2 $\pm$ 1.2
esr1	3.3 $\pm$ 0.8	3.9 $\pm$ 0.0	3.6 $\pm$ 0.2	7.1 $\pm$ 0.3	72.2 $\pm$ 0.7	17.5 $\pm$ 1.0	28.4 $\pm$ 0.5
esr2	3.6 $\pm$ 1.5	3.7 $\pm$ 0.7	3.1 $\pm$ 0.4	7.6 $\pm$ 0.9	75.1 $\pm$ 0.8	16.5 $\pm$ 2.3	29.3 $\pm$ 1.2
gcr	10.7 $\pm$ 0.0	6.1 $\pm$ 0.0	5.5 $\pm$ 0.3	14.0 $\pm$ 0.5	85.5 $\pm$ 0.1	31.2 $\pm$ 0.8	39.0 $\pm$ 0.4
mcr	0.0 $\pm$ 0.0	2.6 $\pm$ 0.9	2.7 $\pm$ 0.5	7.3 $\pm$ 0.6	76.9 $\pm$ 1.4	11.8 $\pm$ 1.8	27.1 $\pm$ 1.2

Appendix

Table 31: TopU-LBVS-Full GIN+FP hybrid model results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	3.2 $\pm$ 1.4	2.9 $\pm$ 0.9	2.4 $\pm$ 0.6	4.9 $\pm$ 0.9	66.1 $\pm$ 3.2	10.5 $\pm$ 4.3	23.3 $\pm$ 2.4
cp2c9	3.1 $\pm$ 0.4	2.2 $\pm$ 0.1	1.9 $\pm$ 0.1	4.7 $\pm$ 0.8	66.7 $\pm$ 2.6	7.7 $\pm$ 0.9	22.6 $\pm$ 1.4
cp2d6	8.9 $\pm$ 0.3	6.0 $\pm$ 0.1	4.1 $\pm$ 0.1	11.1 $\pm$ 0.3	78.0 $\pm$ 0.6	26.6 $\pm$ 0.3	34.4 $\pm$ 0.4
cp3a4	5.8 $\pm$ 1.9	4.5 $\pm$ 0.1	3.7 $\pm$ 0.2	8.9 $\pm$ 0.7	77.8 $\pm$ 1.0	19.9 $\pm$ 1.4	31.5 $\pm$ 1.1
GPCR

Appendix

Table 32: TopU-LBVS-Full GPS results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	2.4 $\pm$ 2.5	2.2 $\pm$ 0.8	2.5 $\pm$ 0.5	6.0 $\pm$ 0.9	74.8 $\pm$ 1.7	9.5 $\pm$ 4.1	26.3 $\pm$ 2.1
esr1	2.4 $\pm$ 0.8	2.0 $\pm$ 0.1	2.1 $\pm$ 0.1	5.3 $\pm$ 0.2	70.1 $\pm$ 1.3	7.1 $\pm$ 0.8	23.7 $\pm$ 0.7
esr2	2.4 $\pm$ 0.8	1.7 $\pm$ 0.6	1.7 $\pm$ 0.1	4.8 $\pm$ 0.5	69.1 $\pm$ 1.6	5.1 $\pm$ 0.9	22.3 $\pm$ 0.7
gcr	3.0 $\pm$ 0.8	4.8 $\pm$ 0.4	4.5 $\pm$ 0.4	10.6 $\pm$ 0.1	84.7 $\pm$ 1.3	22.5 $\pm$ 2.2	34.8 $\pm$ 1.0
mcr	2.7 $\pm$ 3.9	2.6 $\pm$ 2.4	3.0 $\pm$ 0.0	8.3 $\pm$ 3.2	73.9 $\pm$ 2.4	12.4 $\pm$ 7.0	26.7 $\pm$ 4.0

Appendix

Table 33: TopU-LBVS-Full GPS results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	13.2 $\pm$ 1.8	7.1 $\pm$ 0.6	5.0 $\pm$ 0.3	15.3 $\pm$ 1.2	77.8 $\pm$ 0.4	33.4 $\pm$ 1.4	37.7 $\pm$ 0.6
cp2c9	15.1 $\pm$ 1.2	7.8 $\pm$ 0.5	5.3 $\pm$ 0.1	17.5 $\pm$ 0.8	83.2 $\pm$ 0.1	36.9 $\pm$ 0.9	40.9 $\pm$ 0.4
cp2d6	15.1 $\pm$ 0.6	8.4 $\pm$ 0.6	5.4 $\pm$ 0.4	19.2 $\pm$ 1.6	81.8 $\pm$ 0.9	38.5 $\pm$ 2.5	41.4 $\pm$ 1.4
cp3a4	16.4 $\pm$ 0.3	8.9 $\pm$ 0.3	6.2 $\pm$ 0.3	24.0 $\pm$ 0.4	82.5 $\pm$ 0.6	43.2 $\pm$ 0.3	44.3 $\pm$ 0.3
GPCR

Appendix

Table 34: textbfTopU-LBVS-Full GPS+FP hybrid model results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	3.9 $\pm$ 0.7	6.1 $\pm$ 0.3	4.5 $\pm$ 0.2	10.7 $\pm$ 0.4	82.7 $\pm$ 0.7	25.4 $\pm$ 0.9	35.0 $\pm$ 0.5
esr1	3.3 $\pm$ 1.5	4.1 $\pm$ 0.4	3.7 $\pm$ 0.2	7.0 $\pm$ 0.5	71.0 $\pm$ 0.9	17.7 $\pm$ 1.2	28.2 $\pm$ 0.9
esr2	5.9 $\pm$ 2.2	4.0 $\pm$ 0.2	3.5 $\pm$ 0.7	10.3 $\pm$ 0.9	75.7 $\pm$ 0.7	19.2 $\pm$ 0.8	30.7 $\pm$ 0.2
gcr	5.9 $\pm$ 0.8	7.0 $\pm$ 0.7	5.5 $\pm$ 0.2	14.5 $\pm$ 1.3	85.9 $\pm$ 0.5	32.5 $\pm$ 3.0	39.3 $\pm$ 1.4
mcr	2.7 $\pm$ 3.9	3.9 $\pm$ 0.0	3.7 $\pm$ 0.5	7.8 $\pm$ 0.5	77.7 $\pm$ 0.8	13.6 $\pm$ 1.4	28.0 $\pm$ 0.7

Appendix

Table 35: textbfTopU-LBVS-Full GPS+FP hybrid model results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	11.7 $\pm$ 1.8	7.5 $\pm$ 0.7	5.1 $\pm$ 0.4	16.7 $\pm$ 1.5	80.6 $\pm$ 1.0	34.3 $\pm$ 2.6	39.0 $\pm$ 1.6
cp2c9	11.4 $\pm$ 2.4	7.7 $\pm$ 0.7	5.5 $\pm$ 0.3	16.7 $\pm$ 3.1	83.3 $\pm$ 1.2	35.6 $\pm$ 2.8	40.2 $\pm$ 2.2
cp2d6	16.1 $\pm$ 1.8	8.6 $\pm$ 0.5	5.8 $\pm$ 0.3	20.5 $\pm$ 2.4	84.4 $\pm$ 0.3	41.0 $\pm$ 3.0	43.3 $\pm$ 1.6
cp3a4	14.9 $\pm$ 1.3	9.5 $\pm$ 0.5	6.0 $\pm$ 0.1	22.4 $\pm$ 1.8	85.6 $\pm$ 0.5	43.7 $\pm$ 1.6	44.9 $\pm$ 1.0
GPCR

Appendix

Table 36: TopU-LBVS-Full DMPNN results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	4.4 $\pm$ 1.2	5.2 $\pm$ 0.6	4.1 $\pm$ 0.0	9.6 $\pm$ 0.3	82.2 $\pm$ 0.6	22.8 $\pm$ 0.9	33.6 $\pm$ 0.4
esr1	5.9 $\pm$ 1.5	4.5 $\pm$ 0.6	3.3 $\pm$ 0.2	8.1 $\pm$ 0.8	76.3 $\pm$ 0.2	19.3 $\pm$ 1.7	29.9 $\pm$ 1.3
esr2	4.2 $\pm$ 1.7	3.2 $\pm$ 0.8	2.7 $\pm$ 0.4	7.4 $\pm$ 1.4	75.9 $\pm$ 0.6	13.9 $\pm$ 3.0	28.2 $\pm$ 1.3
gcr	8.3 $\pm$ 2.2	7.2 $\pm$ 0.9	5.8 $\pm$ 0.2	15.3 $\pm$ 2.0	87.6 $\pm$ 0.3	34.1 $\pm$ 3.4	40.5 $\pm$ 1.7
mcr	2.7 $\pm$ 3.9	3.2 $\pm$ 0.9	2.3 $\pm$ 0.5	6.4 $\pm$ 1.7	71.2 $\pm$ 2.7	8.9 $\pm$ 3.7	23.9 $\pm$ 2.5

Appendix

Table 37: TopU-LBVS-Full DMPNN results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Cytochrome P450
cp1a2	7.8 $\pm$ 3.1	6.1 $\pm$ 0.4	4.6 $\pm$ 0.2	11.4 $\pm$ 1.9	79.1 $\pm$ 0.4	26.7 $\pm$ 3.2	34.7 $\pm$ 1.8
cp2c9	11.2 $\pm$ 1.4	7.8 $\pm$ 0.9	5.6 $\pm$ 0.4	16.7 $\pm$ 1.4	84.8 $\pm$ 0.1	36.5 $\pm$ 3.8	41.0 $\pm$ 1.3
cp2d6	11.3 $\pm$ 0.9	7.5 $\pm$ 0.5	5.4 $\pm$ 0.2	16.0 $\pm$ 1.4	84.4 $\pm$ 0.6	34.8 $\pm$ 2.2	40.2 $\pm$ 1.2
cp3a4	14.5 $\pm$ 1.8	9.0 $\pm$ 0.3	6.1 $\pm$ 0.1	21.0 $\pm$ 0.7	84.3 $\pm$ 0.1	42.5 $\pm$ 0.9	43.7 $\pm$ 0.4
GPCR

Appendix

Table 38: TopU-LBVS-Full Molformer results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Target	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
Nuclear Receptor
andr	4.9 $\pm$ 1.8	3.2 $\pm$ 0.6	3.6 $\pm$ 0.6	8.3 $\pm$ 1.1	81.1 $\pm$ 1.4	15.9 $\pm$ 2.8	31.2 $\pm$ 1.7
esr1	2.7 $\pm$ 1.5	3.6 $\pm$ 0.8	3.3 $\pm$ 0.2	6.6 $\pm$ 0.8	74.9 $\pm$ 0.9	14.9 $\pm$ 3.2	28.0 $\pm$ 1.6
esr2	2.4 $\pm$ 1.7	3.0 $\pm$ 0.3	3.0 $\pm$ 0.3	6.3 $\pm$ 0.3	74.8 $\pm$ 0.5	12.3 $\pm$ 1.0	27.2 $\pm$ 0.5
gcr	8.3 $\pm$ 2.2	6.5 $\pm$ 1.0	5.3 $\pm$ 0.2	15.7 $\pm$ 2.1	86.5 $\pm$ 0.3	32.4 $\pm$ 2.8	40.4 $\pm$ 1.7
mcr	8.2 $\pm$ 0.0	3.9 $\pm$ 1.6	3.3 $\pm$ 0.9	12.1 $\pm$ 2.1	73.5 $\pm$ 1.5	18.3 $\pm$ 4.3	29.0 $\pm$ 2.0

Appendix

Table 39: TopU-LBVS-Full Molformer results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC (

\alpha

= 20), and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	2.3 $\pm$ 1.6	1.6 $\pm$ 0.6	1.5 $\pm$ 0.3	3.7 $\pm$ 0.2	62.2 $\pm$ 3.6	3.6 $\pm$ 1.5	25.7 $\pm$ 1.7
cp2c9	2	0.0 $\pm$ 0.0	0.3 $\pm$ 0.2	0.7 $\pm$ 0.4	2.5 $\pm$ 0.3	48.6 $\pm$ 5.1	-4.3 $\pm$ 1.4	20.2 $\pm$ 4.5
cp2d6	2	5.2 $\pm$ 2.6	1.9 $\pm$ 0.9	1.5 $\pm$ 0.3	3.7 $\pm$ 1.0	53.9 $\pm$ 2.5	5.6 $\pm$ 4.4	23.8 $\pm$ 3.5
cp3a4	2	1.1 $\pm$ 1.5	0.9 $\pm$ 0.4	1.2 $\pm$ 0.2	3.6 $\pm$ 1.6	49.8 $\pm$ 3.0	-0.4 $\pm$ 2.5	19.9 $\pm$ 3.0
GPCR

Appendix

Table 40: TopU-LBVS-few Morgan-RF results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.7 $\pm$ 0.6	2.5 $\pm$ 0.4	46.0 $\pm$ 9.0	-3.8 $\pm$ 2.0	17.8 $\pm$ 5.3
esr1	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.2	0.8 $\pm$ 0.2	2.6 $\pm$ 0.1	51.1 $\pm$ 2.2	-2.7 $\pm$ 0.9	21.6 $\pm$ 2.0
esr2	2	0.0 $\pm$ 0.0	0.3 $\pm$ 0.3	0.7 $\pm$ 0.7	2.6 $\pm$ 0.5	48.1 $\pm$ 7.7	-3.9 $\pm$ 2.1	23.1 $\pm$ 7.4
gcr	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.1 $\pm$ 0.1	2.4 $\pm$ 0.1	45.4 $\pm$ 2.6	-5.7 $\pm$ 0.8	21.0 $\pm$ 0.8
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	2.4 $\pm$ 0.0	39.6 $\pm$ 2.6	-6.7 $\pm$ 0.0	28.9 $\pm$ 5.0

Appendix

Table 41: TopU-LBVS-few Morgan-RF results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	1.1 $\pm$ 1.6	1.2 $\pm$ 0.3	1.2 $\pm$ 0.5	3.0 $\pm$ 0.2	50.4 $\pm$ 2.2	2.1 $\pm$ 0.4	19.3 $\pm$ 1.0
cp2c9	2	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	0.6 $\pm$ 0.2	2.6 $\pm$ 0.1	51.9 $\pm$ 1.9	-3.5 $\pm$ 0.2	17.7 $\pm$ 0.5
cp2d6	2	4.6 $\pm$ 3.2	3.3 $\pm$ 0.7	1.7 $\pm$ 0.3	4.6 $\pm$ 0.5	54.1 $\pm$ 3.1	8.6 $\pm$ 2.5	23.6 $\pm$ 2.0
cp3a4	2	1.1 $\pm$ 1.5	0.9 $\pm$ 0.3	1.1 $\pm$ 0.2	2.5 $\pm$ 0.3	46.6 $\pm$ 4.1	-0.6 $\pm$ 1.3	16.3 $\pm$ 2.2
GPCR

Appendix

Table 42: TopU-LBVS-few Tanimoto-NN results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	2.2 $\pm$ 0.2	44.3 $\pm$ 4.1	-6.2 $\pm$ 0.2	15.1 $\pm$ 1.8
esr1	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.1 $\pm$ 0.2	2.4 $\pm$ 0.2	50.4 $\pm$ 3.8	-5.0 $\pm$ 0.8	16.5 $\pm$ 2.3
esr2	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	2.4 $\pm$ 0.2	47.2 $\pm$ 4.2	-5.7 $\pm$ 1.1	16.6 $\pm$ 2.7
gcr	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	3.3 $\pm$ 0.3	60.0 $\pm$ 3.3	-4.5 $\pm$ 1.1	24.4 $\pm$ 1.5
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	3.3 $\pm$ 1.1	36.0 $\pm$ 21.2	-6.6 $\pm$ 0.1	20.8 $\pm$ 13.2

Appendix

Table 43: TopU-LBVS-few Tanimoto-NN results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	0.5 $\pm$ 0.7	1.0 $\pm$ 0.9	0.7 $\pm$ 0.5	2.5 $\pm$ 0.2	48.3 $\pm$ 1.4	-2.1 $\pm$ 2.8	16.4 $\pm$ 1.7
cp2c9	2	2.7 $\pm$ 1.9	1.4 $\pm$ 0.4	1.7 $\pm$ 0.6	4.1 $\pm$ 0.9	60.1 $\pm$ 2.7	3.5 $\pm$ 2.9	24.3 $\pm$ 3.0
cp2d6	2	0.0 $\pm$ 0.0	1.2 $\pm$ 0.3	1.2 $\pm$ 0.7	3.0 $\pm$ 0.4	52.5 $\pm$ 2.2	1.4 $\pm$ 3.1	19.9 $\pm$ 1.8
cp3a4	2	0.0 $\pm$ 0.0	0.9 $\pm$ 0.8	1.0 $\pm$ 0.7	3.0 $\pm$ 0.5	54.6 $\pm$ 2.7	-0.4 $\pm$ 4.5	19.4 $\pm$ 2.6
GPCR

Appendix

Table 44: TopU-LBVS-few GAT results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	0.5 $\pm$ 0.7	0.8 $\pm$ 0.6	2.7 $\pm$ 0.2	49.8 $\pm$ 0.8	-3.3 $\pm$ 3.3	18.8 $\pm$ 1.7
esr1	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.4	1.0 $\pm$ 0.7	3.0 $\pm$ 0.3	57.4 $\pm$ 2.3	-2.0 $\pm$ 3.1	20.4 $\pm$ 2.3
esr2	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.8	1.5 $\pm$ 0.8	3.6 $\pm$ 1.8	49.5 $\pm$ 12.7	1.0 $\pm$ 5.8	23.4 $\pm$ 9.6
gcr	2	2.7 $\pm$ 3.9	1.2 $\pm$ 0.8	0.9 $\pm$ 0.7	4.0 $\pm$ 2.5	44.9 $\pm$ 8.3	0.2 $\pm$ 6.4	19.4 $\pm$ 6.3
mcr	1	0.0 $\pm$ 0.0	2.7 $\pm$ 3.9	1.5 $\pm$ 2.1	6.4 $\pm$ 4.1	38.8 $\pm$ 25.6	2.1 $\pm$ 9.8	32.2 $\pm$ 23.2

Appendix

Table 45: TopU-LBVS-few GAT results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	1.1 $\pm$ 1.6	1.5 $\pm$ 0.6	1.1 $\pm$ 0.3	2.6 $\pm$ 0.4	46.7 $\pm$ 3.1	0.0 $\pm$ 2.4	17.0 $\pm$ 2.4
cp2c9	2	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	0.7 $\pm$ 0.2	2.3 $\pm$ 0.1	45.7 $\pm$ 2.2	-3.1 $\pm$ 1.3	16.0 $\pm$ 0.9
cp2d6	2	1.1 $\pm$ 1.6	1.2 $\pm$ 0.3	1.1 $\pm$ 0.3	2.6 $\pm$ 0.4	42.1 $\pm$ 1.7	0.2 $\pm$ 1.9	15.6 $\pm$ 1.4
cp3a4	2	2.1 $\pm$ 1.5	1.1 $\pm$ 0.8	1.0 $\pm$ 0.0	3.0 $\pm$ 0.3	47.2 $\pm$ 6.6	0.2 $\pm$ 2.2	17.4 $\pm$ 1.7
GPCR

Appendix

Table 46: TopU-LBVS-few GAT results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	1.5 $\pm$ 1.2	0.8 $\pm$ 0.6	2.8 $\pm$ 0.2	45.8 $\pm$ 4.9	-0.6 $\pm$ 3.4	19.2 $\pm$ 1.7
esr1	2	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	0.3 $\pm$ 0.4	2.2 $\pm$ 0.4	41.8 $\pm$ 5.5	-3.8 $\pm$ 2.8	14.2 $\pm$ 3.2
esr2	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	2.0 $\pm$ 0.1	35.3 $\pm$ 3.7	-5.6 $\pm$ 0.2	12.7 $\pm$ 1.7
gcr	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.8	0.9 $\pm$ 0.0	2.9 $\pm$ 0.3	50.9 $\pm$ 5.1	-2.3 $\pm$ 1.3	21.2 $\pm$ 2.1
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	2.3 $\pm$ 0.4	21.0 $\pm$ 13.5	-6.7 $\pm$ 0.0	9.3 $\pm$ 6.5

Appendix

Table 47: TopU-LBVS-few GAT results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	0.0 $\pm$ 0.0	1.2 $\pm$ 0.7	0.9 $\pm$ 0.2	2.9 $\pm$ 0.4	53.9 $\pm$ 6.4	-0.6 $\pm$ 2.3	19.4 $\pm$ 2.5
cp2c9	2	1.4 $\pm$ 1.9	1.9 $\pm$ 1.0	1.5 $\pm$ 0.5	3.6 $\pm$ 0.4	55.9 $\pm$ 2.7	3.3 $\pm$ 3.4	23.1 $\pm$ 1.7
cp2d6	2	0.0 $\pm$ 0.0	0.5 $\pm$ 0.3	1.0 $\pm$ 0.4	3.0 $\pm$ 0.2	55.2 $\pm$ 1.4	-0.9 $\pm$ 0.4	20.5 $\pm$ 0.9
cp3a4	2	0.0 $\pm$ 0.0	1.1 $\pm$ 0.3	1.3 $\pm$ 0.3	3.2 $\pm$ 0.6	56.7 $\pm$ 7.7	2.0 $\pm$ 1.6	21.4 $\pm$ 3.6
GPCR

Appendix

Table 48: TopU-LBVS-few GIN results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	0.5 $\pm$ 0.7	0.8 $\pm$ 0.6	2.8 $\pm$ 0.6	51.3 $\pm$ 10.3	-3.4 $\pm$ 1.9	20.2 $\pm$ 4.9
esr1	2	1.4 $\pm$ 1.9	1.7 $\pm$ 0.7	1.6 $\pm$ 0.4	3.6 $\pm$ 1.2	55.4 $\pm$ 8.9	2.7 $\pm$ 4.3	23.3 $\pm$ 5.7
esr2	2	0.0 $\pm$ 0.0	1.2 $\pm$ 0.8	1.2 $\pm$ 0.8	3.1 $\pm$ 0.2	48.8 $\pm$ 1.9	0.1 $\pm$ 3.0	22.5 $\pm$ 1.4
gcr	2	0.0 $\pm$ 0.0	3.0 $\pm$ 0.8	2.7 $\pm$ 0.0	4.9 $\pm$ 0.4	61.7 $\pm$ 5.1	7.6 $\pm$ 0.6	31.5 $\pm$ 2.6
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	3.4 $\pm$ 1.6	31.5 $\pm$ 20.4	-5.7 $\pm$ 1.5	18.7 $\pm$ 15.9

Appendix

Table 49: TopU-LBVS-few GIN results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	2.3 $\pm$ 1.6	1.5 $\pm$ 0.6	1.1 $\pm$ 0.3	3.0 $\pm$ 0.3	52.6 $\pm$ 7.2	0.9 $\pm$ 1.9	19.8 $\pm$ 2.4
cp2c9	2	2.7 $\pm$ 1.9	0.5 $\pm$ 0.4	0.8 $\pm$ 0.6	3.3 $\pm$ 0.9	48.8 $\pm$ 6.7	-0.1 $\pm$ 3.4	19.4 $\pm$ 4.4
cp2d6	2	0.0 $\pm$ 0.0	0.7 $\pm$ 0.6	1.0 $\pm$ 0.3	2.3 $\pm$ 0.1	45.6 $\pm$ 0.6	-1.9 $\pm$ 0.9	15.6 $\pm$ 0.3
cp3a4	2	2.1 $\pm$ 1.5	1.1 $\pm$ 0.3	0.7 $\pm$ 0.4	3.0 $\pm$ 0.4	45.6 $\pm$ 2.2	0.7 $\pm$ 2.3	17.0 $\pm$ 2.0
GPCR

Appendix

Table 50: TopU-LBVS-few GIN-FP results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	2.3 $\pm$ 3.2	0.5 $\pm$ 0.7	0.8 $\pm$ 0.6	2.9 $\pm$ 1.0	42.7 $\pm$ 6.9	-0.6 $\pm$ 4.5	17.9 $\pm$ 4.7
esr1	2	1.4 $\pm$ 1.9	1.1 $\pm$ 0.4	1.2 $\pm$ 0.5	2.8 $\pm$ 0.3	48.0 $\pm$ 4.1	-0.1 $\pm$ 2.1	18.2 $\pm$ 2.2
esr2	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.6 $\pm$ 0.8	2.5 $\pm$ 0.9	43.2 $\pm$ 13.2	-4.0 $\pm$ 3.4	16.8 $\pm$ 7.9
gcr	2	0.0 $\pm$ 0.0	1.2 $\pm$ 0.8	0.9 $\pm$ 0.0	3.1 $\pm$ 0.6	50.2 $\pm$ 2.8	-0.6 $\pm$ 3.3	21.9 $\pm$ 3.2
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	3.2 $\pm$ 1.4	34.2 $\pm$ 24.1	-6.7 $\pm$ 0.1	18.2 $\pm$ 15.6

Appendix

Table 51: TopU-LBVS-few GIN-FP results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	2.3 $\pm$ 1.6	2.0 $\pm$ 0.9	1.6 $\pm$ 0.6	4.3 $\pm$ 0.6	62.6 $\pm$ 1.7	5.5 $\pm$ 4.0	25.5 $\pm$ 2.2
cp2c9	2	1.4 $\pm$ 1.9	1.1 $\pm$ 1.0	1.1 $\pm$ 0.9	4.4 $\pm$ 2.7	51.3 $\pm$ 8.3	0.7 $\pm$ 6.3	19.7 $\pm$ 6.7
cp2d6	2	2.3 $\pm$ 1.6	2.1 $\pm$ 0.6	1.7 $\pm$ 0.6	3.8 $\pm$ 0.6	52.6 $\pm$ 7.6	5.7 $\pm$ 2.5	22.3 $\pm$ 3.9
cp3a4	2	0.0 $\pm$ 0.0	1.5 $\pm$ 0.3	1.2 $\pm$ 0.5	3.2 $\pm$ 0.4	56.1 $\pm$ 2.3	1.6 $\pm$ 2.4	21.2 $\pm$ 2.1
GPCR

Appendix

Table 52: TopU-LBVS-few GPS results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	0.5 $\pm$ 0.7	0.5 $\pm$ 0.4	2.5 $\pm$ 0.3	48.9 $\pm$ 2.8	-4.3 $\pm$ 1.7	17.9 $\pm$ 2.3
esr1	2	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	0.9 $\pm$ 0.7	2.9 $\pm$ 0.2	55.7 $\pm$ 1.3	-1.8 $\pm$ 2.8	20.0 $\pm$ 1.4
esr2	2	2.7 $\pm$ 3.9	1.2 $\pm$ 0.8	1.2 $\pm$ 0.8	3.8 $\pm$ 1.3	47.6 $\pm$ 4.3	1.4 $\pm$ 5.4	22.1 $\pm$ 2.5
gcr	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.8	1.2 $\pm$ 1.1	3.9 $\pm$ 1.4	57.2 $\pm$ 10.9	0.5 $\pm$ 4.9	26.2 $\pm$ 8.5
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	3.8 $\pm$ 1.3	42.5 $\pm$ 24.8	-6.3 $\pm$ 0.4	25.5 $\pm$ 16.0

Appendix

Table 53: TopU-LBVS-few GPS results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	1.1 $\pm$ 1.6	0.5 $\pm$ 0.7	0.5 $\pm$ 0.5	2.6 $\pm$ 0.7	46.2 $\pm$ 3.0	-2.5 $\pm$ 3.8	15.5 $\pm$ 2.8
cp2c9	2	0.0 $\pm$ 0.0	0.8 $\pm$ 0.7	0.7 $\pm$ 0.4	2.9 $\pm$ 0.7	53.8 $\pm$ 9.3	-1.5 $\pm$ 3.5	19.6 $\pm$ 4.9
cp2d6	2	2.3 $\pm$ 1.6	0.9 $\pm$ 0.3	0.8 $\pm$ 0.2	2.4 $\pm$ 0.4	39.4 $\pm$ 2.2	-0.4 $\pm$ 2.5	14.5 $\pm$ 2.1
cp3a4	2	0.0 $\pm$ 0.0	0.2 $\pm$ 0.3	0.6 $\pm$ 0.3	2.0 $\pm$ 0.1	40.0 $\pm$ 3.5	-3.7 $\pm$ 1.4	12.8 $\pm$ 0.8
GPCR

Appendix

Table 54: TopU-LBVS-few GPS-FP results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	0.5 $\pm$ 0.7	0.8 $\pm$ 0.6	2.5 $\pm$ 0.7	39.5 $\pm$ 10.0	-2.2 $\pm$ 4.6	15.4 $\pm$ 5.7
esr1	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	2.1 $\pm$ 0.5	39.5 $\pm$ 9.8	-5.5 $\pm$ 0.9	12.9 $\pm$ 4.1
esr2	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	2.0 $\pm$ 0.1	36.2 $\pm$ 5.1	-6.1 $\pm$ 0.3	12.3 $\pm$ 1.7
gcr	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.8	0.6 $\pm$ 0.4	2.5 $\pm$ 0.4	45.0 $\pm$ 8.9	-3.5 $\pm$ 2.2	17.7 $\pm$ 3.5
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	3.0 $\pm$ 2.1	7.0 $\pm$ 3.8	45.8 $\pm$ 30.6	2.8 $\pm$ 7.3	38.6 $\pm$ 27.2

Appendix

Table 55: TopU-LBVS-few GPS-FP results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	1.1 $\pm$ 1.6	1.5 $\pm$ 1.0	1.6 $\pm$ 0.3	3.8 $\pm$ 0.9	59.1 $\pm$ 3.4	3.6 $\pm$ 4.3	23.5 $\pm$ 2.9
cp2c9	2	0.0 $\pm$ 0.0	1.6 $\pm$ 0.7	1.1 $\pm$ 0.5	3.7 $\pm$ 0.5	62.3 $\pm$ 2.6	1.6 $\pm$ 3.1	24.3 $\pm$ 2.5
cp2d6	2	3.4 $\pm$ 0.0	2.4 $\pm$ 0.3	2.3 $\pm$ 0.7	5.1 $\pm$ 1.3	66.0 $\pm$ 6.5	8.8 $\pm$ 4.0	28.9 $\pm$ 4.8
cp3a4	2	2.1 $\pm$ 1.5	1.1 $\pm$ 0.3	1.6 $\pm$ 0.7	4.2 $\pm$ 0.7	63.9 $\pm$ 5.4	3.7 $\pm$ 2.9	25.4 $\pm$ 3.4
GPCR

Appendix

Table 56: TopU-LBVS-few DMPNN results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	1.0 $\pm$ 0.7	0.5 $\pm$ 0.4	3.2 $\pm$ 0.6	53.8 $\pm$ 5.6	-1.1 $\pm$ 3.3	22.2 $\pm$ 4.8
esr1	2	1.4 $\pm$ 1.9	1.7 $\pm$ 0.7	1.7 $\pm$ 0.7	4.2 $\pm$ 0.8	61.8 $\pm$ 4.0	4.0 $\pm$ 4.0	25.7 $\pm$ 3.7
esr2	2	0.0 $\pm$ 0.0	0.6 $\pm$ 0.8	0.6 $\pm$ 0.8	3.3 $\pm$ 0.8	53.4 $\pm$ 1.7	-1.5 $\pm$ 6.0	22.1 $\pm$ 3.1
gcr	2	0.0 $\pm$ 0.0	3.0 $\pm$ 2.2	3.3 $\pm$ 0.4	6.4 $\pm$ 1.7	70.0 $\pm$ 6.6	10.5 $\pm$ 5.8	37.4 $\pm$ 6.2
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	3.4 $\pm$ 0.6	43.3 $\pm$ 9.0	-6.6 $\pm$ 0.2	23.8 $\pm$ 6.8

Appendix

Table 57: TopU-LBVS-few DMPNN results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Cytochrome P450
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
cp1a2	2	2.3 $\pm$ 1.6	2.0 $\pm$ 0.9	2.2 $\pm$ 0.3	4.5 $\pm$ 0.7	58.3 $\pm$ 3.9	6.3 $\pm$ 2.5	24.4 $\pm$ 1.3
cp2c9	2	0.0 $\pm$ 0.0	1.4 $\pm$ 0.4	1.2 $\pm$ 0.3	3.7 $\pm$ 0.4	63.5 $\pm$ 4.4	1.2 $\pm$ 0.9	24.7 $\pm$ 2.3
cp2d6	2	3.4 $\pm$ 0.0	2.1 $\pm$ 0.0	2.0 $\pm$ 0.3	4.1 $\pm$ 0.7	60.1 $\pm$ 6.7	5.8 $\pm$ 1.5	25.1 $\pm$ 3.2
cp3a4	2	0.0 $\pm$ 0.0	1.1 $\pm$ 0.8	1.4 $\pm$ 0.4	3.7 $\pm$ 0.3	63.7 $\pm$ 1.2	1.4 $\pm$ 3.7	23.9 $\pm$ 1.9
GPCR

Appendix

Table 58: TopU-LBVS-few Molformer results - part 1. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Nuclear Receptor
Target	Tier	EF@1%	EF@5%	EF@10%	PR-AUC	ROC-AUC	BEDROC	LogAUC
andr	2	0.0 $\pm$ 0.0	1.5 $\pm$ 1.2	1.0 $\pm$ 0.9	3.1 $\pm$ 0.8	45.8 $\pm$ 3.8	0.8 $\pm$ 5.6	20.1 $\pm$ 4.7
esr1	2	0.0 $\pm$ 0.0	1.4 $\pm$ 0.4	1.2 $\pm$ 0.4	2.8 $\pm$ 0.4	50.7 $\pm$ 4.0	0.6 $\pm$ 2.4	19.3 $\pm$ 2.7
esr2	2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.3 $\pm$ 0.4	2.5 $\pm$ 0.4	46.3 $\pm$ 7.0	-4.9 $\pm$ 1.8	17.6 $\pm$ 4.1
gcr	2	2.7 $\pm$ 3.9	1.2 $\pm$ 1.7	1.5 $\pm$ 0.8	8.0 $\pm$ 5.5	68.8 $\pm$ 5.4	5.0 $\pm$ 9.1	33.9 $\pm$ 6.5
mcr	1	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	1.5 $\pm$ 2.1	6.4 $\pm$ 1.5	61.5 $\pm$ 8.2	-1.5 $\pm$ 3.4	44.8 $\pm$ 8.7

Appendix

Table 59: TopU-LBVS-few Molformer results - part 2. All values are reported as mean

\pm

std over three random seeds. EF = Enrichment Factor. PR-AUC, ROC-AUC, BEDROC, and LogAUC are reported on a 0–100 scale.

Explore similar work

Sep 21, 2026q-bio.BM

Adapting Boltz-2 with limited experimental activity data improves early enrichment in virtual screening

Virtual screening aims to prioritize active compounds from large chemical libraries within a limited experimental budget. When applying Boltz-2 to virtual screening, a key challenge is how to use limited experimental data from the target assay to improve the prioritization of active compounds. We investigated whether fine-tuning the Boltz-2 affinity heads with a small number of binary activity labels could improve early enrichment of active compounds in hit discovery. We compared fine-tuning with 40-300 labels in a retrospective evaluation on eight MF-PCBA targets. With 300 activity measurements, fine-tuning increased the number of actives in the top 1% by a geometric mean of 1.77-fold across the eight targets and improved average precision (AP) by 2.14-fold relative to the control without fine-tuning. We also investigated whether rescoring a subset of candidates could retain the improvement in hit recovery by reranking only the top-ranked Boltz-2 candidates with the fine-tuned head. Restricting rescoring to approximately 10% of the evaluation set retained hit recovery comparable to full rescoring. These findings show that affinity-head fine-tuning with limited activity labels improves early enrichment with Boltz-2 and that this benefit can be retained when rescoring a restricted set of candidates.

Kairi Furui, Masahito Ohue

Department of Computer Science, School of Computing, Institute of Science Tokyo, Yokohama 226-8501, Japan

Screening Drug Discovery Boltzmann Distributions Early Layers

Ask PDF

Oct 28, 2025cs.LG

APEX: Approximate-but-exhaustive search for ultra-large combinatorial synthesis libraries

Make-on-demand combinatorial synthesis libraries (CSLs) like Enamine REAL have significantly enabled drug discovery efforts. However, their large size presents a challenge for virtual screening, where the goal is to identify the top compounds in a library according to a computational objective (e.g., optimizing docking score) subject to computational constraints under a limited computational budget. For current library sizes -- numbering in the tens of billions of compounds -- and scoring functions of interest, a routine virtual screening campaign may be limited to scoring fewer than 0.1% of the available compounds, leaving potentially many high scoring compounds undiscovered. Furthermore, as constraints (and sometimes objectives) change during the course of a virtual screening campaign, existing virtual screening algorithms typically offer little room for amortization. We propose the approximate-but-exhaustive search protocol for CSLs, or APEX. APEX utilizes a neural network surrogate that exploits the structure of CSLs in the prediction of objectives and constraints to make full enumeration on a consumer GPU possible in under a minute, allowing for exact retrieval of approximate top-k sets. To demonstrate APEX's capabilities, we develop a benchmark CSL comprised of more than 10 million compounds, all of which have been annotated with their docking scores on five medically relevant targets along with physicohemical properties measured with RDKit such that, for any objective and set of constraints, the ground truth top-k compounds can be identified and compared against the retrievals from any virtual screening algorithm. We show APEX's consistently strong performance both in retrieval accuracy and runtime compared to alternative methods.

Aryan Pedawi, Jordi Silvestre-Ryan, Bradley Worley +5

Work performed at Numerion Labs · NVIDIA

Drug Discovery Screening Search Algorithms Combinations

Ask PDF

May 3, 2026cs.LG

Benchmarking Single-Pose Docking, Consensus Rescoring, and Supervised ML on the LIT-PCBA Library: A Critical Evaluation of DiffDock, AutoDock-GPU, GNINA, and DiffDock-NMDN

Virtual screening performance depends heavily on the chosen docking and scoring methods. Recent AI-based tools such as DiffDock and NMDN have reported strong benchmark results, but their practical utility on realistic, experimentally-derived datasets remains unclear. Here we perform a large-scale evaluation on the LIT-PCBA library (15 targets, 578,295 ligand-target pairs with experimentally confirmed actives and inactives). We compare AutoDock-GPU and DiffDock for pose generation, followed by rescoring with GNINA and NMDN. We further evaluate rank-based consensus strategies and supervised machine learning models trained on docking features. GNINA rescoring of AutoDock-GPU poses (AutoDock-GNINA) emerged as the strongest single method with a median EF1% of 2.14. DiffDock-based approaches underperformed relative to AutoDock-GNINA, particularly on challenging targets such as OPRK1. Carefully designed consensus ranking improved robustness but did not surpass the best single scorer. Supervised ML re-ranking delivered the largest gains, achieving a median EF1% of 4.49 (+110% over AutoDock-GNINA). Our results highlight that even the best classical+ML hybrid workflows provide only modest early enrichment on realistic benchmarks. We conclude that no single docking method dominates across targets and that rigorously validated, cost-effective combinations with supervised re-ranking currently offer the most practical value for virtual screening.

Youssef Abo-Dahab, Xiaoiang Xiang, Joanne Chun +1

Department of Bioengineering and Therapeutic Sciences, Schools of Pharmacy and Medicine, University of California, San Francisco, San Francisco, CA, USA · UCSF–Stanford Center of Excellence in Regulatory Science and Innovation (CERSI), San Francisco, CA, USA

Screening

Ask PDF