On the Selectivity of Generative Models in Structure-Based Drug Design

Ella Miray Rajaonson; Jungyoon Lee; William Chau; Alan Aspuru-Guzik; Benjamin Sanchez-Lengeling; Dominique Beaini; Kirill Neklyudov; Marta Skreta

cs.LGOct 16, 2026

On the Selectivity of Generative Models in Structure-Based Drug Design

Authors: Ella Miray Rajaonson, Jungyoon Lee, William Chau, Alan Aspuru-Guzik, Benjamin Sanchez-Lengeling, Dominique Beaini, Kirill Neklyudov, Marta Skreta

Organizations: University of Toronto · Vector Institute · Université de Montréal · Mila – Québec AI Institute ·

Abstract

Designing small-molecule drugs that selectively bind to a target protein while avoiding unintended interactions with other proteins is essential for reducing side effects that contribute substantially to clinical attrition. Although structure-based drug design (SBDD) has advanced rapidly with deep learning, most generative models still optimize binding to a single target and disregard selectivity due to limited training data and unknown off-target interactions. Motivated by this gap, we introduce SELECTBENCH, a comprehensive benchmark assessing the selectivity of generative SBDD models. SELECTBENCH provides three curated evaluation tasks reflecting different real-world drug discovery scenarios: literature-based case studies, safety screening panels used in pharmaceutical development, and large-scale structure-based evaluation sets. We further introduce a suite of metrics to quantify selectivity and conduct a systematic evaluation of nine representative SBDD models, including methods explicitly trained for selectivity and widely-used inference-time guidance techniques. Our results show that even selectivity-trained models achieve only modest off-target discrimination and inference-time guidance provides limited gains, underscoring the need for method improvement. SELECTBENCH provides both a characterization of where current generative SBDD stands on selectivity and an open, extensible platform to accelerate progress. The benchmark is available at https://github.com/lovetaro/selectbench.

Figures & tables

Figure 1: Off-target interactions are a major consideration in real-world drug design (top left) and often not accounted for in current generative SBDD (top right). SelectBench provides tasks, metrics, and baselines to assess selectivity in generative SBDD (bottom).

Figure 2: SelectBench task summary. SelectBench -Case focuses on case studies, SelectBench -Safety models safety screening, and SelectBench -XL enables large-scale evaluation.

On-Target Gene	On-Target Application	Off-Target Gene	Off-Target Effect	Computational study
PARP1	Cancer [ 52 ]	PARP2	Hematological toxicity [ 52 ]	CMD-GEN [ 74 ]
BMP2K	Antiviral [ 28 ]	TTK	Hematological toxicity, neutropenia [ 6 ]	DiffSBDD [ 63 ]
CDK9	Cancer [ 51 ]	CDK2	Hematological toxicity, GI issues [ 34 ]	Alchemical free-energy [ 3 ]
MAPK1	Cancer [ 60 ]	CDK2	Hematological toxicity, GI issues [ 34 ]	Alchemical free-energy [ 3 ]
DRD2	Neuropsychiatric [ 72 ]	HRH1	Hypertension, neutropenia [ 23 ]	Docking [ 68 ]

Table 1: Summary of case studies in SelectBench -Case.

Panel of Table 2 — Table 2: Summary of baseline models evaluated in SelectBench .

Figure 3: Baseline performances on SelectBench -Case The mean performance is reported across

\bar{\Delta}^{\mathrm{z}}_{\text{top-}10}

(

\uparrow

) and

\mathrm{SR}^{\mathrm{z}}_{\tau=1}

(

\uparrow

) for each baseline on each on-/off-target pair in SelectBench -Case. The color of the marker indicates the model type, and its shape indicates if inference-time methods were additionally employed. Data points in the top-right region indicate better performance.

Figure 4: Baseline performances on SelectBench -Safety-Priority. The mean performance is reported across

\mathrm{R}_{20}^{\mathrm{z}}\,(\uparrow)

and

\mathrm{SR}_{\tau=1}^{\text{z, avg}}\,(\uparrow)

for each baseline on each on-target POI/priority off-targets case in SelectBench -Safety-Priority. The color of the marker indicates the model type, and its shape indicates if inference-time methods were additionally employed. Data points in the top-right region indicate better performance.

Figure 5: Baseline performances on SelectBench -XL The mean performance is reported across

\bar{\Delta}^{\mathrm{z}}_{\text{top-}10}

(

\uparrow

) and

\mathrm{SR}^{\mathrm{z}}_{\tau=1}

(

\uparrow

) for each baseline on each on-/off-target pairs in SelectBench -XL by difficulty. The color of the marker indicates the model type, and its shape indicates if inference-time methods were additionally employed. Data points in the top-right region indicate better performance.

Figure 6: Multi-objective metrics exploration on one case study of SelectBench -Case The mean performance is reported across

\Delta^{\mathrm{z}}

(

\uparrow

-s_{\mathrm{on}}

(

\uparrow

) and

\mathrm{SA}

(

\uparrow

) for each baseline on the PARP1/PARP2 case study The color of the marker indicates the model type, and its shape indicates if inference-time methods were additionally employed. Data points in the top-right region indicate better performance. The potency/selectivity trend in can be captured by introducing the potency-and-selectivity metric

\mathrm{PS}=-s_{\mathrm{on}}*\Delta^{\mathrm{z}}

(

\uparrow

). We report

\mathrm{PS}

across all other evaluation tasks in SelectBench in App. G . Additional metrics assessing the quality of generated molecules beyond selectivity and potency for each task are reported in App. H .

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

ID	On-target	Off-target	ID	On-target	Off-target	ID	On-target	Off-target
717	4tqr_A	14gs_A	3498	1r1h_A	2rma_A	2308	1coy_A	2pqw_A
823	1h0i_A	14gs_A	4773	5mgl_A	4keu_A	2508	2jjg_A	1djy_A
981	4rlu_A	2v3r_A	4886	4q8b_B	3ej8_A	2515	2rma_A	2jjg_A
2167	4zfa_A	3g51_A	5895	1k9t_A	4p6p_A	2978	2pqw_A	3pnm_A
2219	4kcq_A	3g51_A	6354	4lfu_A	3l3n_A	3311	2pc8_A	5bur_A
8550	1ai4_A	3hy9_B	8989	4gvd_A	5mgl_A	9741	3af2_A	3u9f_C

Appendix

Table A1: Easy pairs sampled from SelectBench -XL

$\bar{s}_{\text{on}}$	$\bar{\Delta}$	$\mathrm{SR}_{\tau=0}$	$\mathrm{SR}_{\tau=1}$	$\bar{s}_{\text{on}}^{z}$	$\bar{\Delta}^{z}$	$\mathrm{SR}^{z}_{\tau=0}$	$\mathrm{SR}^{z}_{\tau=1}$
$-11.008\pm 0.969$	$-0.033\pm 0.502$	$0.556$	$0.000$	$-1.631\pm 1.009$	$0.207\pm 0.540$	$0.667$	$0.111$

Appendix

Table A2: PARP1/PARP2 positive control evaluation We report: the mean on-target Vina score

\bar{s}_{\text{on}}

(

\downarrow

);

\bar{\Delta}

over selected molecules and the selectivity ratio

\mathrm{SR}_{\tau}

for

\tau\in\{0,1\}

kcal/mol. Additionally, we report the z-score normalized analogues of the corresponding raw metrics, standardized against a reference score distribution:

\bar{s}_{\text{on}}^{z}

\bar{\Delta}^{z}

\mathrm{SR}^{z}_{\tau=0}

and

\mathrm{SR}^{z}_{\tau=1}

Ligands	$\bar{s}_{\text{on}}$	$\bar{\Delta}$	$\mathrm{SR}_{\tau=0}$	$\mathrm{SR}_{\tau=1}$	$\bar{s}_{\text{on}}^{z}$	$\bar{\Delta}^{z}$	$\mathrm{SR}^{z}_{\tau=0}$	$\mathrm{SR}^{z}_{\tau=1}$
CDK9-selective	$-8.050\pm 0.121$	$-0.691\pm 0.365$	$0.000$	$0.000$	$-0.117\pm 0.166$	$0.172\pm 0.439$	$0.750$	$0.000$
CDK9 non-selective	$-7.822\pm 0.741$	$-1.074\pm 0.594$	$0.000$	$0.000$	$0.195\pm 1.014$	$-0.323\pm 0.648$	$0.250$	$0.000$

Appendix

Table A3: CDK9/CDK2 positive control evaluation We report: the mean on-target Vina score

\bar{s}_{\text{on}}

(

\downarrow

);

\bar{\Delta}

over selected molecules and the selectivity ratio

\mathrm{SR}_{\tau}

for

\tau\in\{0,1\}

kcal/mol. Additionally, we report the z-score normalized analogues of the corresponding raw metrics, standardized against a reference score distribution:

\bar{s}_{\text{on}}^{z}

\bar{\Delta}^{z}

\mathrm{SR}^{z}_{\tau=0}

and

\mathrm{SR}^{z}_{\tau=1}

Method	SelectBench -Case	SelectBench -Safety	SelectBench -XL
DecompDiff	2.4%	3.5%	3.1%
DiffSBDD	1.6%	1.4%	5.2%
TargetDiff	0.6%	0.7%	2.3%
MolSnapper	2.5%	5.4%	6.7%
MolForm	0.6%	2.2%	9.2%
Pocket2Mol	0.6%	0.4%	0.4%

Appendix

Table A4: Pre-docking failure rates by method and evaluation track. Failures occur when a wrapper-validated molecule is rejected by Meeko/RDKit ligand preparation before AutoDock Vina runs. Aggregated as

1-(N_{\text{docked}}/N_{\text{generated}})

over all (method, target) cells in each track. SelectBench -Safety totals are computed over the full 85-pair panel.

Method	$\bar{s}_{\text{on}}$ $(\downarrow)$	$\bar{\Delta}$ $(\uparrow)$	$\bar{\Delta}_{\text{top-}5}$ $(\uparrow)$	$\bar{\Delta}_{\text{top-}10}$ $(\uparrow)$	$\mathrm{SR}_{\tau=0}$ $(\uparrow)$	$\mathrm{SR}_{\tau=1}$ $(\uparrow)$	$\bar{s}_{\text{on}}^{z}$ $(\downarrow)$	$\bar{\Delta}^{z}$ $(\uparrow)$	$\mathrm{SR}^{z}_{\tau=0}$ $(\uparrow)$	$\mathrm{SR}^{z}_{\tau=1}$ $(\uparrow)$
\hlcustom Blue1 Case: PARP1 (on) / PARP2 (off)
Standard SBDD models
DecompDiff	$-7.43_{\scriptscriptstyle\pm 0.57}$	$-0.03_{\scriptscriptstyle\pm 0.42}$	$1.05$	$0.84$	$0.41$	$0.02$	$2.15_{\scriptscriptstyle\pm 0.60}$	$0.02_{\scriptscriptstyle\pm 0.41}$	$0.46$	$0.04$
DiffSBDD	$-8.44_{\scriptscriptstyle\pm 1.67}$	$-0.05_{\scriptscriptstyle\pm 0.43}$	$0.90$	$0.77$	$0.46$	$0.01$	$1.09_{\scriptscriptstyle\pm 1.76}$	$0.08_{\scriptscriptstyle\pm 0.43}$	$0.57$	$0.03$
TargetDiff	$-8.56_{\scriptscriptstyle\pm 1.22}$	$0.07_{\scriptscriptstyle\pm 0.54}$	$1.52$	$1.16$	$0.55$	$0.05$	$0.96_{\scriptscriptstyle\pm 1.28}$	$0.20_{\scriptscriptstyle\pm 0.53}$	$0.67$	$0.07$
MolForm	$-8.30_{\scriptscriptstyle\pm 1.23}$	$-0.27_{\scriptscriptstyle\pm 0.43}$	$0.59$	$0.45$	$0.29$	$0.00$	$1.24_{\scriptscriptstyle\pm 1.29}$	$-0.15_{\scriptscriptstyle\pm 0.44}$	$0.37$	$0.00$

Appendix

Table A5: SelectBench -Case evaluation. For each method, we report: the mean on-target Vina score

\bar{s}_{\text{on}}

(

\downarrow

);

\bar{\Delta}

over valid molecules; the mean of the

k

largest

\Delta^{(i)}

(

k=5,10

)

\bar{\Delta}_{\text{top-}k}

; the selectivity ratio

\mathrm{SR}_{\tau}

for

\tau\in\{0,1\}

kcal/mol. Additionally, we report the z-score normalized analogues of the corresponding raw metrics, standardized against a reference score distribution:

\bar{s}_{\text{on}}^{z}

\bar{\Delta}^{z}

\mathrm{SR}^{z}_{\tau=0}

\mathrm{SR}^{z}_{\tau=1}

. The

\tau=1

cutoff in z-space corresponds to

\Delta^{z}\geq 1

(one reference-

\sigma

in the off-target’s own scale).

Method	$\bar{s}_{\text{on}}$ $(\uparrow)$	$\bar{\Delta}$ $(\uparrow)$	$\bar{\Delta}_{\text{top-}5}$ $(\uparrow)$	$\bar{\Delta}_{\text{top-}10}$ $(\uparrow)$	$\mathrm{SR}^{\mathrm{rank}}_{\tau=0}$ $(\uparrow)$	$\mathrm{SR}_{s_{\text{on}}>0.6,s_{\text{off}}<0.4}$ $(\uparrow)$	$\mathrm{SR}^{\mathrm{rank}}_{p_{\text{on}}>0.6,p_{\text{off}}<0.4}$ $(\uparrow)$
\hlcustom Blue1 Case: PARP1 (on) / PARP2 (off)
Standard SBDD models
DecompDiff	$0.396_{\pm 0.126}$	$0.085_{\pm 0.083}$	$0.257$	$0.241$	$0.490$	$0.010$	$0.080$
DiffSBDD	$0.421_{\pm 0.214}$	$0.050_{\pm 0.094}$	$0.300$	$0.253$	$0.444$	$0.010$	$0.020$
TargetDiff	$0.372_{\pm 0.158}$	$0.069_{\pm 0.088}$	$0.284$	$0.241$	$0.500$	$0.010$	$0.030$
MolForm	$0.323_{\pm 0.162}$	$0.038_{\pm 0.070}$	$0.174$	$0.163$	$0.500$	$0.000$	$0.010$

Appendix

Table A6: SelectBench -Case Boltz-2 evaluation. For each method, we report: the mean on-target predicted probability score

\bar{s}_{\text{on}}

(

\uparrow

);

\bar{\Delta}

over valid molecules; the mean of the

k

largest

\Delta^{(i)}

(

k=5,10

)

\bar{\Delta}_{\text{top-}k}

; the selectivity ratio

\mathrm{SR}^{\mathrm{rank}}_{\tau=0}

over percentile-normalized ranks as well as the threshold-based SR over predicted binding probabilities

\mathrm{SR}_{\tau_{\mathrm{on}}=0.6,\;\tau_{\mathrm{off}}=0.4}

and percentile-normalized ranks

\mathrm{SR}^{\mathrm{rank}}_{\tau_{\mathrm{on}}=0.6,\;\tau_{\mathrm{off}}=0.4}

Method	$\bar{s}_{\text{on}}$ $(\downarrow)$	$\bar{\Delta}_{\text{worst-mol}}$ $(\uparrow)$	$\mathrm{SR}_{\tau=0}^{\text{worst-mol}}$ $(\uparrow)$	$\mathrm{SR}_{\tau=1}^{\text{worst-mol}}$ $(\uparrow)$	$\mathrm{R}_{20}$ $(\uparrow)$	$\bar{s}_{\text{on}}^{z}$ $(\downarrow)$	$\bar{\Delta}^{z}$ $(\uparrow)$	$\mathrm{SR}_{\tau=0}^{z,\text{avg}}$ $(\uparrow)$	$\mathrm{SR}_{\tau=1}^{z,\text{avg}}$ $(\uparrow)$	$\mathrm{R}_{20}^{\mathrm{z}}$ $(\uparrow)$
\hlcustom Blue1 POI: 5HT1B (priority)
Standard SBDD models
DecompDiff	$-6.73_{\scriptscriptstyle\pm 1.69}$	$-1.16_{\scriptscriptstyle\pm 1.71}$	$0.24$	$0.02$	$0.64$	$2.45_{\scriptscriptstyle\pm 1.75}$	$0.74_{\scriptscriptstyle\pm 1.85}$	$0.68$	$0.44$	$0.43$
DiffSBDD	$-6.83_{\scriptscriptstyle\pm 1.76}$	$-1.32_{\scriptscriptstyle\pm 1.05}$	$0.07$	$0.00$	$0.38$	$2.34_{\scriptscriptstyle\pm 1.82}$	$-0.08_{\scriptscriptstyle\pm 1.18}$	$0.51$	$0.21$	$0.21$
TargetDiff	$-7.21_{\scriptscriptstyle\pm 1.53}$	$-1.20_{\scriptscriptstyle\pm 1.42}$	$0.19$	$0.04$	$0.50$	$1.95_{\scriptscriptstyle\pm 1.59}$	$1.04_{\scriptscriptstyle\pm 1.77}$	$0.65$	$0.41$	$0.35$
MolForm	$-7.00_{\scriptscriptstyle\pm 2.64}$	$-2.00_{\scriptscriptstyle\pm 2.27}$	$0.09$	$0.02$	$0.45$	$2.17_{\scriptscriptstyle\pm 2.73}$	$0.72_{\scriptscriptstyle\pm 2.01}$	$0.63$	$0.40$	$0.34$

Appendix

Table A7: SelectBench -Safety-Priority evaluation For each method, we report: the mean on-target Vina score

\bar{s}_{\text{on}}

(

\downarrow

); the worst per-molecule

\bar{\Delta}_{\text{worst-mol}}

averaged across

T=22

off-targets (least Selectivity-trained models); the fraction of molecules

\mathrm{SR}_{\tau=0}^{\text{worst-mol}}

\mathrm{SR}_{\tau=1}^{\text{worst-mol}}

whose worst per-target

\Delta

\geq 0

\geq 1

, respectively; the fraction

\mathrm{R}_{20}

of ligands whose on-target ranks in the top-20% across all 23 targets. Additionally, we report the z-score normalized analogues of the corresponding raw metrics, standardized against a reference score distribution:

\bar{s}_{\text{on}}^{z}

\bar{\Delta}^{z}

\mathrm{SR}_{\tau=0}^{z,\text{avg}}

\mathrm{SR}_{\tau=1}^{z,\text{avg}}

\mathrm{R}_{20}^{z}

\mathrm{SR}_{\tau}^{z,\text{avg}}

is the panel-average of per-target z-SRs.

\tau=1

in z-space corresponds to one reference-

\sigma

margin in the off-target’s own scale

Figure A1: Baseline performances on SelectBench -Safety-All The mean performance is reported across

\mathrm{R}_{20}^{\mathrm{z}}\,(\uparrow)

and

\mathrm{SR}_{\tau=1}^{\text{avg}}\,(\uparrow)

for each baseline on each on-target POI/all off-targets case in SelectBench -Safety-All. The color of the marker indicates the model type, and its shape indicates if inference-time methods were additionally employed. Data points in the top-right region indicate better performance.

Method	$\bar{s}_{\text{on}}$ $(\downarrow)$	$\bar{\Delta}_{\text{worst-mol}}$ $(\uparrow)$	$\mathrm{SR}_{\tau=0}^{\text{worst-mol}}$ $(\uparrow)$	$\mathrm{SR}_{\tau=1}^{\text{worst-mol}}$ $(\uparrow)$	$\mathrm{R}_{20}$ $(\uparrow)$	$\bar{s}_{\text{on}}^{z}$ $(\downarrow)$	$\bar{\Delta}^{z}$ $(\uparrow)$	$\mathrm{SR}_{\tau=0}^{z,\text{avg}}$ $(\uparrow)$	$\mathrm{SR}_{\tau=1}^{z,\text{avg}}$ $(\uparrow)$	$\mathrm{R}_{20}^{\mathrm{z}}$ $(\uparrow)$
\hlcustom Blue1 POI: 5HT1B
Standard SBDD models
DecompDiff	$-6.73_{\scriptscriptstyle\pm 1.69}$	$-1.60_{\scriptscriptstyle\pm 1.68}$	$0.11$	$0.01$	$0.70$	$2.45_{\scriptscriptstyle\pm 1.75}$	$0.61_{\scriptscriptstyle\pm 1.79}$	$0.66$	$0.41$	$0.44$
DiffSBDD	$-6.83_{\scriptscriptstyle\pm 1.76}$	$-1.57_{\scriptscriptstyle\pm 0.99}$	$0.00$	$0.00$	$0.44$	$2.34_{\scriptscriptstyle\pm 1.82}$	$-0.13_{\scriptscriptstyle\pm 1.12}$	$0.50$	$0.19$	$0.19$
TargetDiff	$-7.21_{\scriptscriptstyle\pm 1.53}$	$-1.54_{\scriptscriptstyle\pm 1.32}$	$0.06$	$0.01$	$0.62$	$1.95_{\scriptscriptstyle\pm 1.59}$	$0.79_{\scriptscriptstyle\pm 1.56}$	$0.64$	$0.37$	$0.35$
MolForm	$-7.00_{\scriptscriptstyle\pm 2.64}$	$-2.35_{\scriptscriptstyle\pm 2.20}$	$0.02$	$0.00$	$0.55$	$2.17_{\scriptscriptstyle\pm 2.73}$	$0.56_{\scriptscriptstyle\pm 1.92}$	$0.61$	$0.38$	$0.33$

Appendix

Table A8: SelectBench -Safety-All evaluation. For each method, we report: the mean on-target Vina score

\bar{s}_{\text{on}}

(

\downarrow

); the worst per-molecule

\bar{\Delta}_{\text{worst-mol}}

averaged across

T=22

off-targets (least Selectivity-trained models); the fraction of molecules

\mathrm{SR}_{\tau=0}^{\text{worst-mol}}

\mathrm{SR}_{\tau=1}^{\text{worst-mol}}

whose worst per-target

\Delta

\geq 0

\geq 1

, respectively; the fraction

\mathrm{R}_{20}

\bar{s}_{\text{on}}^{z}

\bar{\Delta}^{z}

\mathrm{SR}_{\tau=0}^{z,\text{avg}}

\mathrm{SR}_{\tau=1}^{z,\text{avg}}

\mathrm{R}_{20}^{z}

\mathrm{SR}_{\tau}^{z,\text{avg}}

is the panel-average of per-target z-SRs.

\tau=1

in z-space corresponds to one reference-

\sigma

margin in the off-target’s own scale

Method	$\bar{s}_{\text{on}}$ $(\downarrow)$	$\bar{\Delta}$ $(\uparrow)$	$\bar{\Delta}_{\text{top-}5}$ $(\uparrow)$	$\bar{\Delta}_{\text{top-}10}$ $(\uparrow)$	$\mathrm{SR}_{\tau=0}$ $(\uparrow)$	$\mathrm{SR}_{\tau=1}$ $(\uparrow)$	$\bar{s}_{\text{on}}^{z}$ $(\downarrow)$	$\bar{\Delta}^{z}$ $(\uparrow)$	$\mathrm{SR}^{z}_{\tau=0}$ $(\uparrow)$	$\mathrm{SR}^{z}_{\tau=1}$ $(\uparrow)$
\hlcustom Blue1 Hard
Standard SBDD models
DecompDiff	$-5.48_{\scriptscriptstyle\pm 0.66}$	$0.10_{\scriptscriptstyle\pm 0.40}$	$0.85$	$0.74$	$0.51$	$0.09$	$3.39_{\scriptscriptstyle\pm 0.76}$	$0.07_{\scriptscriptstyle\pm 0.45}$	$0.53$	$0.12$
DiffSBDD	$-5.98_{\scriptscriptstyle\pm 1.16}$	$0.08_{\scriptscriptstyle\pm 0.46}$	$1.12$	$0.92$	$0.54$	$0.06$	$2.76_{\scriptscriptstyle\pm 1.37}$	$0.07_{\scriptscriptstyle\pm 0.53}$	$0.53$	$0.09$
TargetDiff	$-6.42_{\scriptscriptstyle\pm 1.15}$	$0.12_{\scriptscriptstyle\pm 0.57}$	$1.27$	$1.08$	$0.55$	$0.11$	$2.24_{\scriptscriptstyle\pm 1.32}$	$0.10_{\scriptscriptstyle\pm 0.62}$	$0.55$	$0.11$
MolForm	$-6.75_{\scriptscriptstyle\pm 1.30}$	$0.10_{\scriptscriptstyle\pm 0.52}$	$1.13$	$0.97$	$0.53$	$0.10$	$1.89_{\scriptscriptstyle\pm 1.51}$	$0.10_{\scriptscriptstyle\pm 0.58}$	$0.53$	$0.11$

Appendix

Table A9: SelectBench -XL evaluation. Metrics are defined as in the case-study table and computed per on-/off-target pair. Values are means over all pairs in the tier; where shown,

\pm

is the standard deviation across molecules within a pair, averaged over pairs.

\bar{s}_{\text{on}}^{z}

\bar{\Delta}^{z}

\mathrm{SR}^{z}_{\tau=0}

\mathrm{SR}^{z}_{\tau=1}

: z-score variants computed against per-pocket reference distributions;

\tau=1

in z-space corresponds to one reference-

\sigma

margin.

Method	\hlcustom Blue1 PARP1	\hlcustom Pink1 BMP2K	\hlcustom Yellow2 CDK9	\hlcustom Purple1 ERK2	\hlcustom Blue2 DRD2
Standard SBDD models
DecompDiff	$0.14$	$-3.63$	$0.49$	$0.74$	$-8.27$
DiffSBDD	$0.65$	$1.13$	$-0.05$	$-0.16$	$-6.87$
TargetDiff	$1.74$	$-0.16$	$2.10$	$0.50$	$-4.11$
MolForm	$-1.23$	$-2.76$	$1.04$	$3.09$	$-0.28$
MolSnapper	$0.91$	$0.82$	$0.53$	$1.25$	$-8.65$

Appendix

Table A10: SelectBench -Case potency and selectivity evaluation. For each method, we report

\mathrm{PS}

; higher values indicate methods yielding jointly potent and selective molecules.

Method	\hlcustom Blue1 PARP1	\hlcustom Pink1 JAK2	\hlcustom Yellow2 BRAF	\hlcustom Purple1 5-HT1B	\hlcustom Blue2 FA7
Standard SBDD models
DecompDiff	$-4.23$	$0.91$	$-1.86$	$4.09$	$-1.79$
DiffSBDD	$2.81$	$2.03$	$-1.17$	$-0.88$	$-1.73$
TargetDiff	$6.09$	$1.01$	$1.20$	$5.70$	$1.78$
MolForm	$8.36$	$3.25$	$7.25$	$3.89$	$-1.41$
MolSnapper	$1.87$	$-1.45$	$0.57$	$-2.77$	$3.32$

Appendix

Table A11: SelectBench -Safety potency and selectivity evaluation. For each method, we report

\mathrm{PS}

; higher values indicate methods yielding jointly potent and selective molecules.

Method	\hlcustom Blue1 Easy	\hlcustom Pink1 Medium	\hlcustom Yellow2 Hard
Standard SBDD models
DecompDiff	$1.98$	$0.59$	$0.38$
DiffSBDD	$0.36$	$0.48$	$0.42$
TargetDiff	$0.19$	$1.14$	$0.64$
MolForm	$0.50$	$1.20$	$0.68$
MolSnapper	$0.86$	$0.07$	$0.78$

Appendix

Table A12: SelectBench -XL potency and selectivity evaluation. For each method, we report

\mathrm{PS}

; higher values indicate methods yielding jointly potent and selective molecules.

Method	QED $\,(\uparrow)$	SA $\,(\uparrow)$	LogP	Lipinski $\,(\uparrow)$	MW	Diversity $\,(\uparrow)$
\hlcustom Blue1 PARP1 (on) / PARP2 (off)
Standard SBDD models
DecompDiff	$0.22_{\scriptscriptstyle\pm 0.12}$	$0.57_{\scriptscriptstyle\pm 0.07}$	$1.75_{\scriptscriptstyle\pm 2.01}$	$3.48_{\scriptscriptstyle\pm 0.74}$	$412.72_{\scriptscriptstyle\pm 18.46}$	$0.799$
DiffSBDD	$0.39_{\scriptscriptstyle\pm 0.21}$	$0.58_{\scriptscriptstyle\pm 0.12}$	$1.29_{\scriptscriptstyle\pm 1.94}$	$4.16_{\scriptscriptstyle\pm 1.20}$	$415.26_{\scriptscriptstyle\pm 134.94}$	$0.677$
TargetDiff	$0.25_{\scriptscriptstyle\pm 0.17}$	$0.50_{\scriptscriptstyle\pm 0.09}$	$2.05_{\scriptscriptstyle\pm 1.93}$	$3.42_{\scriptscriptstyle\pm 1.39}$	$495.64_{\scriptscriptstyle\pm 94.75}$	$0.667$
MolForm	$0.23_{\scriptscriptstyle\pm 0.10}$	$0.56_{\scriptscriptstyle\pm 0.11}$	$9.00_{\scriptscriptstyle\pm 1.43}$	$2.98_{\scriptscriptstyle\pm 0.73}$	$483.83_{\scriptscriptstyle\pm 100.48}$	$0.728$

Appendix

Table A13: SelectBench -Safety evaluation on additional molecular generation metrics.

Method	QED $\,(\uparrow)$	SA $\,(\uparrow)$	LogP	Lipinski $\,(\uparrow)$	MW	Diversity $\,(\uparrow)$
\hlcustom Blue1 POI: 5HT1B
Standard SBDD models
DecompDiff	$0.18_{\scriptscriptstyle\pm 0.09}$	$0.45_{\scriptscriptstyle\pm 0.09}$	$6.31_{\scriptscriptstyle\pm 2.35}$	$2.56_{\scriptscriptstyle\pm 0.75}$	$605.66_{\scriptscriptstyle\pm 10.96}$	$0.662$
DiffSBDD	$0.43_{\scriptscriptstyle\pm 0.21}$	$0.58_{\scriptscriptstyle\pm 0.14}$	$2.87_{\scriptscriptstyle\pm 1.72}$	$4.44_{\scriptscriptstyle\pm 0.86}$	$405.97_{\scriptscriptstyle\pm 156.40}$	$0.721$
TargetDiff	$0.36_{\scriptscriptstyle\pm 0.16}$	$0.50_{\scriptscriptstyle\pm 0.13}$	$5.98_{\scriptscriptstyle\pm 2.04}$	$3.51_{\scriptscriptstyle\pm 0.93}$	$495.19_{\scriptscriptstyle\pm 100.06}$	$0.690$
MolForm	$0.21_{\scriptscriptstyle\pm 0.08}$	$0.52_{\scriptscriptstyle\pm 0.14}$	$11.19_{\scriptscriptstyle\pm 1.55}$	$2.81_{\scriptscriptstyle\pm 0.57}$	$522.51_{\scriptscriptstyle\pm 117.95}$	$0.726$

Appendix

Table A14: SelectBench -Safety evaluation on additional molecular generation metrics.

Method	QED $\,(\uparrow)$	SA $\,(\uparrow)$	LogP	Lipinski $\,(\uparrow)$	MW	Diversity $\,(\uparrow)$
\hlcustom Blue1 Hard
Standard SBDD models
DecompDiff	$0.44_{\scriptscriptstyle\pm 0.12}$	$0.68_{\scriptscriptstyle\pm 0.10}$	$1.92_{\scriptscriptstyle\pm 2.24}$	$4.44_{\scriptscriptstyle\pm 0.77}$	$270.66_{\scriptscriptstyle\pm 163.66}$	$0.80_{\scriptscriptstyle\pm 0.08}$
DiffSBDD	$0.50_{\scriptscriptstyle\pm 0.05}$	$0.64_{\scriptscriptstyle\pm 0.04}$	$0.83_{\scriptscriptstyle\pm 0.82}$	$4.77_{\scriptscriptstyle\pm 0.17}$	$270.73_{\scriptscriptstyle\pm 67.72}$	$0.82_{\scriptscriptstyle\pm 0.05}$
TargetDiff	$0.47_{\scriptscriptstyle\pm 0.07}$	$0.58_{\scriptscriptstyle\pm 0.06}$	$1.44_{\scriptscriptstyle\pm 1.74}$	$4.53_{\scriptscriptstyle\pm 0.32}$	$337.44_{\scriptscriptstyle\pm 110.22}$	$0.72_{\scriptscriptstyle\pm 0.09}$
MolForm	$0.45_{\scriptscriptstyle\pm 0.07}$	$0.67_{\scriptscriptstyle\pm 0.11}$	$5.62_{\scriptscriptstyle\pm 1.93}$	$4.17_{\scriptscriptstyle\pm 0.56}$	$309.89_{\scriptscriptstyle\pm 117.62}$	$0.75_{\scriptscriptstyle\pm 0.08}$

Appendix

Table A15: SelectBench -Safety evaluation on additional molecular generation metrics.

Figure A2: SelectBench -Case on-/off-target pairs protein structures overlay and their associated similarity. This visually demonstrates the high structural similarity of pairs in SelectBench -Case. Co-crystallized ligands are displayed in yellow sticks. Reported ProBiS Z-scores quantify global structural similarity of the structures, with higher values indicating stronger similarity.

Figure A3: Heatmaps of structural similarity between POIs and the off-targets panel in SelectBench -Safety. Each row corresponds to an on-target, each column corresponds to an off-target, and color intensity reflects the ProBiS Z-score, with darker shades indicating higher local structural similarity. ProBiS Z-score of zero or below are set to white. Priority targets of SelectBench -Safety-Priority are highlighted by the yellow box, the HTR1B off-target structure excluded due to overlap with POI 5HT1B is highlighted by the blue box.

Figure A4: Heatmaps of structural similarity between on-/off-target pairs in SelectBench -XL. Each row corresponds to an on-target, each column corresponds to an off-target, and color intensity reflects the ProBiS Z-score, with darker shades indicating higher local structural similarity. ProBiS Z-score of zero or below are set to white. Grey shade indicates self-pairs, which are excluded from SelectBench -XL.

Structure-Based Drug Design

Structure-based drug design uses protein structures and binding-site geometry to find, rank, and create molecules that bind therapeutic targets.

See all 44 papers

Generative molecule evolution using 3D pharmacophore for efficient Structure-Based Drug Design
Jul 27, 2025Yi He, Ailun Wang, Zhi Wang +3Structure-Based Drug Design Molecular Generation
ByteDance Seed
ShallowBench: Benchmarking Generative Drug Design Models on Shallow-Pocket Targets
Jun 4, 2026Saket Reddy, Shiwei LiuBenchmark Design Structure-Based Drug Design
University of Illinois - Urbana-Champaign
Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization
Jul 14, 2026Ruoxi Gao, Jiangweizhi Peng, Ziqi Chen +10Molecular Optimization Structure-Based Drug Design
Computer Science and Engineering, The Ohio State University, Columbus, OH 43210. · Industrial and System Engineering, University of Minnesota, Minneapolis, MN 55455. · Google Cloud, Google LLC, Mountain View, CA 94043. +6

Jul 27, 2025J/K move · Enter open · S save

Generative molecule evolution using 3D pharmacophore for efficient Structure-Based Drug Design

Yi He, Ailun Wang, Zhi Wang, Yu Liu+2

ByteDance Seed

Structure-Based Drug Design Molecular Generation Molecular Optimization

Open paper AskPDF arXiv

Recent breakthroughs in generative models have shown great promise for drug discovery, yet Structure-Based Drug Design (SBDD) faces two critical challenges: the scarcity of high-quality protein-ligand complex data and the practical synthesizability of generated molecules. While ligand-based approaches benefit from billion-scale small molecule datasets, they often lack the structural context of the target protein. To address these dual challenges of data scarcity and synthetic inaccessibility, we propose MEVO, a generative molecular evolutionary framework that operates within a synthesizable chemical space. MEVO introduces a reaction-aware generative model trained on reaction routes constructed by applying predefined reaction templates to Enamine Building Blocks. The model jointly generates molecules and corresponding synthesis sketches consistent with these templates. By integrating this generative capability into a pocket-aware evolutionary strategy, MEVO iteratively refines molecular structures within the synthesizable chemical space, guided by a hierarchical scoring function that combines fast approximation with rigorous Absolute Binding Free Energy Perturbation (ABFEP) calculations. This framework efficiently generates high-affinity, synthetically feasible binders for diverse protein targets. We validate MEVO's performance through rigorous ABFEP calculations, demonstrating its ability to design potent inhibitors for challenging targets such as KRAS

^{\textrm{G12D}}

, achieving affinities comparable to known active inhibitors. By integrating the data richness of ligand-based learning with the precision of structure-based optimization, MEVO offers a holistic solution that ensures both high binding affinity and synthetic accessibility, paving the way for more efficient and practical structure-based drug design.

Abstract

Structure-Based Drug Design

Explore similar work

Generative molecule evolution using 3D pharmacophore for efficient Structure-Based Drug Design

ShallowBench: Benchmarking Generative Drug Design Models on Shallow-Pocket Targets

Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization