RangeFinder: Predicting Ligand Property Ranges from Protein Pockets

Alan Nadelsticher Ruvalcaba; Chris Alvin; Supratik Mukhopadhyay; Adam Bess

cs.LGNov 2, 2026

RangeFinder: Predicting Ligand Property Ranges from Protein Pockets

Authors: Alan Nadelsticher Ruvalcaba, Chris Alvin, Supratik Mukhopadhyay, Adam Bess

Organizations: Georgia Institute of Technology · Furman University · Louisiana State University

Abstract

Identifying ligands for a protein target requires searching a vast chemical space for a comparatively small set of compatible molecules. Existing machine learning approaches primarily address this problem by generating candidate ligands or scoring individual protein-ligand pairs, but provide limited information about the broader chemical space compatible with a binding pocket. We introduce RangeFinder, an SE(3)-equivariant transformer pocket-level model, trained and evaluated on BioLiP2 data, that predicts ranges over nine physicochemical ligand properties directly from pocket structure, chemistry, and evolutionary context. Using an intersection-over-union objective and evaluation metric (range-IoU), RangeFinder achieves a range-IoU of $0.464 \pm 0.002$ , outperforming all deployable baselines while requiring only $12.2$ ms per pocket. We also quantify our method's ability to reduce the ligand search space while retaining compatible binders; in this regard, RangeFinder achieves a Reduction-Recall F-Score of $0.543 \pm 0.011$ --- once again outperforming all deployable baselines. Ablation experiments indicate that RangeFinder's performance depends more strongly on the density and quality of ligand supervision than on any individual input modality or specialized training strategy. Together, these results position RangeFinder as an inexpensive and effective molecular filtering prior, with demonstrated potential in narrowing the ligand search space of target pockets before docking, generation, or synthesis. The code for this project can be found at: https://github.com/alannadels/RangeFinder.

Figures & tables

Method	Range-IoU	Reduction	Recall	RRF	Latency
Global average	0.402	$3.0\times$	0.205	0.313	$<0.001$ ms
Foldseek	0.390	$4.2\times$	0.170	0.278	$127$ ms
Pocket cluster *	$0.403\pm 0.000$	$3.3\times$	$0.222\pm 0.004$	$0.338\pm 0.004$	$2.83$ ms
Feature $k$ -NN *	$0.409\pm 0.000$	$4.1\times$	$0.214\pm 0.002$	$0.333\pm 0.002$	$3.01$ ms
APoc	0.420	$5.3\times$	0.287	0.424	$26.3$ s
RangeFinder (Ours) *	$0.464\pm 0.002$	$2.1\times$	$0.557\pm 0.026$	$0.543\pm 0.011$	$12.2$ ms

Table 1: Main experimental results on BioLiP2 data. RangeFinder was trained with seeds

42

–

46

(five seeds); the other marked methods were run with seeds

0

–

4

(five seeds). Bold / underline : best/second-best per column, excluding the non-deployable reference.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Property	Definition
MW	Molecular weight
logP	Octanol–water partition coefficient (lipophilicity)
HBD	Number of hydrogen-bond donors
HBA	Number of hydrogen-bond acceptors
TPSA	Topological polar surface area
Rotatable bonds	Number of rotatable bonds

Appendix

Table 2: The nine ligand physicochemical properties predicted by RangeFinder.

Figure 1: Stepwise visualization of the structural pocket–ligand representation on which the model operates. The protein shown in the diagram, N-acyl-homoserine-lactone lactonase (PDB ID 6N9Q), represents a real example of a protein containing only one pocket as an active binding site.

Figure 2: Stage 1 – Input embedding. Each pocket atom receives $46$ local chemistry and structural descriptors together with its residue’s $1280$ -dimensional ESM-2 embedding. The resulting $1326$ -dimensional feature vector is projected to $32$ invariant scalar channels, while $16$ equivariant vector channels are initialized to zero.

Figure 3: Stage 2 – Equivariant attention. The zero-initialized vector channels acquire directional information from neighboring atoms within $5$ Å, while the scalar channels remain rotation invariant. The hidden-state width remains at $80$ components per atom: $32$ scalars and $16$ three-dimensional vectors.

	Train	Validation	Test
Pockets ( $n_{\mathrm{lig}}\geq 3$ )	2,627	562	564
Structure-views	10,905	2,313	3,823

Appendix

Table 3: BioLiP2-based dataset split by protein. A structure-view is one observed protein–ligand entry associated with a pocket.

Architecture
Setting	Value
Encoder	SE(3)-Transformer ( e3nn ); $4$ layers, $4$ heads, $\ell_{\max}=1$
Hidden irreducible reps. (width)	$32\times 0e+16\times 1o$ ( $80$ components per atom)
Radius graph / radial basis	$5$ Å cutoff; $16$ Gaussian radial basis functions
Prediction head	MLP $96\!\to\!48\!\to\!19$ ( $18$ interval parameters $+1$ aromatic logit)
Input features and normalization

Appendix

Table 4: RangeFinder training, loss, pocket-construction, and CMA-ES search configuration.

	BioLiP2	PDBbind
Global-average floor	0.402	0.403
Pocket cluster $K{=}6$ (geometry-blind)	0.403	0.431
RangeFinder	0.467	0.437
RangeFinder gain over geometry-blind baseline	+0.064	+0.006

Appendix

Table 5: Range-IoU under two supervision sources, with the model held fixed. Each column uses its own test split; compare each model’s margin over its same-dataset baselines. RangeFinder variants were trained under the same fixed seed (seed

=42

Variant	ESM	Chemistry	3D	Feature dim.	Range-IoU	Latency
Full model	✓	✓	✓	1326	0.467	12.2 ms
No ESM	—	✓	✓	46	0.464	12.2 ms
No chemistry	✓	—	✓	1280	0.464	12.2 ms
No 3D	✓	✓	—	1326	0.455	6.8 ms

Appendix

Table 6: Leave-one-out input ablation. All other training and evaluation choices are held fixed. All ablation variants shown were trained under the same fixed seed (seed

=42

). Bold / underline : best/second-best per column.

Variant	Strategy	Range-IoU	Latency
RangeFinder	Single model	0.467	12.2 ms
Denoising-pretrained encoder	Self-supervised pretraining	0.461	12.2 ms
Mixture of experts ( $K{=}6$ )	Per-cluster specialization	0.442	15 ms

Appendix

Table 7: Training-strategy ablation. Alternative training strategies and architectures for RangeFinder on the same

564

BioLiP2 test pockets. All ablation variants shown were trained under the same fixed seed (seed

=42

). Bold / underline : best/second-best per column.

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 47 papers

Structure-guided molecular design with contrastive 3D protein-ligand learning
Apr 21, 2026Carles Navarro, Philipp Tholke, Gianni de FabritiisContrastive Learning Virtual Screening
Acellera Labs, C/ Doctor Trueta 183, 08005 Barcelona, Spain · Computational Science Laboratory, Universitat Pompeu Fabra, PRBB, C/ Doctor Aiguader 88, 08003 Barcelona, Spain · Instituci´o Catalana de Recerca i Estudis Avan¸cats (ICREA), Passeig Llu´ıs Companys 23, 08010 Barcelona, Spain
zPocket: Better Pocket Finding with Frozen Co-Folding Features
Oct 13, 2026Tally Portnoi, Allison M Keys, Benjamin R DiFrancesco +1Drug Discovery Structure-Based Drug Design
MIT · Genesis Molecular AI
Pyrite: Building and Benchmarking Protein-Ligand Pose Ranking Agents
Nov 10, 2026Megha Srivastava, Michael H. Lim, Allison M. Keys +2Learning to Rank Molecular Docking
Stanford University · Genesis Molecular AI

Apr 21, 2026J/K move · Enter open · S save

Structure-guided molecular design with contrastive 3D protein-ligand learning

Carles Navarro, Philipp Tholke, Gianni de Fabritiis

Acellera Labs, C/ Doctor Trueta 183, 08005 Barcelona, Spain · Computational Science Laboratory, Universitat Pompeu Fabra, PRBB, C/ Doctor Aiguader 88, 08003 Barcelona, Spain · Instituci´o Catalana de Recerca i Estudis Avan¸cats (ICREA), Passeig Llu´ıs Companys 23, 08010 Barcelona, Spain

Contrastive Learning Virtual Screening Structure-Based Drug Design 3D Molecular Generation 3D Representation Learning Molecular Generation+2

Open paper AskPDF arXiv

Structure-based drug discovery faces the dual challenge of accurately capturing 3D protein-ligand interactions while navigating ultra-large chemical spaces to identify synthetically accessible candidates. In this work, we present a unified framework that addresses these challenges by combining contrastive 3D structure encoding with autoregressive molecular generation conditioned on commercial compound spaces. First, we introduce an SE(3)-equivariant transformer that encodes ligand and pocket structures into a shared embedding space via contrastive learning, achieving competitive results in zero-shot virtual screening. Second, we integrate these embeddings into a multimodal Chemical Language Model (MCLM). The model generates target-specific molecules conditioned on either pocket or ligand structures, with a learned dataset token that steers the output toward targeted chemical spaces, yielding candidates with favorable predicted binding properties across diverse targets.

Abstract

Structure-Based Drug Design

Explore similar work

Structure-guided molecular design with contrastive 3D protein-ligand learning

zPocket: Better Pocket Finding with Frozen Co-Folding Features

Pyrite: Building and Benchmarking Protein-Ligand Pose Ranking Agents