acceptodds
Under review as a conference paper at ICLR 2027

HBDrug3D: A Unified Benchmark Dataset of Heterobifunctional Drugs for Geometric Linker Design

Abstract

Heterobifunctional drugs, including proteolysis-targeting chimeras (PROTACs), peptide-drug conjugates (PDCs), and antibody-drug conjugates (ADCs), have emerged as important therapeutic modalities. However, the scarcity of publicly available 3D data and standardized evaluation protocols limits the development and comparison of machine learning methods for these molecules. We introduce HBDrug3D, one of the largest publicly available datasets of heterobifunctional drugs, comprising 70,393 molecular conformations across PROTACs, PDCs, and ADCs. Each molecule is annotated at the component level, identifying its linker and modality-specific functional groups, including warheads, payloads, and E3 ligands. Based on HBDrug3D, we establish a standardized benchmark for geometric linker design, a central task in heterobifunctional drug development. The benchmark evaluates generated linkers in terms of validity, novelty, and physicochemical properties relevant to medicinal chemistry. By integrating diverse molecular modalities, component-level annotations, and reproducible evaluation protocols, HBDrug3D provides a common foundation for developing and systematically comparing machine learning approaches to heterobifunctional drug design.

Primary area:
datasets and benchmarks
Keywords:
Heterobifunctional Molecules, PROTACs, 3D Molecular Conformations, Geometric Deep Learning, 3D Molecular Generation, Benchmark Dataset

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8edMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HBDrug3D, a 3D dataset for linker generation in PROTACs, PDCs, and ADCs. The authors generate 3D conformations from curated molecules and evaluate four existing linker generation methods on the dataset. The benchmark includes both random and scaffold splits, with metrics for validity, recovery, 3D geometry, and molecular properties.

Critical strengths

The paper studies a useful setting that is not well covered by standard small-molecule linker benchmarks. The results also show that existing methods struggle with longer linkers, especially for ADCs, which I found useful.

Critical weaknesses
  1. I am not fully convinced that the 3D benchmark reflects the relevant linker design setting. The conformations are generated from complete molecules using ConfGen, without the biological context in which the linker needs to function. For PROTACs in particular, the relevant geometry depends on the target–PROTAC–E3 complex. It is therefore unclear to me how well performance on this benchmark reflects useful linker design in practice.
  2. I have concerns about the data split. The paper states that each conformation is treated as a separate data point in the random split. This seems to allow different conformations of the same molecule to appear in both training and test sets. The scaffold split is more convincing, but I could not find enough detail on how the “fragment scaffold” is defined or how many unique molecules are present in each split.
Questions for authors

1.How are the relative 3D positions of the input fragments obtained? Are they inherited directly from the ConfGen conformations of the complete molecule?

2.Can different conformations of the same molecule appear in both training and test sets under the random split?

3.How exactly is the “fragment scaffold” defined in the scaffold split?

Reviewer_aHLAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript provides a combined dataset and benchmark for linker generation for heterobifunctional drugs (HBDs), i.e. poteolysis targeting chimeras (PROTACs), antibody-drug conjugates (ADCs), and pepdite-drug conjugates (PDCs). Starting from a collection of around 7K reported HBDs, the dataset first annotates functional groups, reconstruct possible ADC conjugation constructs, and generate molecular conformers in absense of a protein pocket, resulting in around 70K datapoints. The authors then retrain and test generative models on generated conformers, benchmarking chemical properties, novelty and geometric validity. They identify decrease in performance with longer linkers and plausible 3D geometry.

Critical strengths

The paper constitutes a useful resource for a specific domain of application. The combination of different HBD modalities, and the curated annotations lowers the effort required to develop and compare 3D generative models for linker generation. The presented shortcomings of the shown baselines in generating long, flexible linkers could help improve generative models in this specific task.

Critical weaknesses
  • The motivation behind the setup of this linker design benchmark is not validated. The relative fragment positioning comes from computationally generated, pocket-agnostic conformers. The benchmark task then asks whether models can generate compatible linkers for these fragments, but this doesn’t necessarily relate to biologically relevant linker design: Isolated conformations do not capture the protein-complex of PROTACS or ADCs. In contrast to other pocket-agnostic methods (like DeLinker), the starting fragments don’t come from an observed geometry but freely generated conformers and therefore add little value above 2D methods.
  • For ADCs, the paper describes molecule generation depending on conjugation types, but does not include details if the attachment was inferred from experimental records, or inferred from general conjugation rules.
  • The benchmark and experimental setup lacks clarity (see below).
Questions for authors
  • For the ADC subset, which attachment atoms and product structures are supported from the experimental/source recods, and which are inferred?
  • Is the scaffold for the split defined on the full molecule, or on eigher fragment?
  • How was bond annotation performed for methods without explicit generation of bond types?
  • What is the reason for 3Dlinker’s zero PB.val score?
  • Does retraining on HBDrug improve baseline performance?
Reviewer_XHscOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces HBDrug3D, a dataset of 70,393 computed conformations covering PROTACs, peptide–drug conjugates, and antibody-conjugation-derived structures. The benchmark evaluates DeLinker, 3DLinker, DiffLinker, and LinkerNet, trained from scratch on individual modalities and the combined dataset under random and scaffold splits.

Critical strengths
  1. The proposed dataset may broaden the chemical settings available for evaluating linker-generation models. The inclusion of PROTACs, ADC-derived conjugation structures, and peptide-containing compounds exposes models to substantially different linker sizes and molecular contexts.

  2. The paper also makes some initial efforts to the systematic evaluation through scaffold-grouped partitions, multiple metrics, and repeated runs.

Critical weaknesses
  1. The most critical issue of this work is that the contribution is quite limited to the community. The resource’s incremental contribution and reference-data quality are not established sufficiently, and the positioning overstates the absence of prior work. The introduction states that existing heterobifunctional resources lack 3D structural information and that the discussed methods have only been evaluated on general small-molecule datasets. However, the cited PROTAC-DB 3.0 resource includes experimental and predicted ternary structures, and the original LinkerNet paper already evaluates LinkerNet, DiffLinker, and DeLinker on PROTAC-DB. It is quite clear that this work does not provide much new insights to the community.

  2. The scale comparison also needs consistent units: Table 1 places HBDrug3D’s conformer count alongside molecule or processed-example counts for other datasets. The final number of distinct reconstructed ADC molecular graphs and the handling of source records that map to the same CLP structure are not clearly reported.

  3. The primary generalization protocol is quite ambiguous. Section 2.1 refers to molecular scaffolds, whereas Section 3.1 refers to fragment scaffolds. Although the main results use scaffold splitting, the paper does not establish precisely which chemical entities are held out.

Questions for authors
  1. What is the real contribution of this work compared to some prior work, like PROTAC-DB 3.0. The current paper version is clearly overstating the contribution.

  2. Please specify how each model’s final coordinates and bonds are obtained, whether fragment poses remain fixed, and what embedding or relaxation is applied before geometric evaluation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.