Beyond Affinity: A Benchmark of 1D, 2D, and 3D Methods Reveals Critical Trade-offs in Structure-Based Drug Design

TL;DR

This work addresses the lack of cross-paradigm benchmarks in structure-based drug design by evaluating fifteen models spanning 1D (SMILES/SELFIES), 2D (graph-based) and 3D (pocket-aware) representations under a unified framework. It introduces a multi-faceted benchmark that combines docking scores, pose validity, and pharmaceutical-property metrics, revealing a clear affinity-validity trade-off: 3D methods excel at binding affinity but often struggle with chemical validity and pose quality, while 1D methods are reliable on standard molecular metrics but rarely achieve top affinities, and 2D methods offer a balanced middle ground. The study also shows that simple hybrids (e.g., TamGen) do not automatically resolve these trade-offs, highlighting the need for hybrid architectures that robustly enforce chemistry during generation alongside 3D structural constraints. Overall, the benchmark provides actionable insights and a public codebase to guide future development of SBDD methods that balance affinity with chemical and structural validity, accelerating practical drug discovery.

Abstract

Currently, the field of structure-based drug design is dominated by three main types of algorithms: search-based algorithms, deep generative models, and reinforcement learning. While existing works have typically focused on comparing models within a single algorithmic category, cross-algorithm comparisons remain scarce. In this paper, to fill the gap, we establish a benchmark to evaluate the performance of fifteen models across these different algorithmic foundations by assessing the pharmaceutical properties of the generated molecules and their docking affinities and poses with specified target proteins. We highlight the unique advantages of each algorithmic approach and offer recommendations for the design of future SBDD models. We emphasize that 1D/2D ligand-centric drug design methods can be used in SBDD by treating the docking function as a black-box oracle, which is typically neglected. Our evaluation reveals distinct patterns across model categories. 3D structure-based models excel in binding affinities but show inconsistencies in chemical validity and pose quality. 1D models demonstrate reliable performance in standard molecular metrics but rarely achieve optimal binding affinities. 2D models offer balanced performance, maintaining high chemical validity while achieving moderate binding scores. Through detailed analysis across multiple protein targets, we identify key improvement areas for each model category, providing insights for researchers to combine strengths of different approaches while addressing their limitations. All the code that are used for benchmarking is available in https://github.com/zkysfls/2025-sbdd-benchmark

Figures (7)

• Figure 1: The bar chart of average generated molecules that are calculated by our selected oracles for each model across all target proteins under given time. 1D methods are colored orange, blue is used to indicate 3D methods, and green represents 2D methods.
• Figure 2: The cumulative density function (CDF) of strain energy of each model
• Figure 3: The heatmap based on the average of each model's Top-10 docking score for each target protein.
• Figure 4: The cumulative density function (CDF) of RMSD for all 3D models
• Figure 5: The clashes box plot for 1D models