AI-guided Antibiotic Discovery Pipeline from Target Selection to Compound Identification

TL;DR

The study presents an end-to-end AI-guided antibiotic discovery pipeline that moves from structure-based target identification to real-world compound realization. By clustering predicted pathogen proteomes and benchmarking six 3D-structure-aware generative models, it demonstrates practical guidance for integrating DL tools into early-stage antibiotic design. DeepBlock and TamGen emerge as top performers across usability and output quality, with rigorous post-generation curation reducing vast output to synthesizable candidates. The work provides a benchmark, blueprint, and actionable framework for deploying AI in antibiotic discovery, with significant implications for accelerating the identification of novel antibacterial agents and reducing synthesis barriers.

Abstract

Antibiotic resistance presents a growing global health crisis, demanding new therapeutic strategies that target novel bacterial mechanisms. Recent advances in protein structure prediction and machine learning-driven molecule generation offer a promising opportunity to accelerate drug discovery. However, practical guidance on selecting and integrating these models into real-world pipelines remains limited. In this study, we develop an end-to-end, artificial intelligence-guided antibiotic discovery pipeline that spans target identification to compound realization. We leverage structure-based clustering across predicted proteomes of multiple pathogens to identify conserved, essential, and non-human-homologous targets. We then systematically evaluate six leading 3D-structure-aware generative models$\unicode{x2014}$spanning diffusion, autoregressive, graph neural network, and language model architectures$\unicode{x2014}$on their usability, chemical validity, and biological relevance. Rigorous post-processing filters and commercial analogue searches reduce over 100 000 generated compounds to a focused, synthesizable set. Our results highlight DeepBlock and TamGen as top performers across diverse criteria, while also revealing critical trade-offs between model complexity, usability, and output quality. This work provides a comparative benchmark and blueprint for deploying artificial intelligence in early-stage antibiotic development.

Figures (12)

• Figure 1: Foldseek search similarity of target proteins to CrossDocked2020 training set. We show the top1 result of each Foldseek search for case study proteins from the respective model publications ($\alpha=0.6$) and our 3 targets ($\alpha=1$) against the CrossDocked2020 training set. For DiffSBDD and ResGen multiple case study proteins were scored. Pocket2Mol and TargetDiff were omitted as the respective publications show now independent case studies. Three metrics are reported: Probability score (Foldseek’s estimated probability of the top1 hit to be homologous to the query structure), the global TM score, and LDDT score. The performed case studies per model were: DeepBlock with NCEH1; DifSBDD with BIKE and MPSK1; ResGen with AKT1 and CDK2; TamGen with CLPP2_MYCTU.
• Figure 2: Model implementation ranking.a) We ranked the ease of installation from the respective GitHub repositories. b) We rated the documentation of the repositories and the underlying codebases. c) We compared the models' usability, focused on sampling with pretrained models. d) We rated the input options and ease of input generation. e) We evaluated the output format, especially the contents and ease of further data processing. f) We showed an overall rating averaged across the previous characteristics.
• Figure 3: Overview of generated molecules.a) We visualize the number of actually generated molecules based on model and target protein. We prompted all models to generate 10000.0 molecules per target. b) We show the validity of molecules in relation to the SBDD model. The validity is determined by our SMILESCleaner class.
• Figure 4: Model-based chemical space analysis of the generated molecules. The UMAP space is presented for each model separately (a--f), based on ECFPs. Colors represent the different protein targets. We sampled up to 1000 molecules per subplot. Only valid molecules were considered.
• Figure 5: Structural alerts of generated molecules. Here, we analysed the number of Dundee alerts based on model (a) as well as target (b). Only valid molecules were considered.