AI approaches for the discovery and validation of drug targets

Aaron Wenteler, Claudia Cabrera, Wei Wei, Victor Neduva, Michael R. BarnesCambridge 2024 · DOI

Abstract

Artificial intelligence (AI) holds immense promise for accelerating and improving all aspects of drug discovery, not least target discovery and validation. By integrating a diverse range of biological data modalities, AI enables the accurate prediction of drug target properties, ultimately illuminating biological mechanisms of disease and guiding drug discovery strategies. Despite the indisputable potential of AI in drug target discovery, there are many challenges and obstacles yet to be overcome, including dealing with data biases, model interpretability and generalisability, and the validation of predicted drug targets, to name a few. By exploring recent advancements in AI, this review showcases current applications of AI for drug target discovery and offers perspectives on the future of AI for the discovery and validation of drug targets, paving the way for the generation of novel and safer pharmaceuticals.

Introduction

The pharmaceutical industry currently faces a "chemistry arms race" that is increasingly unsustainable. Historically, drug discovery relied on phenotypic assays—testing compounds on cells to observe effects without necessarily understanding the underlying mechanism. While the industry has shifted toward a "target-first" approach, identifying specific proteins or genes to modulate, this has created a significant bottleneck in target validation. Poorly validated targets are primary drivers of high attrition rates in clinical trials, leading to billions of dollars in losses during late-stage failures. Furthermore, the "druggable genome" remains remarkably narrow, with over 50% of current drugs targeting just four protein families. This review explores how artificial intelligence (AI) is dismantling these bottlenecks by integrating multi-modal data to navigate the "unknome"—the vast, unexplored portion of the human genome—to identify novel, safe, and efficacious targets.

Methodological Frameworks

AI architectures are transforming the field through several key approaches:

  • Ensemble Learning: Tools like DrugnomeAI combine multiple classifiers to rank candidate genes based on druggability features, such as protein-protein interaction networks.
  • Foundation Models & Transfer Learning: Transformers like Geneformer, trained on massive datasets such as the 30-million-cell Genecorpus-30M, learn the "language" of gene expression to predict disease responses.
  • Graph Neural Networks (GNNs): Models such as EMOGI treat biological systems as networks, identifying cancer-driving genes within complex multiomics data.
  • Generative AI: Used for de novo protein design and molecular docking, enabling the discovery of new binding sites and structures.
  • Multimodal Integration: Late fusion techniques allow the simultaneous processing of disparate data types, including histopathology images, chemical notations (SMILES), and electronic health records.

Data Ecosystems

The efficacy of these models relies on foundational "Big Data" repositories:

  • Genomics/Tabular: UK Biobank, Genes & Health, Open Targets, and TCGA (The Cancer Genome Atlas).
  • Structural/Sequence: PDB (Protein Data Bank), UniProt, and ChEMBL.
  • Single-Cell Transcriptomics: Large-scale datasets like Genecorpus-30M.
  • Knowledge Graphs: STRING, STITCH, and Reactome for mapping biological relationships.

Strategic Insights: The AI "Sweet Spot"

The efficacy of AI in this domain is maximized when three specific criteria intersect: Scale, Complexity, and Non-linearity. Drug target discovery uniquely satisfies all three, making it an ideal candidate for AI intervention.

The "Sweet Spot" for AI impact in drug discovery, defined by the intersection of scale, complexity, and non-linearity.

A critical insight is that drug targets supported by genetic evidence are 80% more likely to succeed in clinical trials. AI excels at extracting this evidence from population-scale data. Furthermore, AI facilitates the identification of allosteric sites—binding pockets away from a protein's main active site—allowing for more selective drugs. Models also enable "in silico perturbation," or digital experiments that simulate gene knockdown in virtual cells to predict outcomes before physical laboratory work begins.

Clinical Impact and Results

The transition to AI-driven discovery is yielding tangible results. It is estimated that in silico environments provide 25% to 50% cost and time savings up to the preclinical stage.

Current landscape of AI-derived therapeutics in clinical pipelines, categorized by company and development stage.

As shown in the clinical pipeline data, companies such as Insilico Medicine, Recursion, and BenevolentAI have successfully moved compounds into Phase 1 and Phase 2 trials across oncology, neurology, and rare diseases. Additionally, AI has identified "Newly Predicted Cancer Genes" (NPCGs) that traditional mutation-based studies overlooked, significantly expanding the pool of therapeutic targets.

Challenges and Future Directions

Despite these advances, several hurdles remain:

  • The Diversity Gap: Genomic datasets are heavily skewed toward European populations, which may lead to biased models that fail to generalize across ethnicities.
  • The "Black Box" Problem: Model interpretability is a significant hurdle; biologists require transparency to trust predicted targets.
  • Data Bias: AI risks "historical momentum," where models focus on well-studied protein families due to data availability rather than exploring the "unknome."
  • Validation Requirements: While AI can predict and prioritize, experimental "wet-lab" validation remains essential to provide the ground truth labels necessary for iterative model improvement.

Summary updated

Abstract

Artificial intelligence (AI) holds immense promise for accelerating and improving all aspects of drug discovery, not least target discovery and validation. By integrating a diverse range of biological data modalities, AI enables the accurate prediction of drug target properties, ultimately illuminating biological mechanisms of disease and guiding drug discovery strategies. Despite the indisputable potential of AI in drug target discovery, there are many challenges and obstacles yet to be overcome, including dealing with data biases, model interpretability and generalisability, and the validation of predicted drug targets, to name a few. By exploring recent advancements in AI, this review showcases current applications of AI for drug target discovery and offers perspectives on the future of AI for the discovery and validation of drug targets, paving the way for the generation of novel and safer pharmaceuticals.
BibTeX
@inproceedings{wenteler2024ai,
  author = {Aaron Wenteler and Claudia Cabrera and Wei Wei and Victor Neduva and Michael R. Barnes},
  title = {{AI approaches for the discovery and validation of drug targets}},
  booktitle = {Cambridge},
  year = {2024},
  doi = {10.1017/pcm.2024.4},
}
Figure