AI indelibly transforms solid tumor drug development

Artificial intelligence (AI) is fundamentally reshaping the landscape of solid tumor (ST) drug development. By integrating multi-omics data, spatial transcriptomics, and advanced computational models, AI has dramatically accelerated the discovery and validation of novel therapeutic targets, compressing the traditional decade-long research and development cycle to just two to three years. Generative AI platforms have enabled the optimization of small molecule inhibitors, biologics, and messenger RNA (mRNA) vaccines, leading to breakthroughs in overcoming tumor heterogeneity, enhancing drug efficacy, and predicting resistance mechanisms. However, the clinical translation of AI-driven discoveries still faces significant challenges, including data bias, algorithmic transparency, and the validation gap between computational models and real-world human physiology. This review systematically examines the transformative impact of AI across the ST drug development continuum and advocates for interdisciplinary collaboration and ethical frameworks to realize the full potential of precision oncology.

AI-driven target discovery

Advances in single-cell and spatial omics

The integration of single-cell RNA sequencing (scRNA-seq) with AI has become a cornerstone for deciphering tumor heterogeneity. For instance, in pancreatic ductal adenocarcinoma (PDAC), spatial transcriptomics combined with deep learning models like SELFormer identified TRAILR1 as a key driver of immune escape, leading to the repurposing of mTOR inhibitors to enhance apoptosis. Similarly, scConGraph models revealed GDF15 as a chemoresistance factor, validated functionally as a therapeutic target. Spatial technologies such as Vistum and CODEX have further elucidated lactate metabolism gradients and immune checkpoint co-localization, informing combination therapies and immunotherapy response prediction with over 80% accuracy.

Targeting historically undruggable targets

AI is revolutionizing the targeting of once "undruggable" oncoproteins like KRAS and MYC. Through AlphaFold2 and reinforcement learning, allosteric inhibitors such as sotorasib and adagrasib were developed, with glecirasib recently approved for KRAS G12C-mutant cancers. AI also facilitates the design of proteolysis-targeting chimeras (PROTACs) and molecular glues. For example, PROTAC-RL optimizes degraders for BRD4, while quantum-classical hybrid models enable targeting of non-G12C KRAS mutants. KT-333, a STAT3-targeting PROTAC, shows promising clinical responses in early trials.

AI-optimized drug design

Small-molecule drugs

Generative AI platforms like PandaOmics and Chemistry42 have drastically accelerated hit identification and toxicity prediction. Novel inhibitors such as ISM5939 (targeting ENPP1) and ISM3091 (targeting USP1) were developed in under 30 months, with significantly reduced synthesis efforts. AI models also address drug resistance-e.g., designing fourth-generation EGFR inhibitors-and optimize pharmacokinetics through structure-based generative approaches, reducing design timelines to weeks.

Biologics

AI enhances antibody-drug conjugate (ADC) design by optimizing target selection, antibody humanization, and linker chemistry. Platforms like RADR® predict payload efficacy and patient-specific responses. Next-generation ADCs like Enhertu, optimized via AI, achieve higher drug-antibody ratios and expanded efficacy in HER2-low cancers, demonstrating pan-cancer potential.

mRNA vaccines

AI improves neoantigen prediction and mRNA vaccine design. Tools like PISTE achieve over 90% accuracy in TCR–antigen binding prediction, while LinearDesign optimizes mRNA stability and codon usage in minutes. Machine learning models further refine lipid nanoparticle (LNP) formulations for enhanced delivery and thermostability, advancing personalized cancer vaccines.

Challenges in clinical translation

Validation gaps and biases

The in vitro–in vivo gap is addressed by organ-on-a-chip systems like InSMAR-chip, which preserve tumor-immune interactions and predict clinical responses. Racial biases in genomic data are mitigated via adversarial debiasing and population-specific AI models. Drug resistance prediction is improved through dynamic models like DiSyn and JointSyn, which simulate clonal evolution and synergistic drug effects.

Data and algorithmic biases

Underrepresentation in training data and domain shifts between healthcare settings undermine model generalizability. Techniques such as federated domain adaptation and incremental learning (e.g., CODE-AE) help align models with diverse clinical environments and emerging therapeutic targets.

Future perspectives

In the short term (2–3 years), multimodal foundation models like MoLFormer and quantum-accelerated PROTAC design will enhance target degradation and immunotherapy response prediction. Long-term (5–10 years), AI-enabled closed-loop systems-integrating robotic biopsy, nanopore sequencing, and on-demand LNP production-could deliver personalized therapies within 72 hours. AI-guided chimeric antigen receptor macrophages (CAR-M) are also poised to reprogram the tumor microenvironment, offering new immunotherapeutic strategies.

Conclusion

AI has indelibly transformed ST drug development, from target discovery to clinical trial design. Yet, challenges in data equity, model interpretability, and clinical validation remain. Addressing these will require interdisciplinary collaboration, ethical oversight, and continuous learning systems. With emerging technologies such as quantum computing and multimodal AI, the future of oncology promises more personalized, effective, and equitable cancer care.