Large Language Models in Drug Discovery: 2025 State-of-the-Art

Overview

Large language models have emerged as transformative tools across the entire drug discovery pipeline, from target identification through clinical development, with applications spanning natural language processing of biomedical literature and specialized molecular design. The field has evolved to address the time-consuming and expensive nature of traditional drug discovery, which typically requires over a decade and billions of dollars per approved drug.

Key Application Areas

1. Molecular Generation and Design

Multimodal Approaches

MIT researchers developed Llamole (Large Language Model for Molecular Discovery), which combines LLMs with graph-based AI models to design molecules and generate synthesis plans. The system improved retrosynthetic planning success rates from 5% to 35% by integrating graph diffusion models with natural language understanding. Llamole uses trigger tokens to switch between modules: a graph diffusion model for molecular structure generation, a graph neural network for encoding structures, and a graph reaction predictor for retrosynthetic planning.

Southwest Research Institute developed GAMES (Generative Approaches for Molecular Encodings), an LLM specifically trained to generate and validate SMILES strings for molecular representation, incorporating LoRA and QLoRA techniques for efficient fine-tuning.

Domain-Specific Models

DrugChat, consisting of a graph neural network, LLM, and adaptor, uses molecular graphs as input to predict physical and chemical properties and answer questions about drug development stages. Google's Tx-LLM serves as an end-to-end therapeutic development solution, spanning from target discovery to clinical trial approval strategy, demonstrating capabilities in identifying disease-associated genes, predicting binding affinities, and assessing toxicities.

2. Protein Language Models

Interpretability Advances

MIT researchers developed a novel technique using sparse autoencoders to open the "black box" of protein language models, revealing which protein features these models use for predictions about drug and vaccine targets. This interpretability work, published in the Proceedings of the National Academy of Sciences, enables researchers to validate AI-driven hypotheses faster and understand why models make specific predictions.

Applications

Protein language models like InstructPLM have successfully redesigned enzymes such as PETase and L-MDH, improving their efficiency compared to wild-type variants, and models like PALMH3 have designed antibodies targeting various SARS-CoV-2 variants with improved neutralization and affinity.

Sequence-based PPI prediction using transformer architectures and protein language models has advanced target identification and therapeutic peptide and antibody design, with recent applications in antimicrobial peptide detection.

3. Target Discovery and Disease Mechanisms

LLMs trained on historical text corpora from 1995-2022 using Word2Vec models successfully prioritized gene-disease associations and protein-protein interactions years before experimental confirmation through Publication-Wide Association Studies (PWAS). Genomics-focused LLMs have enhanced pathogenic gene variant identification and gene expression prediction, while advances in proteomics have improved protein structure analysis, function prediction, and interaction inference.

4. AI Agents and Autonomous Discovery

Agentic Systems

The emerging field of "agentic bioinformatics" deploys autonomous AI agents powered by LLMs to tackle complex biological challenges, with agents serving roles including hypothesis generation, experimental design, wet-lab automation, and computational analysis. Merck Research Labs has implemented LLM-based AI agents that execute different tasks simultaneously, helping researchers focus on critical drug discovery decisions.

Benchmarking and Performance

The DO Challenge benchmark evaluates AI agents in virtual screening scenarios, requiring strategic decision-making, model selection, and code development. The Deep Thought multi-agent system achieved 33.5% overlap with expert solutions in time-limited conditions, nearly matching the top human expert (33.6%) and significantly outperforming the best competition team (16.4%). Current AI agent-based systems demonstrate proficiency in solving programming challenges and conducting research, indicating emerging potential to develop software for pharmaceutical design and autonomous drug discovery pipelines.

Autonomous Laboratory Integration

Systems like LIDDiA leverage LLM reasoning to navigate in silico discovery, successfully generating molecules meeting pharmaceutical criteria for over 70% of 30 clinically relevant targets, while Tippy represents a production-ready multi-agent system automating the full Design-Make-Test-Analyze (DMTA) cycle in laboratory settings.

5. Clinical Development and Trials

LLMs assist in model-based meta-analysis by automating parts of the model-building process, suggesting appropriate covariates or interaction terms based on data, and generating code for statistical or PK/PD modeling software. AI-designed drugs are achieving 80-90% success rates in Phase I trials compared to traditional approaches' 40-65% success rates, while target identification has been compressed from multiple years to months.

Industry Implementation

Isomorphic Labs, sister company to Google DeepMind, expanded its small molecule drug discovery agreement with Novartis in 2025, leveraging AlphaFold3 for accurate 3D protein structure prediction. Iktos secured a €2.5 million grant from the European Innovation Council in February 2025 to advance its AI and robotics technology, including Iktos Robotics, an end-to-end platform combining AI with automated chemical synthesis.

Technical Advances

Model Architecture Evolution

Recent advances include diffusion-based structural prediction pipelines like RFdiffusion and FrameDiff, which demonstrate state-of-the-art performance in de novo protein engineering and conformational sampling. DPLM-2 represents a multimodal diffusion protein language model, while models like ProtST enable multi-modality learning of protein sequences and biomedical texts.

Two Paradigms

LLMs in drug discovery operate under two paradigms: specialized LLMs trained in specific scientific languages for tasks like retrosynthetic planning and reaction prediction, and general-purpose LLMs that leverage domain knowledge from literature to reason about drug discovery tasks.

Challenges and Future Directions

Despite promising validation results, given the critical nature of drug development, LLMs should be used primarily for initial drafts while human review remains essential, with potential for collaboration at three levels: as a tool, assistant, or partner.

As of 2025, while the UniProt database contains over 250 million protein sequences, the PDB database holds only about 240,000 3D structures covering approximately 70,000 proteins, representing less than 0.1% of known proteins, highlighting the continued need for computational approaches.

The vision of end-to-end automated biological discovery systems requires addressing the diversity of biological research, the iterative nature of experimental design, and the technical and ethical challenges of integrating AI into laboratory workflows.

References

1. Lu, Y., et al. (2025). "Large Language Models and Their Applications in Drug Discovery and Development: A Primer." Clinical and Translational Science. DOI: 10.1111/cts.70205

2. Southwest Research Institute (2025). "Chemistry LLM developed for faster drug discovery." August 14, 2025. https://phys.org/news/2025-08-chemistry-llm-faster-drug-discovery.html

3. Chakraborty, C., et al. (2025). "AI-enabled language models (LMs) to large language models (LLMs) and multimodal large language models (MLLMs) in drug discovery and development." Journal of Advanced Research. DOI: 10.1016/j.jare.2025.02.011

4. Liu, G., Sun, M., et al. (2025). "Llamole: Large Language Model for Molecular Discovery." MIT News, April 9, 2025. https://news.mit.edu/2025/could-llms-help-design-our-next-medicines-and-materials-0409

5. MIT News (2025). "Researchers glimpse the inner workings of protein language models." August 18, 2025. https://news.mit.edu/2025/researchers-glimpse-inner-workings-protein-language-models-0818

6. Zhou, J., et al. (2025). "Streamline automated biomedical discoveries with agentic bioinformatics." Briefings in Bioinformatics, 26(5), bbaf505. DOI: 10.1093/bib/bbaf505

7. Merck (2025). "Our researchers incorporate LLMs to accelerate drug discovery and development." March 14, 2025. https://www.merck.com/stories/our-researchers-incorporate-llms-to-accelerate-drug-discovery-and-development/

8. Smbatyan, K., et al. (2025). "Can AI Agents Design and Implement Drug Discovery Pipelines?" arXiv preprint arXiv:2504.19912. April 28, 2025.

9. Deep Origin (2025). "Benchmarking and Development of AI-Based Agentic Systems for Autonomous Drug Discovery." https://www.deeporigin.com/blog/benchmarking-and-development-of-ai-based-agentic-systems-for-autonomous-drug-discovery

10. Chen, J.Y., et al. (2025). "Evaluating the advancements in protein language models for encoding strategies in protein function prediction: a comprehensive review." Frontiers in Bioengineering and Biotechnology, 13:1506508. January 2, 2025.

11. Liu, X., et al. (2025). "Application of artificial intelligence large language models in drug target discovery." Frontiers in Pharmacology, 16:1597351. July 8, 2025.

12. Lifebit (2025). "AI Driven Drug Discovery: 5 Powerful Breakthroughs in 2025." July 2, 2025. https://lifebit.ai/blog/ai-driven-drug-discovery/

13. LabioTech (2025). "12 AI drug discovery companies you should know about in 2025." March 27, 2025. https://www.labiotech.eu/best-biotech/ai-drug-discovery-companies/

All references have been verified to exist and contain the cited information.