Yu-Xiu Lin Ph.D.(c) & Thai-Yen Ling, Ph.D.
Department and Graduate Institute of Pharmacology, College of Medicine, National Taiwan University, Taipei, Taiwan

Introduction

Over the past several years, generative artificial intelligence has primarily been used to produce text, images, and computer code in response to user instructions. As large language models gain stronger reasoning and tool-use capabilities, AI is evolving from a passive chatbot into an AI agent capable of receiving a goal, breaking it into smaller tasks, retrieving information, operating external software, evaluating results, and adjusting its strategy based on feedback. An AI agent is therefore more than an advanced language model. A complete agentic system also requires memory, workflow orchestration, external tools, and a secure execution environment. [1].

From Digital Assistants to Scientific Partners

AI agents are beginning to move beyond literature summarization and administrative support. A representative example is Robin, a multi-agent scientific discovery system reported in Nature. Robin integrates literature retrieval, hypothesis generation, experimental planning, and data analysis into an iterative research cycle. Human researchers conduct the physical experiments and return the results to the system, which interprets the data and proposes updated hypotheses [2].

Nevertheless, current AI scientists should not be viewed as autonomous replacements for human researchers. Biological research involves tacit knowledge, undocumented experimental failures, biological variability, and practical limitations that may not be represented in training data.

AI may also generate incorrect references, overlook data bias, or propose experiments that are theoretically plausible but difficult to execute. A more realistic role for AI is therefore that of a co-scientist: AI can continuously search, calculate, simulate, and generate candidate solutions, while human researchers determine their biological significance and decide which experiments should be performed.

Nemotron and BioNeMo: Reasoning With Biological Tools

Within NVIDIA’s architecture, Nemotron provides the reasoning foundation that enables an agent to understand objectives, plan tasks, and select appropriate tools. Nemotron is not a complete agent by itself; it operates with orchestration frameworks and secure execution environments that manage memory, tool access, permissions, and workflow control [3].

General-purpose language models can understand biomedical literature but do not necessarily know how to correctly perform protein-structure prediction, molecular docking, genomic analysis, or therapeutic design. The BioNeMo Agent Toolkit addresses this limitation by packaging biological and chemical models, libraries, and computational workflows into tools that AI agents can call [4]. In this framework, the reasoning model determines what should be done, while BioNeMo provides the domain-specific tools needed to perform the scientific computation. These tools include protein-structure prediction, molecular docking, generative chemistry, genomic analysis, protein design, biomarker discovery, and virtual screening.

The main advance is not simply faster computation. BioNeMo can connect tools that researchers previously operated separately across different software packages and data formats. An agent may gather evidence, prepare model inputs, run computational experiments, compare outputs, and recommend the next analytical step within a repeatable workflow. However, these outputs remain computational hypotheses and must still be confirmed through biochemical assays, cell experiments, animal studies, and clinical validation.

Digital Biology and Drug Development

Digital biology converts DNA, RNA, proteins, cellular images, single-cell omics, spatial data, and clinical phenotypes into computational representations that AI models can learn from and simulate. Earlier biological models were usually designed for individual tasks, such as predicting protein structures or classifying cell types. New biological foundation models attempt to learn broader biological patterns and predict gene perturbations, drug responses, and changes in cellular states. This development has contributed to the concept of the AI virtual cell [5,6].

Virtual-cell models may help researchers simulate how cells respond to genetic modification, environmental changes, or drug exposure. However, they remain incomplete representations of real cells. Cellular behavior is influenced by culture medium, extracellular matrices, cell density, mechanical signals, oxygen levels, cell–cell interactions, and time—factors that are often missing or incompletely recorded in existing datasets. Virtual cells should therefore be regarded as tools for generating hypotheses and narrowing the experimental search space rather than as substitutes for biological experiments. AI agents could simulate gene knockouts or drug treatments, identify the most informative conditions, and recommend validation using CRISPR assays, organoids, or cell cultures. Experimental results could then be returned to the model, forming a closed design–build–test–learn cycle [7].

In early drug discovery, agents could integrate biomedical literature, disease databases, genomic evidence, protein-interaction networks, and single-cell data to identify potential therapeutic targets and biomarkers. Candidate targets could then be passed to structural-analysis and virtual-screening agents. During compound design, agents could combine protein-structure prediction, molecular generation, docking, molecular dynamics, and ADMET models. Rather than relying on one model to generate a “perfect drug,” multiple specialized tools could iteratively optimize potency, selectivity, solubility, toxicity, and synthetic feasibility.

In preclinical and clinical development, agents may also assist with toxicology review, dose–response analysis, biomarker identification, patient stratification, clinical-trial matching, and statistical programming. When AI-generated evidence is used to support decisions regarding the safety, effectiveness, or quality of a drug or biological product, however, its context of use must be clearly defined. The Good AI Practice principles jointly developed by the U.S. Food and Drug Administration and the European Medicines Agency emphasize human-centered design, risk-based assessment, data governance, performance evaluation, documentation, and life-cycle management for AI used in drug development [8].

The Special Value of AI Agents in Cell Therapy

Unlike small-molecule drugs with relatively stable chemical structures, cell-therapy products are living populations that change over time and in response to their environment. Donor or patient variability, starting-cell characteristics, culture conditions, cell density, genetic modification, metabolic state, and harvest timing can all influence the final product. Cell therapy is therefore particularly suitable for multimodal AI systems capable of integrating cellular images, environmental parameters, metabolite concentrations, flow-cytometry results, gene-expression profiles, and potency assays.

The first potential application is cell and product design. AI can integrate single-cell transcriptomics, protein expression, and functional assays to identify cellular states associated with therapeutic activity, persistence, differentiation capacity, immunomodulation, or exhaustion. In CAR-T development, AI-guided design has already been experimentally explored to improve CAR durability, reduce antigen escape, and optimize antitumor function [9].

The second application is process development and parameter optimization. An agent could analyze relationships between critical process parameters and critical quality attributes throughout cell culture. Researchers have used live-cell bright-field imaging and machine learning to recognize cellular states during pluripotent stem-cell differentiation, predict outcomes, and adjust culture conditions to reduce variability between cell lines and production batches [10]

The third application is non-destructive, real-time quality monitoring. Conventional cell-product testing often requires sampling, staining, or destruction of cells, while functional potency assays may take several days. Image-based AI can instead examine the morphology, arrangement, movement, and heterogeneity of living cells [11].

In the future, an AI agent could continuously monitor an entire culture process rather than analyze one image at a time. When abnormal morphology, confluence, or growth kinetics are detected, the agent could integrate manufacturing records, environmental parameters, metabolic measurements, and imaging data before recommending additional sampling or deviation investigation.

In a GMP environment, the data used by the agent, model version, recommendations, and subsequent actions would need to remain traceable. Process changes and final batch-release decisions should remain under the responsibility of qualified personnel.

From N1X to Physical AI and Automated Cell Factories

Before its formal launch, NVIDIA’s Arm-based AI PC platform was widely discussed in technology media under the name N1X. NVIDIA officially introduced the platform as the RTX Spark Superchip, which combines an Arm-based CPU, a Blackwell GPU, and unified memory to support local AI agents and large models [12].

For biomedical research, local agent computing could potentially allow sensitive experimental or patient data to be analyzed near laboratory instruments rather than being transferred entirely to the cloud. Large-scale model training and molecular screening would still require data-center infrastructure, but selected image-analysis, monitoring, and agent workflows could run locally. While N1X or RTX Spark brings agents closer to researchers and instruments, Cosmos 3 extends AI into the physical environment. Cosmos 3 combines visual reasoning, world generation, and action prediction, allowing it to function as a vision-language model, world model, or foundation for robotic control [13].

Its role is not to analyze DNA or proteins directly. Instead, it can help robots understand physical environments, predict the consequences of actions, and practice tasks in simulation before transferring those skills to real equipment. In an automated cell-therapy facility, Physical AI could support the transport of culture vessels, tubing connections, sampling, centrifugation, medium exchange, microscopy, and cold-chain logistics. Because trial-and-error learning in sterile or GMP environments is costly and risky, developers can first create a laboratory digital twin in which equipment, robotic arms, consumables, and workflow layouts are simulated.

Multiply Labs has used NVIDIA Omniverse to construct digital twins of cell-therapy manufacturing environments and Isaac Sim to test robotic laboratory operations through thousands of virtual iterations before real-world deployment.

Developments in Taiwan

Publicly documented NVIDIA collaborations in Taiwan currently focus more on clinical agents, smart hospitals, and medical robotics than on the direct adoption of BioNeMo for drug discovery. Foxconn has integrated NVIDIA technologies including Nemotron, NemoClaw, Omniverse, and Isaac into its healthcare platforms.

Its CoDoctor and CoDoClaw systems coordinate specialized clinical agents for applications such as breast imaging, electrocardiography, retinal imaging, and coronary-artery analysis [14]. Foxconn’s Nurabot nursing collaborative robot has also completed field validation and is being introduced into hospitals and long-term care settings.

Nurabot supports repetitive clinical tasks such as medication and specimen transport, demonstrating how Physical AI can be integrated into real healthcare workflows [15]. Compal Electronics is developing the POLYMEDX smart-hospital platform with NVIDIA technologies. POLYMEDX integrates AI orchestration, digital twins, edge AI, Omniverse, Isaac ROS, and robotic systems to simulate, validate, and deploy hospital workflows [16].

These projects are not yet equivalent to BioNeMo-based drug discovery or automated cell-therapy manufacturing. Nevertheless, they demonstrate Taiwan’s growing ability to integrate AI agents, simulation, robotics, and healthcare operations. The same infrastructure could eventually be extended to laboratory automation and advanced biomanufacturing.

Conclusion

The greatest impact of AI agents on drug development and cell therapy may not come from accelerating one isolated analysis. Their broader value lies in connecting literature review, computational modeling, experimental design, cell culture, quality monitoring, and robotic operations into a continuously learning workflow. A future research environment may involve human scientists defining research questions and acceptable risk boundaries, AI agents organizing information and planning workflows, and Physical AI systems performing selected laboratory operations. In medicine and cell therapy, the most valuable AI agent will not be an uncontrolled autonomous scientist, but a traceable and reliable digital research partner whose evidence, actions, and limitations can be reviewed by human experts.

References

• NVIDIA. (2026, June 1). NVIDIA GTC Taipei 2026 keynote. YouTube.
• Ghareeb, A. E., Chang, B., Mitchener, L., et al. (2026). A multi-agent system for automating scientific discovery. Nature, 655, 497–505.
• NVIDIA Research. (2026, June 4). NVIDIA Nemotron 3 Ultra.
• NVIDIA. (2026d, June 23). NVIDIA announces BioNeMo Agent Toolkit—Tools for agents to accelerate scientific discovery.
• Qian L. et al. Grow AI virtual cells: three data pillars and closed-loop learning. Cell Research, 2025.
• Ma, C., Zhang, H., Rao, Y., et al. (2026). AI-driven virtual cell models in preclinical research: Technical pathways, validation mechanisms, and clinical translation potential. npj Digital Medicine, 9, Article 25.
• National Academies of Sciences, Engineering, and Medicine. (2025). The age of AI in the life sciences: benefits and biosecurity considerations.
• U.S. Food and Drug Administration. Guiding Principles of Good AI Practice in Drug Development, 2026.
• Ansari, M. S., Chauhan, V., Singh, A., et al. (2026). AI-guided CAR designs and targeted pathway modulation to enhance multi-antigen CAR T cell durability and overcome antigen escape. Nature Communications, 17, 1549.
• Yang, X., Chen, D., Sun, Q., et al. (2023). A live-cell image-based machine learning strategy for reducing variability in PSC differentiation systems. Cell Discovery, 9,
• Imboden, S., Liu, X., Lee, B. S., et al. (2021). Investigating heterogeneities of live mesenchymal stromal cells using AI-based label-free imaging. Scientific Reports, 11, 6728.
• Kampman, J. (2026, May 29). Nvidia and Microsoft tease “a new era of PC” ahead of Computex 2026. Tom’s Hardware.
• NVIDIA. (2026e, May 31). NVIDIA launches Cosmos 3, the open frontier foundation model for Physical AI.
• NVIDIA. AI’s Next Revolution: Multiply Labs Is Scaling Robotics-Driven Cell Therapy Biomanufacturing Labs, 2026.
• Hon Hai Technology Group. (2026, June 1). Foxconn presents its Physical AI healthcare strategy at NVIDIA GTC Taipei COMPUTEX 2026.
• Compal Electronics. (2026, June 2). Compal collaborates with NVIDIA to deepen its Physical AI ecosystem and accelerate the deployment of the POLYMEDX smart-hospital platform.