Biomedical Technology · global
Virtual Cells Need Real Experiments: AI Drug Discovery Competition Shifts Toward Data Quality
Organoid chips provide biological responses, while virtual cells attempt to learn from them. An industry forum in Shanghai outlined a complementary research and development direction for the two, but consistency across batches of data and validation of predictions for specific uses remain hurdles to practical implementation.
How a drug will act after entering the human body is one of the hardest questions to answer in advance during new drug development. AI virtual cells and organoid chips seek answers through computational models and biological experiments, respectively. A recent industry forum in Shanghai suggested that the two approaches could complement each other. For this collaboration to genuinely help screen drugs, the key questions are whether the experimental data are reliable and which questions the models can actually answer.
According to a September 23 report by Cailian Press, the “Building a Virtual Human” roundtable at the Jefferies Asia Pacific Healthcare Summit discussed the respective roles of the two technologies. Information released before the event by the Shanghai Comprehensive Service Platform for Enterprises Going Global listed the inaugural summit as scheduled for September 20–21 in Zhangjiang Science City, with artificial intelligence and organoid-driven drug discovery both on the agenda.
The idea of complementarity stems from their different ways of working: virtual cells use models to estimate cellular responses, while organoid chips provide measurable biological experimental results. XtalPi co-founder Lipeng Lai said at the forum that the value of organoid chips lies in supplying data highly relevant to biological questions, potentially supporting continued model improvement. This was a research and development direction proposed by participants; the report did not include a performance comparison following the combination of the two technologies.
Whether the data can be compared is a practical challenge for this approach. Lai noted that maintaining consistency across different experimental batches and instruments remains unresolved. If measurement differences become mixed into training data, models may struggle to distinguish changes arising from biological responses from those caused simply by experimental conditions. Increasing the volume of data therefore cannot replace quality control.
The intended use also determines the validation threshold. Participants emphasized that whether animal experiments can be reduced or replaced must be assessed for specific scenarios. Even if a model performs well in one test, this does not directly establish that it can address other safety questions. The report did not provide model accuracy figures, independent validation data, or evidence that a specific platform had been accepted by a regulatory authority, leaving insufficient grounds to determine its ability to replace animal experiments.
In the early stages of commercialization, companies may begin by offering research and development services. Jiajie Wu of 耀速科技 said that technical services and joint development can help startups understand what kinds of models the industry needs through actual drug development work. This approach also gives data requirements a specific purpose, although the forum report did not disclose related contract values or measurable research and development outcomes.
What emerged from this discussion was a more specific set of questions for the next stage of AI drug discovery: whether experiments can be reproduced consistently and whether models can withstand validation for clearly defined biological questions. The proposed complementarity between virtual cells and organoid chips offers one path forward. How far it can go still needs to be answered by comparable data and actual drug development results.