Sheng Wang’s group assigns GPT-6 Astra tasks that take hours; drug discovery awaits laboratory measurements

Sheng Wang’s group assigns GPT-6 Astra tasks that take hours; drug discovery awaits laboratory measurements

On October 19, AI and computational biology researcher Sheng Wang told the Silicon Valley 101 podcast that his group assigns GPT-6 Astra tasks lasting 8–24 hours and uses its computer-control capabilities to operate older, specialized biomedical software. In drug discovery, the next decision depends on results from cell and animal experiments that the laboratory must carry out.

According to Sheng Wang, AI can already outperform humans on tasks with predefined inputs, a test, and a correct answer. In his group, GPT-6 Astra handles lengthy computational tasks and operates specialized software. Drug discovery requires a different cycle: the laboratory tests how a selected candidate behaves in cells or an organism, then supplies the measurement to inform the next decision.

To design such an experiment, the model needs to account for the laboratory’s specific conditions: materials, equipment, temperature, humidity, and budget. Sheng Wang describes the task this way:

“For us, one of the main use cases is whether the model can plan an experiment: how to conduct it, at what temperature, with which molecule, at what humidity, and under what conditions.”

The laboratory conducts the experiment and obtains a measurement; the model uses these data to select the next hypothesis or candidate. Cell and mouse experiments take several weeks, while testing in primates takes at least four months. This schedule sets the pace of drug discovery, even when the computational work takes hours.

Adaptyv Bio has integrated laboratory testing into its Rosalind Workbench protein design service so that binding test results feed into the next design round. This exchange of data allows the model to choose the next experiment based on the results of the previous one.