
Most AI models in biology are good at one kind of question: given this genetic change, will a disease appear? Far fewer can say why it happened or what to do about it. Researchers at Texas A&M University are working on that gap with a model called PertReason, which has now been accepted to the Conference on Neural Information Processing Systems (NeurIPS).
The problem PertReason targets
As described in the Texas A&M Engineering news story, today’s AI models can predict whether someone with a genetic change will develop a disease, but they cannot answer two critical questions: why the disease occurred, and how to design a treatment. The team, in the university’s Department of Electrical and Computer Engineering, is building AI virtual cell technology that predicts how genetic and drug effects change cellular behavior and also uncovers the biological mechanisms behind those changes.
Professor Yang Shen is involved in the project, supporting postdoctoral researcher Dongkwan Kim and Ph.D. student Yiming Gao, who are spearheading the initiative. Ph.D. student Yining Yang is assisting. The work is funded by the National Institute of General Medical Sciences.
How the approach works
Many traditional methods focus on individual biological components. According to the university, PertReason lets researchers trace how drugs and genetic changes affect cellular networks, so they can see a wider chain of cause and effect. That broader view is meant to help explain why a biological change leads to disease and to point toward better ways to counter it.
Shen says the group’s earlier work centered on fundamental molecules such as proteins, and the new work extends to the cellular level. In his words, the team is expanding the scale of what it studies. The project builds on AI models including large language models, with the aim of keeping them explainable and trustworthy while they reach accurate conclusions.
Checking the reasoning, not only the answer
One detail stands out. The team created a benchmark called PertReasonQA that tests whether a model’s reasoning is right, not just its final answer. It is designed to catch cases where a model gets the right answer without using the proper data to reach it. For scientific uses, that distinction is important, because a correct guess for the wrong reason can mislead the next experiment.
The group also trained an AI model to reason through biological pathways, which the university says helps it make more trustworthy predictions. An earlier version of the work was presented at a workshop on multimodal foundation models and large language models for life sciences during the International Conference on Machine Learning in Seoul in July.
Why researchers care
The team began collaborating early with clinical scientists working with breast cancer patients. They observed that some patients respond to treatment at first and then suddenly develop resistance, and they wanted to help address that rather than only study outcomes. Shen noted that good numerical metrics are one thing, but impact on people is another.
The practical promise is prioritization. By helping scientists narrow down which pathways to investigate first, the system could reduce trial and error and save time and resources. This is research news about a laboratory tool. The story does not describe any clinical use, and nothing here is medical advice.
What it means
For computational biology, PertReason is an example of a trend toward models that show their work. Benchmarks that grade reasoning, such as PertReasonQA, give reviewers and lab teams a way to judge whether to trust a prediction before they spend bench time on it.
For research groups outside this lab, the lesson is about evaluation. As language models enter life-science workflows, checking the path to an answer may matter as much as accuracy scores. A shared habit of testing reasoning could make AI-suggested experiments easier to defend.
What comes next
Shen and his team plan to keep expanding the model’s capabilities. With Texas A&M’s new VISION supercomputer recently arrived on campus, he intends to use it to push AI virtual cell models toward virtual replicas of the human body, known as digital twins, to inform medical decision making. His stated goal is multiscale digital twins for patients and personalized medicine, supporting clinicians and improving outcomes. That is a long-term vision, and the story gives no timeline.
Conclusion
PertReason is early-stage research, but its focus on explanation rather than prediction alone is a sensible direction. With a NeurIPS acceptance and a benchmark aimed at honest reasoning, the Texas A&M team has given the virtual cell idea a clearer way to be tested.