Dr. Jo Varshney believes the next competitive advantage in AI-driven drug discovery will not be better predictions, but better answers to a much harder question: can we explain why a drug should work in humans?
Artificial intelligence is rapidly changing how drugs are discovered, but Dr. Jo Varshney believes the industry’s biggest challenge is no longer generating possibilities. It is deciding which ones deserve to move forward. As CEO and Founder of VeriSIM Life, Varshney works at the intersection of AI, computational biology, and drug development, using modelling to understand how medicines might behave in the human body before costly clinical trials begin. Her argument is deceptively simple: as AI makes prediction easier, scientific explainability becomes more important, not less. The real advantage, she says, will belong to platforms that can connect an AI prediction to biology, evidence, uncertainty, and ultimately, human outcomes.
As drug discovery goes AI-native across the US, Europe and Asia, does explainability become a genuine competitive edge, or just a compliance checkbox most platforms will eventually tick?
Explainability will become expected. The real differentiator will be whether you can explain the science, not just the AI.
It is relatively easy to show why a model produced a particular score or prediction. The harder question in drug development is: why should we believe this will work in humans?
That requires connecting a prediction back to biology, mechanism, exposure, safety, and the experimental evidence behind it. It also means being clear about where the uncertainty is and what needs to be tested next.
As more of drug discovery becomes AI native, generating predictions will not be the bottleneck. Deciding which predictions are worth acting on will be.
So yes, explainable AI may eventually become a compliance checkbox. But scientific explainability that helps teams make better translational decisions will remain a competitive advantage.
Regulators in Asia are approaching AI-driven drug development quite differently from the FDA. Does BIOiSIM’s approach travel across those regulatory philosophies, and does explainability matter more in some markets than others?
Yes, because BIOiSIM is not built around one regulator’s framework. It is built around the underlying scientific questions regulators care about everywhere: how a drug behaves, what evidence supports that conclusion, where the uncertainty is, and whether the result can be reproduced and validated.
The emphasis may differ by market. Some regulators may focus more heavily on model transparency and traceability, while others may put more weight on validation against experimental or clinical data. But the core requirement is similar: you need to show why the output should be trusted.
That is where explainability matters. Not as a feature added for regulatory review, but as part of how the platform works. BIOiSIM connects predictions back to mechanism, pharmacology, and supporting data so teams can understand what is driving a result and test it.
That makes the approach portable across regulatory environments, even as the specific expectations continue to evolve.
PT001 has developed largely within a US clinical framework. What would it take for a platform like BIOiSIM to confidently model human translation for populations with different genetics, environments, or healthcare infrastructure, say in India or Southeast Asia?
The core biology does not change by geography, but the inputs can.
For BIOiSIM to model translation in India, Southeast Asia, or any other population, we would incorporate the factors that could materially change drug response, including genetics, physiology, disease characteristics, comorbidities, diet, environmental exposures, and standards of care.
The important part is not assuming that a model developed using one population automatically generalizes to another. You need to identify which variables are likely to matter, test their impact computationally, and validate against local clinical or real world data where available.
For PT001, that could mean asking whether differences in patient characteristics change lung exposure, systemic exposure, efficacy, or safety and then determining whether those differences are large enough to affect dose or trial design.
Healthcare infrastructure is a slightly different question. It may not change the pharmacology of the drug, but it can absolutely change how a therapy needs to be developed and delivered.
That is where I see computational approaches becoming especially useful: not creating a different model for every country, but understanding which differences actually matter for translation and which do not.
You’ve framed the real challenge as knowing which molecules will work in humans, not generating them. As more countries build their own AI drug discovery ecosystems, is that translation gap universal, or does it look different depending on where a candidate is developed?
The translation gap is universal. Biology does not become easier because a molecule was discovered in Boston, Shanghai, or Bangalore.
What can differ is the data available, the patient populations being studied, how preclinical studies are designed, and the clinical and regulatory path used to develop the drug. Those differences can change where uncertainty enters the program, but the fundamental question is the same: will what we see in a model or an animal actually hold up in humans?
That is why I think the next phase of AI drug discovery will be less about who can generate the most molecules and more about who can eliminate the wrong ones earlier.
If we can understand why a candidate should translate, where it might fail, and what evidence we need before moving forward, we can make drug development much more efficient regardless of where that molecule originated.
If explainable AI becomes the industry standard for predicting clinical success, what happens to smaller biotech markets outside the US that don’t have the same depth of clinical trial data BIOiSIM likely draws on?
I actually think this is where explainable approaches can be most useful. You should not need an enormous local clinical dataset before you can make a meaningful translational assessment.
BIOiSIM does not rely on clinical trial data alone. We bring together different types of evidence, including mechanistic biology, pharmacology, preclinical and clinical data, and known physiological differences across populations. The goal is to use what is known to understand what can reasonably translate, while being very clear about what remains uncertain.
For smaller biotech markets, that can help identify which local data are actually important to generate rather than trying to recreate the depth of data available in the US.
The advantage of explainability is that you can see where the evidence is strong, where assumptions are being made, and where additional validation is needed. That makes limited data a problem you can work through, rather than simply a reason the model cannot be used.
