AI tools bridging preclinical to FIH are now a necessity, not a nice to have

Main video credit: TAW4/Shutterstock.com

A​​​​​​​s artificial intelligence (AI) becomes more established in clinical development, one area showing particular promise is the transition between preclinical studies and first-in-human (FIH).

While preclinical trials are evolving and can give sponsors a good indication of how their drug will behave, there are still instances where outcomes can differ significantly once it is tested in humans.

Given the high cost of drug development, the ability to derisk a candidate or identify ways to improve an asset will be an invaluable tool for both pharma and biotech.

There are already companies in the space working to better predict toxicity and the likelihood of clinical trial success before a drug is even dosed in humans, helping companies decide before enrolling patients whether a candidate is viable.

Even “Acts of God” reveal limitations in supply chain agility. A case in point was the severe shortage of IV fluids in the US after damage to Baxter’s North Carolina-based facility during Hurricane Helene in September 2024.

With market dynamics in constant flux, Pharmaceutical Technology spoke with Stefan Verheyden, CEO of Sanner, a German-based CDMO for pharmaceutical and medical device packaging, to discuss how the healthcare industry is built to withstand disruption.

The interview was conducted ahead of Pharmapack Europe, a pharma packaging conference held from 21 to 22 January in Paris.

This interview has been edited for length and clarity.

Paul Grady, CEO of Crucial Data Solutions, argues that simulated, AI‑driven trials can give sponsors a fairly good idea of how a trial will evolve, before enrolling a single human.

“We’re not saying that [you should] use AI to replace biological participants… but what it really does is it allows you, before spending a billion dollars on your clinical trial, to get a pretty good idea of how your trial is going to go, and… do that in a matter of hours.”

The need for reform is echoed by Theodore Alexandrov, CEO of DeepCyte, which utilises AI within mass spectrometry to evaluate single-cell-level toxicity concerns, particularly around animal models, pointing out that even regulators are ‘screaming for reformation’ in toxicity testing.

Alexandrov believes that toxicity is a core problem in drug development, on par with efficacy, aiming to cut costly late‑stage failures and reduce reliance on animal testing.

DeepCyte’s approach aligns with global regulatory momentum, including the US Food and Drug Administration (FDA) Modernization Act 2.0 and the European Medicine Agencies (EMAs) efforts to push non‑animal “new approach methodologies” (NAMS) in safety testing.

The single‑cell mass spectrometry technology “listens” to what cells are actually doing in real time, offering mechanistic insight rather than just black‑box toxicity predictions. The cells are hit with a laser, causing molecules to “fly off” and be measured with a mass spectrometer.

This allows the company to focus on metabolomics and determine what the cell is doing at that exact moment, rather than just a blueprint or machinery (DNA/RNA/proteins).

Alexandrov says: “If you only look at late‑stage markers, it’s like checking the foundations when your house is already collapsing. It tells you to run, but it doesn’t tell you that termites were the real problem — and by then it’s too late to do anything about it.”

When embarking on a clinical journey, companies have to consider the millions of dollars it will take to get a trial through multiple phases, which could lead to a flawed trial design causing failure, says Grady.

Crucial Data Solutions’ AI “Floyd”, which utilises Gemini, can simulate 10-year clinical trials in hours, generating synthetic participants and longitudinal datasets before the first human is enrolled. Floyd is able to predict both efficacy and adverse events (AEs), with the AE data having closely mirrored real‑world outcomes.

Grady says that the system is not intended to replace biological participants, but to be utilised after preclinical investigations and before FIH as a risk-based approach. It can provide data to help sponsors refine biostatistics plans, optimise back-end analyses and improve overall trial design and operational efficiency.

As a result, it could be particularly helpful for early‑stage biotechs to prioritise their most promising preclinical assets, project endpoints and identify weak spots in trial plans while resources are still scarce.

In a funding environment that has been volatile since 2022, investors are increasingly demanding deeper scientific validation before committing capital. While there has been some upswing in funding activity in 2026, many companies are still struggling to get investment in very early stage therapies due to a lack of data, making it near impossible for them to gain enough capital to initiate FIH studies.

According to Karolina Makovskyte, symptoms project manager at Vugene, the mechanistic understanding of a therapeutic candidate is no longer a “nice to have” - it is becoming a prerequisite, and this is somewhere AI can help.

“For investors, having a mechanistic understanding of what a drug does to the system at quite an early stage is very crucial in order for them to have a more de-risked, let's say, pipeline of portfolio of companies and assets, so we really see it as a crucial thing"

Vugene’, AI‑driven multi-omics are being used as a form of proof-of-concept. By elucidating the mechanism of action at a comparatively early stage - often between preclinical work and FIH trials - companies can present a more de‑risked story to investors.

For investors, this level of mechanistic clarity supports better portfolio decisions; for companies, it becomes a critical asset in competitive fundraising processes.

As AI becomes more embedded in R&D, the emphasis is shifting from mere automation to explainability and trust.

While trust in AI is improving, it does continue to be a real barrier for AI implementation. As a result, AI must be not just predictive but explainable, Grady says.

Alexandrov agrees, stating that explainability is deeply mechanistic, and that both regulators and biologists don’t just want a ‘toxic / not toxic’ label but insight into what’s going wrong in the cell.

“There is demand for not just doing prediction but also explaining mechanisms. ‘Hey, it should be backed with evidence… explain what this prediction is based on? What are specific molecular problems that the drug triggers?’ Because we know that we can better generalise to a human context,” Alexandrov explains.

Makovskyte adds that AI and its explainability are vital; however, she highlights that Vugene is deliberately avoiding a “fully automated black box.” Instead, it will combine agentic AI capabilities to synthesise, compare, and interpret complex datasets, including cross‑project insights alongside its team of experienced bioinformaticians who review and validate outputs, ensuring that the biological narratives are robust and meaningful.

Grady sees a lot of theory in the industry, such as digital twins, but is seeing very little evidence of practical deployment overall. As a result, Grady is sceptical that simply duplicating data via digital twins is necessary, arguing that many proponents underestimate AI’s reasoning capabilities; “you don’t always need a “twin” if the model already reasons over rich data”, Grady says.

Large tech companies, such as NVIDIA, will likely be instrumental if the industry moves toward synthetic participants as a replacement, and Crucial Data Solutions expects to be a small part of that broader ecosystem, Grady believes. ​​​​​​​

Despite AI often being discussed as a technology which is “on the horizon,” all experts agree that the paradigm shift to significant utilisation of AI is already underway.

“Although we’ve been talking about this sort of technology, and many companies talk about it as if it will be available in five years, I’m just saying it’s available today. Do with it as you please,” Grady concludes.

A growing number of companies, both large and small, are building internal AI models and actively seeking ways to integrate AI into existing research and development workflows.

Crucially, this shift is not just about new algorithms; it’s about maximising the value of existing data. Many organisations already sit on large volumes of experimental and clinical data that are underutilised.

When companies are planning how to move into clinical-stage development, better utilisation of this data will save both time and cost, something that, in this difficult funding environment, will be incredibly attractive to investors, proving that AI in this time period is just as important, in some cases even more so, than its utilisation in clinical trials. ​​​​​​​