Autolomous CEO Argues AI Must Link Shared Failure Data In Cell And Gene Therapy, After Sarepta Safety Setback

Cell and gene therapy companies are generating large amounts of data, but much of it still sits in separate silos. In an opinion article, Autolomous founder and CEO David Venables argues that this fragmentation is now a strategic problem for the sector, because safety events can reset risk perceptions across an entire modality rather than staying confined to one company.

Venables compares the field to aviation before the industry adopted the “black box” model of shared learning from failures. He points to the 2025 deaths from acute liver failure following treatment with Sarepta Therapeutics’ AAV-based gene therapies, and the interruption in Elevidys shipments and FDA investigation that followed, as an example of how one company’s setback can alter the risk conversation for every AAV program.

The Data Problem AI Is Supposed To Solve

Venables’ argument is not that AI replaces scientific judgment. He says intelligence “resides in humanity,” while modern computing provides the processing power needed to make sense of the volume of information CGT already produces.

He describes autologous therapies as especially data-intensive because each patient effectively generates their own “book” with hundreds of pages and thousands of data points. In his view, the field needs an integrated model that combines three areas of data capture: scientific process information, manufacturing information and post-treatment results. Today, he says, those are still treated as discrete silos.

That leaves CGT trying to optimize with tools borrowed from traditional pharma even though the underlying process is different. Venables argues that CGT requires live, momentary data, yet teams are still manually recording information on pen and paper or using 16 different kits to obtain a single data point. His conclusion is that AI is most useful as the connective layer between those data lakes, allowing treatment outcomes to feed back into process optimization and manufacturing decisions.

What The Piece Says Is Holding The Field Back

The article identifies three recurring mistakes. The first is a zero-sum approach to competition, where companies protect even non-IP lessons from failed work. Venables’ point is that mistakes are rarely patentable, but they are still the information peers need to avoid spending years on dead-end research.

The second is neglecting what he calls a “fourth leg” of data: genetic elements identifiable before disease manifests. He argues that oncology and post-birth rare disorders receive most of the attention, even though decades of digital information could help connect family history with disease outcomes.

The third is adopting off-the-shelf AI without a clearly defined operational purpose. Venables likens that to using a 24-wheel lorry to transport three people, arguing that companies should choose tools based on the specific optimization task instead of adding AI to workflows that were never designed for digital use.

The industry signal is straightforward: in this view, AI’s bottleneck in CGT is not model availability but data architecture and incentives. If companies keep treating failures as private and operations as disconnected, more computing power alone is unlikely to improve development efficiency or rebuild confidence after safety shocks.