Why In Vivo Therapies Fail When We Treat Them Like Ex Vivo Ones

By Wei Zhu, independent strategy and architecture advisor, cell and gene therapy

As in vivo cell and gene therapies gain momentum, the industry has leaned heavily on comparisons to established ex vivo models. The logic is intuitive: if ex vivo CAR-T and gene-modified cell therapies succeeded by exerting tight manufacturing and process control, then in vivo approaches should follow a similar trajectory, just with different delivery vehicles. But this assumption is increasingly proving fragile.

In vivo therapies are not a manufacturing shortcut, nor a linear evolution of ex vivo platforms. They represent a fundamentally different operating environment, one in which biological systems, not manufacturing processes, carry the burden of control. When we fail to acknowledge that shift, we introduce a class of risks that cannot be mitigated by better vectors, higher doses, or incremental process optimization.

The Illusion Of Control In Ex Vivo Thinking

Ex vivo paradigms offer a comforting sense of determinism. Cells are collected, engineered, expanded, tested, and released under tightly defined conditions. Variability is addressed upstream, and uncertainty is minimized before the therapy ever enters the patient.

In low-entropy disease settings, such as certain hematologic malignancies, this approach has worked remarkably well. However, when this mental model is applied to in vivo therapies, something subtle but critical happens: uncertainty does not disappear, it simply shifts downstream.

Manufacturing control may still be excellent. Vector design may be elegant. But once the therapy is released into the body, it enters a biological system defined by heterogeneity, feedback loops, spatial variation, immune surveillance, and time-dependent state transitions.

In vivo therapies are not executed on biology, they are executed by biology, and that distinction changes everything.

Why In Vivo Is A High-Entropy Operating Environment

In vivo therapeutic systems operate inside environments characterized by:

• spatial heterogeneity (tumor microenvironments, tissue gradients, immune niches)
• temporal variability (disease progression, immune activation cycles, inflammation)
• state-dependent responsiveness (cell exhaustion, activation thresholds, epigenetic memory)
• irreversibility (once certain biological trajectories are entered, reversal is costly or impossible).

In such environments, the question is no longer: “Can we execute this intervention precisely?” It becomes: “Can we govern when, where, and under what conditions intervention is allowed to occur?” This is not an execution problem, but instead a decision-architecture problem.

The Failure Pattern We Keep Mislabeling

Across in vivo CAR-T, gene therapy, and immune-modulating platforms, a recurring pattern appears in clinical outcomes:

• Initial biological activity is observed.
• Partial efficacy is achieved.
• Unintended toxicity, loss of persistence, or relapse follows.
• The system stabilizes into a suboptimal, often irreversible state.

These outcomes are frequently attributed to:

• weak vectors
• insufficient potency
• poor target selection
• manufacturing variability.

But in many cases, these explanations describe what failed, not why the system failed in the way it did. What is often missing is a framework for understanding how therapeutic actions interact with biological state transitions over time.

Execution Without Governance Creates Trajectory Risk

In ex vivo systems, guardrails are implicit. Decisions are embedded in manufacturing gates, release criteria, and process controls. In in vivo systems, those gates largely disappear. When activation, expansion, differentiation, or payload expression occurs without adequate sensing and permission, therapies can push biological systems into trajectories that look promising early, but later become traps.

Examples include:

• immune activation that overshoots into toxicity
• tumor engagement that accelerates exhaustion
• gene expression that stabilizes maladaptive cellular states
• partial remodeling that locks the system into brittle equilibria.

These are not failures of execution quality. They are failures of when-to-act logic.

Why Better Tools Alone Don’t Fix This

A common response to in vivo failure is to add more capability:

• Higher doses
• More potent constructs
• Additional genetic modifications
• Combination regimens

While these tools can increase short-term efficacy, they often amplify long-term instability. Without mechanisms to sense biological state, gate intervention, and stabilize outcomes, more powerful tools simply drive the system faster, sometimes into failure modes that are harder to unwind. In high-entropy systems, capability without operational rigor increases risk.

Reframing The Problem: From Modality To Architecture

The industry often debates in vivo therapies at the modality level:

• Viral vs. non-viral
• Ex vivo vs. in vivo
• Single-shot vs. repeat dosing

But these comparisons obscure the deeper issue. The real question is not which modality is superior but: “Is the therapeutic system designed to manage biological state transitions over time?” This requires shifting from a tool-centric mindset to an architecture-centric one, where sensing, decision logic, orchestration, and stabilization are treated as first-class design elements.

When therapies are designed as governed systems rather than isolated interventions, several design priorities shift:

1. State Awareness Becomes Central. Therapies must infer not just target presence but biological readiness, stress signals, and contextual cues before acting.
2. Action Is Permissioned, Not Automatic. Intervention should occur only when conditions are permissive, rather than being triggered by a single static input.
3. Timing Is Treated as a Variable, Not a Constant. When action happens can matter as much as what action happens.
4. Stabilization Is a Design Goal. The objective is not merely to induce change but to guide the system toward a durable, healthy attractor state. This does not eliminate uncertainty, but it contains it.

Practical Implications For Development Teams

For teams working on in vivo programs, this reframing leads to several actionable shifts:

• Preclinical design: Ask not only whether a construct works but under what biological conditions it should not act.
• Translational planning: Evaluate whether early signals indicate sustainable trajectories or temporary perturbations.
• Clinical interpretation: Distinguish between “early success” and “trajectory alignment.”
• Portfolio strategy: Recognize that some failures are architectural mismatches, not execution mistakes.

These insights can inform go/no-go decisions earlier, before resources are committed to paths that are biologically difficult to exit.

The Real Risk: Treating In Vivo As Ex Vivo Plus Delivery

The greatest threat to in vivo therapies is not technical infeasibility. It is conceptual inertia. When in vivo systems are framed as ex vivo therapies with better delivery, development teams underestimate the complexity of modulating living systems in situ. The result is predictable: therapies that appear promising until they encounter biological realities they were never designed to manage.

Conclusion: Designing For Governance, Not Just Power

In vivo cell and gene therapies demand a different kind of thinking.

• Not more tools.
• Not tighter manufacturing.
• Not louder signals.

They require architectures capable of governing biological decision-making under uncertainty. Until we design therapies with that premise in mind, the industry will continue to see programs that work briefly, fail unexpectedly, and resist simple fixes. The future of in vivo therapeutics will not be decided by which modality is strongest, but by which architectures are wise enough to act only when biology is ready.

About the Author:

Wei Zhu is an independent consultant who works on the operating systems behind how next-generation cell and gene therapy (CGT) actually function. Her focus is at the architecture layer, where immune states, logic layers, and multi-modal interventions are orchestrated into stable therapeutic attractors. She helps organizations move beyond pipeline and platform thinking to design state-aware, logic-driven therapeutic systems that can adapt, respond, and stabilize in complex biological environments. Wei is the architect of the State-Logic Operating System (SLOS), a framework that upgrades CGT development from project execution to system-level design. Her work integrates insights from cell therapy, in vivo delivery, synthetic biology, and control theory to shape the next generation of CGT innovation.