Predictive, Not Prescriptive: The Real Role of AI in Therapeutic Innovation

Aug 22, 2026

Founder & CEO, Bioada

Predictive, Not Prescriptive: The Real Role of AI in Therapeutic Innovation

Like Be the first to like this

In recent years, a bold narrative has taken hold in the biotech world: that artificial intelligence can now autonomously generate drugs — from design to development — bypassing traditional bottlenecks and collapsing timelines. These claims often culminate in headlines about AI-generated molecules advancing to Phase I clinical trials, with implications that a revolution in drug discovery is not only coming but already underway.

The reality is more nuanced, and more grounded in biology than buzzwords.

AI is undoubtedly reshaping how we approach therapeutic innovation. But the leap from predictive modeling to prescriptive design is a scientifically fragile one. Rather than envisioning AI as a self-contained engine of discovery, the more accurate paradigm is that of AI-Assisted Drug Discovery (AADD) where computational tools are embedded into a human-guided, experimentally validated ecosystem.

In this piece, we explore the true role of AI in therapeutic innovation, the limits of end-to-end automation, and why biological context remains indispensable even when molecules enter clinical trials.

Chemically Elegant, Biologically Uncertain

AI excels at generating chemically novel structures. Advanced generative models can design small molecules that are synthetically tractable, conform to medicinal chemistry principles, and exhibit predicted target affinity. These tools operate across vast chemical spaces at a speed and scale impossible for human medicinal chemists alone.

And in at least one widely publicized example, such a molecule — designed entirely in silico using AI tools — has entered a Phase I clinical trial in humans.

This milestone has been widely celebrated as a proof point for AI-led drug discovery. But a closer analysis reveals a critical caveat: entry into Phase I is not validation of therapeutic relevance, let alone efficacy. It simply indicates that the molecule passed preclinical safety assessments and was cleared for first-in-human testing, a bar that thousands of traditionally discovered molecules have also cleared, only to fail in Phase II or III due to lack of efficacy or unforeseen toxicity.

This is the crux: AI models operate in chemical space, but drugs must function in biological space. A molecule that looks “perfect” to a neural network may fail in a human system due to unknown pharmacodynamics, off-target effects, poor metabolic stability, or failure to engage the intended mechanism in vivo.

Indeed, the overwhelming majority of clinical failures — including those for AI-generated candidates — result not from poor chemical design, but from insufficient biological translation.

The Human-AI Loop: Discovery is Iterative, Not Linear

Drug discovery is not a pipeline in the traditional engineering sense. It is a high-dimensional, nonlinear, iterative cycle — one that demands continual feedback between computational hypotheses and biological evidence.

In a mature AADD framework, AI contributes at multiple stages:

Target discovery via mining of multi-omics and clinical data.
Virtual screening to prioritize candidates with desired physicochemical and ADMET properties.
De novo molecular generation guided by learned chemical features.
Lead optimization through reinforcement learning and multi-objective design.

But in all of these, the critical step is validation and that remains deeply human. AI can generate a molecule with a theoretically high binding affinity to a modeled target, but:

Is that target actually disease-modifying in humans?
Is the model of the target protein conformation accurate?
Do downstream pathway effects support therapeutic benefit?
Can the compound evade degradation, immune surveillance, or metabolic conversion?

These questions cannot be answered in silico, no matter how elegant the algorithm. They require experimental systems, mechanistic expertise, and biological insight.

Even the most celebrated AI pipelines quietly rely on extensive human scaffolding — including human-led target prioritization, experimental compound synthesis, in vitro assays, and preclinical pharmacology — all of which shape and constrain what AI can realistically propose.

A Case Study in Misinterpretation

Consider the oft-cited case of an AI-designed molecule reaching human trials within a dramatically shortened timeline. This program was heralded as a new era of autonomous drug discovery, with timelines compressed from years to months and computational engines replacing traditional R&D workflows.

Yet behind the scenes, the process followed nearly all the standard industry steps, albeit aided by more efficient computational tools:

The disease target was chosen by domain experts, not discovered de novo by AI.
The molecule was synthesized, purified, and validated using conventional wet-lab workflows.
Pharmacokinetics, toxicology, and dosing were evaluated in animal models, not solely predicted.
The compound advanced to a Phase I safety study, with no efficacy signal, no biomarker-guided stratification, and no mechanism-of-action confirmation in humans.

This is not a critique of the effort, it is a critique of the narrative surrounding it. The actual science was sound, but the claim that AI discovered the drug “end-to-end” is a distortion of how drug discovery works.

It also overlooks a historical reality: many drugs that succeed in Phase I fail in Phase II and III. Success in early safety studies, even if accelerated, does not imply clinical utility. If AI is to contribute meaningfully to therapeutic innovation, its predictions must survive the brutal crucible of biological complexity and be tested accordingly.

The Risk of Overpromising

AI is a powerful engine for exploration, hypothesis generation, and decision support. But framing it as a prescriptive authority, one that can bypass empirical science, invites two dangers:

Scientific Disillusionment: Overhyped claims that dissolve under scrutiny damage credibility and may provoke backlash when high-profile programs fail.
Misallocated Investment: Resources may flow into AI-generated leads that are chemically sound but biologically inert, instead of being directed toward programs with deeper mechanistic insight.

It is especially important to avoid conflating chemical novelty with clinical relevance. The former is what AI generates with ease. The latter is what biology ultimately demands.

AADD: A Scientifically Grounded Paradigm

AI-Assisted Drug Discovery offers a realistic and productive framework, one that recognizes both the strengths and limitations of current technology. In this paradigm:

AI accelerates human decision-making, but does not replace it.
Models are validated against real-world biological systems, not just computational benchmarks.
Discovery is context-aware, grounded in pathophysiology, disease heterogeneity, and therapeutic selectivity.

Crucially, AADD emphasizes iterative refinement, a back-and-forth process where algorithms inform experiments and experiments inform algorithm tuning. This feedback loop is where true innovation emerges, not from digital alchemy, but from the disciplined fusion of computation and biology.

From Promise to Progress

AI is not discovering drugs in isolation, and that’s not a failure. It’s a reflection of the complexity of biology and the rigor required to bring safe, effective therapies to patients.

The idea that AI can prescriptively design drugs end-to-end is not supported by clinical outcomes, biological evidence, or mechanistic reasoning. What AI can do — and increasingly does well — is help researchers navigate complexity, generate new hypotheses, and reduce the search space for viable candidates.

But the final arbiters of therapeutic relevance are not algorithms, they are cells, tissues, organisms, and ultimately patients.

If we want to unlock AI’s full potential, we must resist the urge to mythologize it. Instead, we should focus on integrating it intelligently into discovery ecosystems that respect, and work within, the boundaries of biological reality.