AI in Biopharma: Promise and Caution

The past few years have been extraordinary for AI in life sciences. AlphaFold solved a problem that haunted structural biology for fifty years. Major pharmaceutical companies are announcing partnerships with frontier AI labs. Academic labs are training models on biological data with results that would have seemed impossible a decade ago. The energy in the space is real, and the potential is genuinely thrilling.

I work in this field. I’ve seen the progress firsthand. And I need to be direct: we should be excited, but we should also be cautious.

The caution isn’t about whether AI can help drug development. It absolutely can. The caution is about the gap between what looks promising in a research context and what actually works when bringing a drug to patients. That gap appears wider than many realize from outside the industry, and it’s driven by three structural realities that algorithms alone cannot overcome.

The Feedback Loop Problem

Tech progress tends to be fast partly because feedback is immediate and unambiguous. Code either compiles or it doesn’t. Results are observable in seconds. This tight feedback loop enables rapid iterative improvement.

Biological systems don’t work this way.

In drug development, the causal chain is long, complex, and often hidden. A compound is designed and tested in cells. Cellular responses are only partly understood because cellular behavior depends on numerous variables, many of which remain unmeasurable. Testing then moves to animals, which have immune systems, vascular systems, livers, and microbiomes—elements that matter for human biology but are absent in cell cultures. Observed outcomes are difficult to trace back to the original hypothesis. When moving to human trials, complexity increases further because human biology encompasses variables like psychology, behavior, diet, genetics, and disease heterogeneity.

Each step typically takes months or years. Confounding factors are numerous. Direct causality is rarely obvious. The iterative path to understanding is far less direct than it is with code.

This doesn’t mean AI lacks utility. Rather, it highlights a fundamental constraint: the feedback loop is long, the signal is noisy, and understanding causality in biological systems presents challenges distinct from other hard engineering problems.

The Data Generation Problem

AI advances typically follow a familiar pattern: obtain more data, increase model scale, achieve better results. This formula works when generating data is inexpensive. It works with images and text because both are abundant online.

Biological data operates differently. This difference is significant and often underappreciated.

Good sources of human data exist—biobanks, electronic health records, naturally occurring disease variation. But these capture correlation rather than causation. Understanding whether a drug works requires perturbing a biological system and observing outcomes. For humans, this occurs through controlled trials, which are expensive, time-consuming, and ethically constrained. Ethical boundaries limit the scope of human experimentation with candidate compounds.

Researchers therefore build model systems—cell lines, organoids, animal models. Each involves compromises. Cell lines lack immune systems, vascular systems, and the complexity that determines safety or efficacy in humans. Animal models are closer to human biology but remain distinct. Critically, using these systems at scale to generate training data requires perturbing living cells and organisms in ways that raise ethical considerations. Questions about appropriate scale—how many cells to test, how many animals, what constitutes justifiable use versus waste—lack clear consensus.

This creates a fundamental tension. Data needed to train sophisticated models is expensive to generate and subject to ethical constraints that technology companies building models on scraped data typically don’t encounter. Data that is easier and cheaper to generate tends to be biologically less relevant. These constraints—the sparse high-dimensional biological data, the limited ability to perturb humans for experimentation, the fundamental gap between model systems and human biology—represent core obstacles to scaling AI in drug development, as explored in detail by Ankit Garg’s analysis of data challenges in biological research.

The “sim-to-real gap”—the difference between model system performance and human performance—represents a meaningful constraint on the improvements any algorithm can deliver.

The Culture Problem

The third obstacle is the hardest to talk about because it’s not a technical problem, which means it’s harder to solve.

Biopharma is a regulated industry. This regulation is necessary and important—patient safety is at stake. But regulation encourages caution, which shapes organizational structures and decision-making processes toward risk aversion. Over decades, risk aversion becomes embedded in organizational culture.

The technology sector operates differently. Tech is often driven by move-fast-and-break-things dynamics. Data is treated as a business asset. Digital capabilities are integrated into competitive strategy. When a tech company identifies a missing capability, the response tends to be rapid: build it, buy it, or hire for it. In pharma, analogous decisions typically pass through multiple approval layers, risk committees, governance boards, and constraints imposed by legacy systems.

This isn’t a moral judgment about either industry. Risk aversion in pharma is appropriate given its stakes. Speed in tech is appropriate for non-critical applications. But the cultural difference creates a meaningful capability gap. Pharma companies have, on average, moved more slowly on cloud infrastructure, data engineering practices, modern analytics, and data-driven decision-making. This matters because AI functions best when embedded in organizations that treat data as currency and iteration as a learning mechanism.

A culture optimized around regulatory compliance and operational stability tends to encounter obstacles when adopting AI at scale. Such cultures typically don’t shift rapidly in response to new hiring or technology deployment alone.

What Actually Needs to Happen

The opportunities are significant. Drug discovery is slow, expensive, and prone to failure. If AI can accelerate key steps or reduce failure rates, the potential value is substantial. Capturing that value requires moving beyond initial excitement into sustained, difficult work.

Three categories of work emerge as particularly important.

1. Strengthen the Fundamentals

Effective AI deployment requires basic infrastructure. This includes:

• A modern tech stack. Not for prestige but for capability. Cloud data platforms, modern analytics tooling, and engineering practices that enable iteration. Some large pharma companies continue building analytics on systems designed decades ago. Sophisticated AI performs poorly on unstable foundations.

• Data-backed decision making. This is cultural rather than technical. It involves shifting the default question from “who has authority to approve this” to “what data supports this decision.” It requires building organizational capacity for measurement and feedback, even when feedback arrives slowly.

• Documented processes. Optimization is difficult without measurement, and measurement is difficult without clear process definition. Significant portions of drug development expertise reside within individuals, embedded in practice rather than documented. Applying AI to a process first requires understanding and describing it explicitly.

This work sounds unglamorous. For many, it is. Yet it is foundational and a pre-cursor to AI transformation. The pattern holds across digital transformation efforts: the assumption that late movers can skip generations of capability and leap ahead rarely holds. The late mover still must build the organizational and technical substrate—the infrastructure, the culture, the data literacy—that early movers have already struggled to construct. There are no shortcuts to capability.

2. Choose Impact Over Flashiness

Projects that receive approval often emphasize excitement or cutting-edge positioning. “AI for target discovery” has more apparent appeal than “AI to reduce data preparation time.” Yet the latter often delivers more concrete value.

The applications likely to deliver substantial returns address specific bottlenecks. They shorten timelines, reduce failure rates, or eliminate manual work at scale. Within drug development, these opportunities often appear foundational rather than flashy: standardizing data taxonomy across clinical trials, shortening candidate-to-first-human timelines, improving information quality for regulatory submissions.

These applications lack glamour. But their impact tends to be measurable and the feedback loop, while still extended, is tighter than for applications further along the development pipeline.

3. Make Data and Technology External-Facing

Many biopharma companies treat data and technology primarily as internal capabilities. But they can function as competitive differentiators and revenue sources.

Possibilities include: monetizing clinical trial data to accelerate broader research efforts, building platforms for external researcher access to certain datasets, publishing insights that strengthen positioning with regulators and investors, and creating differentiated stakeholder experiences through improved physician information or patient support programs.

External-facing data and technology strategies shift incentives in important ways. They create faster feedback loops, provide market validation, and build accountability. Capabilities advance because markets reward them, not solely because compliance requires them. This market-driven acceleration appears more powerful for cultural change than mandates alone.

Why I’m Still Optimistic

The obstacles are real. So is the opportunity. And disruption from outside the industry cannot be underestimated.

AlphaFold originated from researchers at DeepMind, not structural biologists within pharma companies. Significant advances in computational biology frequently come from researchers unfamiliar with established practices, approaching problems from different angles.

Future breakthroughs might emerge from roboticists reconceptualizing how experiments are conducted, from economists applying market design principles to development timelines, or from software engineers trained outside pharma who bring different mental models of what’s feasible.

An influx of people and ideas from faster-moving industries—those accustomed to more radical technology adoption and less deference to established practice—could catalyze meaningful change.

This process is already underway. It will likely introduce cultural friction. It will likely accelerate pace. It will likely create the kind of productive tension that organizations require for genuine transformation.

The potential of AI in drug development is substantial. Realizing it requires less emphasis on excitement and more on fundamentals, less focus on flashiness and more on measurable impact, and greater openness to outside perspectives disrupting conventional approaches.