Demis Hassabis, co-founder of DeepMind and Isomorphic Labs as well as recent Nobel Laureate, has articulated a vision for the future of drug discovery: building an artificial in silico cell that could simulate human biology with perfect fidelity. His goal and execution revolve around creating sophisticated computational models that drug developers could use to test millions of potential drugs virtually, predicting their effects with the same accuracy as laboratory experiments but at a fraction of the time and cost. This digital cell would theoretically massively improve the throughput of potential novel drug molecules being tested by orders of magnitude. But it doesn’t actually solve medicine’s single most persistent challenge: the high fraction of failure of potential drug molecules when applied to human beings, despite stellar performance on human cells.
Despite exponential improvements in cellular modeling capabilities over two decades, clinical trial success rates remain at 5-15% across therapeutic areas. Modern laboratories can grow human brain organoids with measurable electrical activity, create patient-specific cancer models using induced pluripotent stem cells, and test drug candidates against cell lines from thousands of genetically diverse individuals. Pharmaceutical companies routinely demonstrate dramatic therapeutic effects in these sophisticated systems. Yet 85-95% of drugs showing promise in cellular models fail when tested in humans. This failure rate has not improved meaningfully despite technological advances that have increased laboratory capabilities by several orders of magnitude since 2000.
The translation problem stems not from technical limitations in cellular modeling, but from fundamental conceptual limitations about the ability for individual cells to recapitulate safety and efficacy of drugs when applied to the entirety of the human body. These barriers cannot be solved by building more sophisticated models of human cells, regardless of computational power or biological fidelity. The solution requires complex, potentially AI-driven, “top-down” models of human patients that capture potential heterogeneity of individuals as well as heterogeneity of disease. These top-down models complement and synergize with the “bottom-up” models of biology that start from the cell, in helping scientists and therapeutics companies develop drugs that are safer, more effective, and more personalized for patients.
Understanding Bottom-Up vs. Top-Down Approaches
Before diving deeper into why cellular models fail, we need to clarify the distinction between “bottom-up” and “top-down” approaches to drug development. Bottom-up modeling starts with the smallest biological units—molecules, proteins, cells—and attempts to build understanding of human disease by studying these components in isolation or simple combinations. This reductionist approach assumes that if we understand all the parts, we can predict how the whole system behaves. It’s the dominant paradigm in pharmaceutical research: identify a molecular target, design a drug to hit that target, test it in cells, then animals, then humans.
Top-down modeling takes the opposite approach: starting with observations of complete human systems and using pattern recognition to identify what works without necessarily understanding every mechanistic detail. Instead of asking “how does this drug affect this cell?” it asks “which patients with which biological signatures respond to this treatment?” This approach embraces the complexity of human biology rather than trying to reduce it to manageable components. The pharmaceutical industry has invested decades and hundreds of billions of dollars perfecting bottom-up approaches. The time has come to seriously consider whether top-down strategies might offer a complementary path to therapeutic success.
Why Perfect Cells Don’t Make Perfect Predictions
The pharmaceutical industry invests approximately $120 billion annually in research and development, with individual drug programs consuming $1-3 billion from target identification through regulatory approval. To put this in perspective, the cost of developing a single approved drug now exceeds the GDP of small nations, yet our success rates haven’t improved since the 1990s. For every drug that reaches patients, 10-20 enter clinical trials, and hundreds more fail in preclinical development. Each failure represents not just financial loss but lost time for patients waiting for effective treatments.
The industry has responded by making cellular models ever more sophisticated. We can now grow miniature human brains in dishes, create “organs-on-chips” that mimic tissue functions, and use CRISPR to create cellular models of virtually any genetic disease. These aren’t primitive tools—they represent genuine scientific achievements that would seem miraculous to researchers from previous generations. Yet none of these advances have substantially improved our ability to predict which drugs will work in actual patients. The fundamental success rates remain unchanged because we’re perfecting our ability to model the wrong thing: individual cellular responses rather than integrated human physiology.
The cardiovascular field offers clear examples of this disconnect. Researchers can demonstrate that a compound improves heart cell contractility, reduces oxidative stress, and enhances calcium handling—all markers that should translate to better heart function. In isolated cardiac myocytes, these drugs perform beautifully, generating dose-response curves that suggest clear therapeutic benefit. But when tested in patients with heart failure, the same compounds might worsen outcomes. The drugs do exactly what the cellular models predicted at the cellular level, but they also trigger compensatory mechanisms—increased oxygen demand, altered kidney function, and neurohormonal activation—that overwhelm any benefits. Human cardiovascular function emerges from interactions between the heart, blood vessels, kidneys, nervous system, and hormones. No cellular model, no matter how sophisticated, captures this systems-level complexity.
Take torcetrapib, a drug designed to raise HDL (“good”) cholesterol by inhibiting an enzyme called CETP. In cellular assays, it perfectly blocked its target. In early human studies, it raised HDL levels dramatically—exactly as the cellular models predicted. Pharmaceutical executives were so confident they started planning factory expansions before Phase III trials completed. Then the data came in: patients taking torcetrapib were dying from cardiovascular disease at higher rates than those on placebo. The cellular models were completely right about the drug’s molecular action but completely wrong about its therapeutic value. The drug’s effect on blood pressure regulation and aldosterone secretion—invisible in cellular models—proved fatal.
Temporal Compression and Individual Variability
Human diseases unfold over years or decades, while cellular experiments typically last days or weeks. This temporal compression fundamentally alters what we’re studying. Alzheimer’s disease likely begins decades before symptoms appear, involving gradual accumulation of protein aggregates, slow neuronal death, compensation by healthy brain tissue, and eventual cognitive decline. Cellular models can show that a drug reduces amyloid beta production or prevents tau aggregation over a few weeks, but they cannot capture the decades-long interplay between protein accumulation, inflammation, vascular changes, and neural compensation that characterizes the actual disease. A drug that prevents early protein accumulation might be useless once neurons have died. A treatment that helps in mid-stage disease might be harmful early on. Cellular models, locked into experimental timeframes, cannot navigate this temporal complexity.
Cancer provides another stark example of temporal limitations. In culture, we can often eliminate 100% of cancer cells with targeted therapies or combinations. Victory seems absolute. But in patients, these same treatments might shrink tumors for months before resistance inevitably emerges. The cellular models correctly predict initial drug sensitivity but completely miss the evolutionary dynamics that drive treatment failure. Cancer isn’t a disease of cells; it’s a disease of cellular populations evolving under selection pressure over time. A tumor contains billions of cells with different mutations, and treatment creates intense selection for resistant variants. This evolutionary process plays out over months or years, not the days or weeks of laboratory experiments.
Perhaps the most underappreciated aspect of drug development is the enormous variability in human response to identical treatments. We discuss “the patient” as if humans were standardized biological machines, but two patients with identical diagnoses can respond oppositely to the same drug. In oncology, patients with the same cancer type, identical driver mutations, similar age and health status might show completely different responses to targeted therapy. One experiences complete remission while another shows no benefit. Our best cellular models, even those derived from the patients’ own tumors, often cannot predict these differences.
The numbers reveal the magnitude of this challenge. Even for drugs where we’ve identified genetic markers that affect response—like CYP2C19 variants for clopidogrel metabolism—these markers typically explain less than 25% of the variability in patient outcomes. The remaining 75% of variability remains unexplained by any factors we can measure in cellular models. This isn’t a failure of effort or technology; it reflects the fundamental complexity of human biology. Drug response depends on immune system status, microbiome composition, metabolic state, prior exposures, stress levels, and countless other variables that make each patient biologically unique. Cellular models necessarily standardize these variables to achieve reproducible results, but in doing so, they eliminate the very factors that determine real-world therapeutic outcomes.
Real-World Failures: When Cellular Success Misleads
The drug development graveyard is filled with compounds that looked revolutionary in cellular models but failed catastrophically in humans. Beyond torcetrapib, consider the MEK inhibitors in cancer. These drugs potently block a key growth signaling pathway that’s hyperactive in many tumors. In cellular models, they stop cancer growth cold. In patients, they can paradoxically accelerate tumor growth by triggering compensatory activation of parallel pathways. The cellular models weren’t wrong about MEK inhibition—they were wrong about its therapeutic relevance in the context of redundant signaling networks.
In Alzheimer’s disease, we’ve seen drug after drug successfully clear amyloid plaques both in cellular models and in patient brains, yet none have improved cognitive function. Solanezumab, aducanumab, and others hit their cellular targets perfectly but failed their clinical endpoints. The cellular models correctly predicted target engagement but couldn’t predict that clearing amyloid might be therapeutically irrelevant once cognitive decline begins. Similarly, BACE inhibitors prevent amyloid production in cellular models and reduce brain amyloid in patients, but some actually worsened cognitive function. The reductionist focus on amyloid, validated by countless cellular experiments, may have sent the entire field down an expensive dead end.
The Promise of Top-Down Discovery
Top-down drug discovery starts with the end result—which patients get better with which treatments—and works backward to understand patterns, even without complete mechanistic knowledge. This approach has already yielded unexpected successes. Metformin, originally a diabetes drug, shows promise in cancer prevention and longevity—discoveries that came from population observations, not cellular predictions. Viagra’s transformation from failed angina drug to erectile dysfunction treatment came from alert clinicians noticing unexpected effects. Beta-blockers’ use for performance anxiety emerged from musicians discovering that their heart medication also calmed stage fright.
Modern AI systems excel at finding patterns in complex, high-dimensional data—exactly what human biology presents. Instead of trying to understand every protein and pathway, these systems can identify which combinations of biological markers predict treatment response. They can find patterns that human scientists miss because they don’t fit our mechanistic frameworks. Imagine AI trained on millions of patient records, including complete molecular profiles, clinical histories, and treatment responses. Such systems might identify that patients with specific combinations of gene expression patterns, metabolite levels, and clinical features respond well to particular treatments—even if we don’t understand why.
Building Comprehensive Biological Profiles
The key to successful top-down modeling is comprehensive biological characterization that goes far beyond simple genetic testing. RNA profiling provides dynamic readouts of which genes are active in different tissues, changing with disease progression and treatment. Proteomics directly measures protein levels and modifications that genetics can’t predict. Metabolomics captures the end products of cellular processes, providing snapshots of what’s actually happening biochemically. Immunomics details the status of the immune system, increasingly recognized as crucial for drug response across all therapeutic areas. The microbiome influences drug metabolism and immune function in ways we’re only beginning to understand. Digital biomarkers from wearable devices provide continuous physiological monitoring that captures dynamics invisible in periodic clinical visits.
At Biostate AI, we’re focusing on RNA-based profiling because it provides a real-time biological readout that integrates genetic, environmental, and disease factors. Unlike DNA, which remains largely static, RNA levels reflect what cells are actually doing at any moment. By building massive datasets linking RNA profiles to treatment outcomes, we can train AI systems to predict which patients will benefit from specific interventions. This approach doesn’t require understanding why certain RNA patterns predict response—only that they do so reliably.
Skeptics might reasonably ask why top-down AI approaches would succeed where other paradigms have failed. Three factors suggest this time might actually be different. First, the scale of biological data now available exceeds previous eras by orders of magnitude. We can measure hundreds of thousands of molecular features (e.g. coding and non-coding RNA species) from individual patients at costs that continue to plummet. This data density approaches what’s needed to capture the true complexity of human biology. Second, AI capabilities have reached inflection points in pattern recognition that make finding subtle signals in noisy biological data feasible. Modern deep learning systems can identify patterns in datasets that would take human scientists lifetimes to analyze. Third, the regulatory and commercial environment has shifted to reward precision rather than broad efficacy. The FDA has approved drugs based on biomarker-defined populations rather than traditional disease categories. Basket trials test drugs across multiple cancer types sharing molecular features. Payers increasingly demand evidence that expensive drugs work in specific patients, creating economic incentives for personalized approaches.
The future isn’t about choosing between cellular models and patient-level AI systems—it’s about intelligent integration. Cellular models remain invaluable for understanding drug mechanisms, identifying initial candidates, and ensuring basic safety. They’re the workbench where scientists can rapidly test ideas and iterate on molecular designs. You can’t run a million experiments on human patients, but you can in cell culture. Top-down AI systems complement this by identifying which patients might benefit from drugs that cellular models suggest could work. They can reveal unexpected uses for existing drugs by finding patterns in patient response data that mechanistic models would never predict. The ideal drug development pipeline might use cellular models for initial discovery and safety testing, animal models for basic pharmacology, and AI-driven patient matching for clinical development.
The Urgency of Change
Every year we continue relying primarily on cellular models represents thousands of failed drug programs and millions of patients without effective treatments. The COVID-19 vaccine development showed what’s possible when we’re willing to challenge traditional paradigms. By running multiple development stages in parallel and using novel platforms, we compressed typical timelines from decades to months. The translation crisis in drug development demands similar urgency and willingness to embrace new approaches.
The current system isn’t just inefficient—it’s inadequate for addressing complex diseases that have resisted decades of cellular modeling approaches. We need to acknowledge that some problems can’t be solved by reductionism, no matter how sophisticated our cellular models become. The diseases that remain unsolved—Alzheimer’s, many cancers, autoimmune conditions—may require fundamentally different approaches that embrace rather than eliminate biological complexity.
The translation crisis in drug development isn’t a technical problem awaiting a technical solution. It’s a conceptual problem requiring a fundamental shift in how we think about drug discovery. Building perfect models of human cells, no matter how sophisticated, cannot solve problems that emerge from the complexity of complete human systems. The path forward requires humility about the limits of reductionist approaches and openness to empirical strategies that work with biological complexity rather than against it.
The revolution in drug development will come not from building better models of human cells, but from building better models of human patients. This doesn’t mean abandoning our hard-won understanding of molecular biology. It means recognizing that emergence—where system properties transcend component properties—is a fundamental feature of human biology that our drug development approaches must respect. The sooner we embrace this conceptual shift, the sooner we can begin addressing the diseases that continue to resist our current approaches.
By David Zhang and Claude 4.0 Sonnet
July 29, 2025
© 2025 David Yu Zhang. This article is licensed under Creative Commons CC-BY 4.0. Feel free to share and adapt with attribution.
Dave Zhang's Startup Blog 