No Perfect Drug Exists, but Maybe Perfect Cocktails Do
Today’s treatments struggle with limited efficacy, with cancer patients often dying within 24 months despite using $100,000+ drugs. Imagine instead walking into your doctor's office in 2035 and receiving a prescription not for a single pill, but for a personalized cocktail of 25 different FDA-approved medications, each dosed precisely for you right then, based on your current disease progression and intimate knowledge of your body's responses to different drugs and combinations of drugs. This prescription isn't the result of a traditional clinical trial testing your specific combination; those trials would cost billions and take decades to complete. Instead, your treatment cocktail was designed by an AI system trained on 100 million human multiomic datasets, capable of predicting how your unique biology will respond to this specific multi-drug combination. This vision represents a fundamental departure from today's pharmaceutical paradigm, where regulatory agencies require separate clinical trials for every new drug combination, making personalized multi-drug therapy economically impossible.
For most human diseases, there may be no "perfect" single drug molecule that cures everyone with minimal side effects. The entire pharmaceutical industry is essentially trapped in incremental optimization, searching through a 10^60+ parameter space of possible molecules for marginally better efficacy and safety profiles. Most drugs today are approved with overall response rates of just 20-30%, often accompanied by significant adverse events--a success rate that would be considered catastrophic failure in any other engineering discipline. This is because most of the low-hanging fruits in the chemical parameter space of potential drugs, such as penicillin, have already been discovered, and so we as a society are spending ever more dollars chasing ever less ripe apples.
But what if we're optimizing for the wrong thing entirely? What if the solution isn't finding better individual molecules, but learning to conduct pharmaceutical orchestras where each drug plays a specific role in a carefully choreographed therapeutic symphony?
The Nutritional Fallacy in Drug Development
There is an important parallel between nutrition and pharmacology: just as no single food can provide optimal human nutrition, no single drug may be capable of addressing the full complexity of a single human patient's disease. Humans require essential amino acids from diverse protein sources, vitamins from various fruits and vegetables, and minerals from different food groups. Even the most "nutritionally complete" single food would eventually lead to malnutrition and death. This principle extends beyond basic nutrition to therapeutic nutrition in clinical settings, where doctors don't prescribe a single "super-nutrient" but carefully balance proteins, carbohydrates, fats, vitamins, and minerals in precise ratios tailored to each patient's specific deficiencies and metabolic capacity.
Our current mono-drug approach to complex diseases creates therapeutic vulnerabilities that become apparent when single targets fail or develop resistance. The pharmaceutical industry's obsession with "magic bullets" traces back to Paul Ehrlich's early 20th-century concept of finding compounds that would selectively kill pathogens without harming the host. This framework worked brilliantly for infectious diseases--penicillin revolutionized medicine by targeting bacterial cell walls that don't exist in human cells.
But chronic diseases like cancer, autoimmune disorders, and neurodegenerative conditions involve our own biological systems malfunctioning, making selective targeting exponentially more challenging. Modern drug development has become trapped in what economists call a "local optimization problem"--continuously refining approaches that worked for simple infectious diseases rather than recognizing that complex chronic diseases require fundamentally different strategies. The result is an industry that celebrates modest improvements in drugs that help only a fraction of patients, while the fundamental approach remains unchanged.
Three Faces of Resistance: When Biology Outsmarts Chemistry
How do diseases outsmart single-drug treatments? Here are three revealing examples, blockbuster drugs across three important disease areas (oncology, immunology, and cardiovascular medicine) from which we can observe significant differences in how and why a particular drug fails. They also hint at how a personalized drug combination approach to disease treatment could potentially provide better patient outcomes.
Osimertinib: The Evolutionary Arms Race in Cancer
Non-small cell lung cancer kills approximately 135,000 Americans annually, making it the leading cause of cancer death worldwide. About 15% of these cases harbor EGFR mutations, which were historically associated with poor outcomes until the development of targeted therapies. Osimertinib, approved by the FDA in 2017, represents one of oncology's greatest precision medicine triumphs. This oral medication, costing approximately $150,000 annually, achieves an remarkable 80% objective response rate in first-line treatment of EGFR-mutated tumors--response rates that seemed impossible just two decades ago. With median progression-free survival of 18.9 months versus 10.2 months for earlier-generation EGFR inhibitors, it appeared we had cracked the code of targeting cancer's genetic vulnerabilities.
But resistance is inevitable, and it's mechanistically fascinating. Despite this highly effective targeted therapy, most patients experience disease progression within 18-19 months through remarkably diverse mechanisms. Tumors develop acquired EGFR mutations like C797S, amplify alternative pathways like MET and HER2, undergo histological transformation to small-cell carcinoma, or activate completely EGFR-independent survival pathways. The cancer essentially maintains what researchers call a "resistance toolkit"--a diverse arsenal of genetic and epigenetic escape routes that can be rapidly deployed when selective pressure is applied. When we block one pathway, tumors don't simply die--they quickly identify and amplify alternative survival mechanisms, often within months.
This illustrates the fundamental challenge of evolutionary resistance. Cancer represents evolution in fast-forward, with millions of mutations creating vast genetic diversity within each tumor. When we apply selective pressure with a single drug, we're essentially performing a natural selection experiment--killing off sensitive cells while inadvertently selecting for resistant variants that were present in tiny numbers before treatment. Historical parallels exist in warfare and pest control: during World War II, the introduction of radar gave Allied forces a decisive advantage until German scientists developed radar-jamming technology and stealth tactics. Similarly, DDT initially decimated mosquito populations until resistant variants became dominant through intensive selection pressure. Single-mechanism interventions, no matter how initially effective, create selection pressure for resistance.
The current oncologic "solution" follows a predictable pattern: osimertinib → chemotherapy → immunotherapy. Each transition gives tumors time to evolve new resistance mechanisms while we chase their latest adaptation. We're always playing catch-up in an arms race where the enemy evolves faster than we can develop new weapons. But researchers have discovered something intriguing: when tumors develop both T790M and C797S mutations on different chromosomes (in trans), they become vulnerable to combinations of first- and third-generation inhibitors simultaneously, suggesting that hitting multiple pathways from the start might prevent the sequential resistance that plagues single-agent therapy.
Ustekinumab: When Immune Systems Route Around Roadblocks
Crohn's disease affects approximately 700,000 Americans with chronic intestinal inflammation that can lead to bowel obstruction, perforation, and life-threatening complications requiring multiple surgeries. Before modern therapies, many patients faced repeated hospitalizations and progressive disability. Ustekinumab, a monoclonal antibody administered by intravenous infusion every 8-12 weeks at an annual cost of approximately $100,000, blocks IL-12 and IL-23 signaling pathways central to inflammatory bowel disease. Clinical trials show initial response rates of 34-56% depending on patient treatment history, with many patients achieving sustained remission when the therapy works.
But the resistance mechanisms that emerge don't follow cancer's evolutionary playbook. Autoimmune resistance occurs through pathway redundancy and cellular cross-talk that allows inflammation to persist despite pathway blockade. One common example resistance mechanism is the expansion of apoptosis-resistant TNFR2+IL23R+ T cells. Initially, ustekinumab effectively neutralizes IL-23, reducing inflammation by blocking IL-23-dependent T cells. Over time, certain T cell populations expressing high levels of TNFR2 and IL23R adapt and expand due to their resistance to apoptosis and reduced reliance on high IL-23 signaling (due to compensatory survival signals from TNFR2). These resistant T cells avoid apoptosis pathways that typically limit inflammatory cell expansion, effectively bypassing the therapeutic action of ustekinumab. As these resistant T cells proliferate, they maintain or exacerbate inflammation despite IL-23 blockade, leading to clinical resistance to ustekinumab. Additional resistance pathways compound the problem, such as IL-22 responses contributing treatment failure through mechanisms that bypass both TNF and IL-23 blockade.
The immune system doesn't evolve resistance like cancer; it activates redundant inflammatory cascades simultaneously, creating what engineers would recognize as "fault-tolerant" biological systems. This explains why many autoimmune patients require dose intensification over time but continue responding to their medications. Unlike cancer resistance, where target mutations make drugs completely ineffective, autoimmune resistance often represents "pharmacokinetic loss of response"--needing higher drug concentrations to suppress increasingly complex inflammatory networks. Single-pathway blockade in autoimmune diseases is like trying to stop a river by blocking one tributary while ignoring the dozens of other water sources feeding the main channel.
Clopidogrel: The Humbling Mystery of Pharmacogenomic "Success"
Cardiovascular disease remains the leading cause of death globally, claiming over 650,000 American lives annually through heart attacks and strokes. Clopidogrel, a generic oral medication costing less than $100 annually, helps prevent these events by inhibiting platelet aggregation--the process that forms dangerous blood clots in arteries. Approved in 1997 and prescribed to millions of patients worldwide, it represents a pharmacogenomic "success story" where we've identified key genetic polymorphisms, developed clinical testing protocols, and issued FDA guidelines based on genetic variants.
CYP2C19 loss-of-function alleles (*2 and *3) are present in approximately 30% of patients and significantly reduce clopidogrel's conversion to its active metabolite. Patients carrying these polymorphisms have 68% higher risk of major adverse cardiovascular events and 75% higher risk of stent thrombosis compared to normal metabolizers. This led to FDA black box warnings and clinical recommendations for genetic testing--precision medicine working exactly as intended. The logic seemed straightforward: test patients for genetic variants, adjust therapy accordingly, and achieve better outcomes through personalized medicine.
But here's the humbling reality: in rigorously controlled studies of healthy subjects with wild-type CYP2C19 genetics, completely free of medications, food interactions, and compliance issues, clopidogrel pharmacokinetics still varied wildly with coefficients of variation exceeding 30-40%. Even more striking, all identified genetic and environmental factors together accounted for only 18% of pharmacokinetic variation and 32-64% of pharmacodynamic variation. In our most "solved" pharmacogenomic example, 80% of why the drug works or doesn't work remains completely mysterious. Additional factors like ABCB1 polymorphisms, P2Y12 receptor variants, PON1 enzyme variations, and drug-drug interactions with proton pump inhibitors explain some additional variability, but the vast majority of therapeutic response remains unexplained by current scientific understanding.
The clopidogrel story should fundamentally humble anyone who believes we understand drug response well enough to optimize it computationally using current approaches. If our most thoroughly studied medication with extensive genetic testing, detailed metabolic pathway mapping, and decades of clinical experience still leaves 80% of therapeutic variability unexplained, what does this mean for AI-driven single-agent drug discovery? This uncertainty has profound implications for how we approach precision medicine. Current AI drug development assumes that if we map target interactions, predict binding affinities, and understand metabolic pathways, we can design better molecules. But clopidogrel demonstrates that even when we identify the correct target (P2Y12 receptor), understand the activation pathway (CYP2C19 metabolism), characterize the major genetic variants, and control for obvious confounding factors, massive biological variability persists through completely unknown mechanisms.
Consider what this means practically: patients with identical genetics, identical drug concentrations, identical baseline physiology, and identical environmental exposures still show wildly different therapeutic responses. What biological processes account for this variability? The honest scientific answer is that we have no idea. This represents both the fundamental limitation of current drug development and the greatest opportunity for combination approaches. Rather than waiting decades to characterize each unknown biological pathway individually, combination therapy could provide therapeutic redundancy across multiple systems simultaneously--essentially creating "fault-tolerant" treatments that work even when we don't understand all the underlying mechanisms. Together, these examples illustrate why targeting multiple biological pathways simultaneously may be essential.
Historical parallels exist in engineering systems where redundancy compensates for incomplete understanding. The Apollo spacecraft used triple-redundant computer systems not because engineers fully understood every possible failure mode, but because they recognized that complex systems could fail in unpredictable ways. When one computer disagreed with the other two, the majority vote prevailed--a principle that could apply to therapeutic combinations. Drug combinations might function similarly: if three different mechanisms each have 70% probability of working independently, the probability that all three fail simultaneously drops to less than 3%. This mathematical principle of redundancy could make combination therapy effective even when individual components show modest efficacy.
Proofs of Concept for Efficacy of Combination Therapy
The pharmaceutical industry has several compelling examples of combination therapy transforming treatment paradigms, demonstrating that the concept isn't theoretical. HIV/AIDS treatment provides the most dramatic success story. In the mid-1980s, HIV was essentially a death sentence despite intensive research efforts and massive pharmaceutical investment. Single-agent approaches like AZT (zidovudine) showed modest temporary benefits but inevitable resistance development. The breakthrough came with Highly Active Antiretroviral Therapy (HAART)--combinations of drugs targeting different viral mechanisms simultaneously. HAART transformed HIV from a fatal disease to a manageable chronic condition precisely because it made viral resistance exponentially more difficult. HIV would need to develop resistance to multiple drug classes simultaneously, requiring multiple independent mutations that rarely occur together.
A more recent, nonobvious example is cancer immunotherapy. Although the drug is a monotherapy, its downstream effects on the human immune system effectively makes it a combination therapy--just one orchestrated by the patient's own biology rather than designed by pharmaceutical companies. When checkpoint inhibitors like pembrolizumab (Keytruda) or nivolumab (Opdivo) work, they don't deploy a single mechanism against cancer. Instead, they unleash CD8+ cytotoxic T cells, CD4+ helper cells, natural killer cells, B cells producing tumor-specific antibodies, complement cascade activation, and dozens of cytokine networks--all attacking the tumor simultaneously. Cancer must escape ALL of these mechanisms to survive, which is exponentially more difficult than escaping a single drug target.
This is why immunotherapy can achieve durable responses that chemotherapy rarely accomplishes. Traditional chemotherapy applies selective pressure for resistance mutations, while immunotherapy creates what researchers call "immune memory"--a persistent surveillance system that can recognize and eliminate cancer cells even if they acquire new mutations. But immunotherapy failures reveal the same principle from a different angle. Even patients with high tumor mutational burden who should theoretically have many targetable neoantigens often don't respond to checkpoint inhibitors. The failures typically result from immune exclusion (T cells can't infiltrate tumors), immunosuppressive networks (regulatory T cells and inhibitory cytokines), antigen presentation defects, or checkpoint redundancy beyond PD-1 and CTLA-4.
Recent combination immunotherapy trials demonstrate the power of addressing multiple failure modes simultaneously. Combinations targeting PD-1 plus CTLA-4, or checkpoint inhibitors plus anti-angiogenic agents, or immunotherapy plus chemotherapy show improved response rates precisely because they address different resistance mechanisms in parallel. The temporal dimension adds another crucial layer: the immune system doesn't just use multiple mechanisms--it sequences them dynamically over time. Initial innate immune responses trigger adaptive responses, leading to memory formation and long-term surveillance. Optimal therapeutic combinations might similarly require temporal orchestration rather than just simultaneous dosing.
The Regulatory Bottleneck: Why Personalized Drug Cocktails Currently Cannot be Approved
The greatest barrier to personalized combination therapy isn't scientific--it's regulatory and economic. Under current FDA frameworks, every new drug combination requires separate clinical trials demonstrating safety and efficacy. Testing even a modest 3-drug combination across different dosing schedules, administration sequences, and patient populations would require thousands of participants and hundreds of millions of dollars. Scale this to the 20+ drug combinations that might be optimal for complex diseases, and the combinatorial explosion becomes financially impossible. A comprehensive clinical trial program for personalized multi-drug therapy would cost more than the entire current pharmaceutical industry spends annually on research and development.
The Phase 1 / Phase 2 / Phase 3 clinical trial regulatory framework made sense when we were developing single agents for infectious diseases, in which we can find large affected populations of patients to evaluate drug safety and efficacy statistically. However, as we move to a personalized N=1 medical treatment future, this approach breaks down because there are no other patients with the exact same biological profile to do the clinical trial on. Requiring traditional clinical trials for each potential combination essentially makes personalized medicine impossible under current regulatory structures. It's like requiring separate safety testing for every possible meal combination before restaurants could serve food--technically thorough but practically prohibitive.
However, there is hope, as FDA have approved "basket trials" traditionally outside the standard clinical trial approach, in order for cancer drugs to be used for many different cancers despite lack of sufficient statistically powered cohorts within each cancer type. Entrectinib, an NTRK inhibitor, received FDA approval for all solid tumors that harbor a gene fusion involving one of the 3 NTRK genes. This is a remarkable precedent, because there are thousands of distinct NTRK gene fusion subtypes, as well as tens of different cancers. Thus, the regulatory framework already exists for approving treatments based on predictive biological signatures rather than traditional population-based clinical trials. We can imagine a potential future where advanced AI, trained on millions of human's biomolecular datasets, can be potentially approved by the FDA to prescribe complex combinations of many different drugs, despite that specific combination never before having been tested in a human.
Building Advanced AI to Create Personalized Drug Cocktails
This convergence of regulatory precedent and AI capability creates an unprecedented opportunity for radically extending the healthy lifespan of people everywhere. At Biostate AI, we believe that we can and will build advanced AI that can accurately prescribe unique combinations of drugs to optimally effectively and safely treat each human patient's disease. This will require a massive amount of what we call "Dynamic Multiome" data from millions of human participants across multiple points in time at different stages of disease progression and conventional medical treatment. No such Dynamic Multiome dataset exists currently at any reasonable scale.
The Dynamic Multiome -- the 100,000+ biomolecules that are present in the blood sample of a human at a particular point in time -- extends beyond the core expertise of Biostate AI in RNA alone. We are currently assembling a consortium of leading life sciences companies and top academic researchers to build this unprecedented dataset, and provide more details on consortium in the coming months. The purpose of this blog post is to detail one ultimate goal of the consortium, as well as research into advanced patient-centered AI.
No perfect drug exists for most diseases; this is why people still die of cardiovascular disease, cancer, respiratory disease, autoimmune disease, and other medical conditions. However, we as a civilization may be on the cusp of being able to build perfect drug combinations to dramatically increase healthy lifespan. Every year we delay the development and deployment of this technology means an extra 60 million people are dying pre-maturely.
By David Zhang and Claude 4.0 Sonnet
June 17, 2025
© 2025 David Yu Zhang. This article is licensed under Creative Commons CC-BY 4.0. Feel free to share and adapt with attribution.