The Single Target Delusion: Why AI Is Optimizing Only for Biology's Edge Cases

In 1907, Paul Ehrlich coined the term "magic bullet"—a single molecule (the drug) that would hit one single protein (the target) and cure a disease. A few years later, he found one: Salvarsan attacked the bacteria causing syphilis without harming the patient. The idea is an elegant paradigm that has shaped drug development for the following century. It is also wrong for the majority of diseases.

The problem is that this magic bullet concept doesn't work for most diseases. But we've spent a century following it anyway.

When Magic Bullets Actually Work (And Why They Don't Usually)

Reality is that magic bullets are the exception and not the rule, simply because it is rare for diseases to be caused by only one protein. Chronic Myelogenous Leukemia (CML) is driven by literally one mutated protein, BCR-ABL, that causes uncontrolled growth of white blood cells. If we hit the mutated protein - we stop the disease. CML is therefore the classical example for a disease solved by a magic bullet that drug developers cite. The magic bullet drug is Gleevec, which earned Novartis $15B in revenue over a decade, and patients who respond have normal life expectancy. That is as close as we've ever gotten to a magic bullet.

Herceptin is a breast cancer drug that was designed as a magic bullet to target HER2 receptors. But in practice, Herceptin is rarely used alone - doctors realized it works best when combined with other chemotherapy drugs. Even when there is a magic bullet to hit one target - that’s not enough. Yet most cancer drugs are still developed for a single target. 

Biology is more complex. Cancer may start with one mutated protein, but as the disease progresses, it accumulates more mutations. Each new mutated protein helps the cancer grow uncontrollably. These mutations can even develop in response to treatment: while a drug targets one protein, the cancer evolves another mutation to activate an alternative pathway and continue growing. This is actually how cancer cells dodge any magic bullet we shoot at them. 

Biology is a system rather than a collection of individual reactions. There are multiple interactions between different proteins and other biomolecules - metabolites, DNA, RNA, and other proteins. It is  a network of multiple signals with many knobs and switches. This is true for our normal biology and also for many diseases, not just cancer: obesity involves metabolic pathways, appetite regulation, and insulin signaling; aging depends on cellular stress responses, mitochondrial function, and DNA repair; the immune system has many checks and balances for activating and deactivating it, that sometimes go wrong and lead to autoimmune diseases. All these interconnected processes require coordinated intervention across multiple targets. Yet we continue designing drugs to hit single targets in isolation.

Multi-Target by Design, Not Chance

Reality is that it is common for drugs to work on several targets. The problem is that it is mostly not by design. Studies found that drugs bind to an average of 6.3 protein receptors, not just the intended target. Unintentionally acting on multiple targets could occasionally lead to surprising positive effects, but mostly to negative ones.

Aspirin, one of the most common and successful drugs today, is not a magic bullet but a “dirty drug” - hitting multiple targets unintentionally. It hits multiple proteins like COX-1 and COX-2, affects platelet aggregation, and has anti-inflammatory effects across multiple pathways. Aspirin’s multi target activity is accidental - the drug was discovered in 1897, before scientists had knowledge of protein targets and specifically before they knew about COX-1 and COX-2 enzymes. We got lucky that this multi target activity led to many clinical benefits.

More commonly, we get negative side effects when drugs hit unintended proteins, sometimes causing toxicity and organ damage. Vioxx, another pain and anti-inflammatory medication that was supposed to be a magic bullet and act selectively on COX-2 protein, was also hitting different proteins in the cardiovascular system, leading to tens of thousands of heart attacks and strokes. Vioxx was pulled from the market in 2004.

We want well designed drugs that work on multiple targets - specific combinations that have synergy and lead to better clinical results. Eli Lilly specifically designed Zepbound as a dual-targeting molecule from day one. Zepbound hits both GLP-1 and GIP receptors, while Ozempic hits a single target - GLP-1. The dual targeting strategy paid off: Zepbound is more effective than Ozempic, with 22% weight loss compared to 17%. 

AI Is Turbocharging the Wrong Approach

AI has given us incredible tools for drug discovery. AlphaFold3 and its clones and successors can predict protein structures with atomic precision, which won the AlphaFold team the Nobel Prize. These models focus on predicting the structure of how a single drug binds to a single protein. While these AI models greatly speed up our understanding of how a protein and a molecule structurally fit, these models actually promote the "magic bullet" reductionist approach, while missing what we really need to develop successful drugs.

The latest AI models focus on accelerating making magic bullets. Recent advances from AlphaFold to generative models that design antibodies like Chai2 and BoltzGen, have focused on optimizing the design of a single molecule that will bind to a target with hope of achieving atomic precision. But this will not lead us to better drugs.

The 85% failure rate in drug discovery isn't solved by AI predicting how a drug fits into a protein and jointly predicting the drug-protein structure. Knowing how a molecule binds to one protein tells us almost nothing about what happens in a living system. A cancer cell's survival doesn't depend on a single binding between a protein and a molecule. It depends on hundreds of proteins working in concert: growth signals, metabolism, immune evasion, all running simultaneously in feedback loops. 

The failure rate persists because we're designing drugs for individual proteins when we need to design them for biological networks, and AI focused on structure prediction isn't solving that fundamental mismatch.

Building AI for How Biology Actually Works

What we need is AI models that could lead to rational multi-target drug design. We need AI models that help us design drugs in a way that fits how biological systems actually work. We need drugs that work on multiple targets while avoiding other, unwanted targets - by design, not by chance.

The current AI architectures have an inherent single-target design bias: These AI models are fundamentally built to optimize binders or drugs against one specific target at a time, reflecting the “magic bullet” paradigm.  

Optimizing for multiple targets increases complexity beyond what current models can handle. Since each protein target has unique structural and chemical requirements, when current AI models try to satisfy two or more targets at once, they are forced to make trade-offs. Adding a second target can cause the model to effectively “forget” or compromise on the first objective.

Beyond the architectural limitations, there's a fundamental speed problem. Current models generate protein-drug structures by modeling every atom, which makes them computationally intensive. The models can't iterate fast enough to explore what happens when hitting multiple targets at once. They also can't efficiently model how a candidate drug might cause side effects by unintentionally interacting with any of thousands of human proteins. For a full drug discovery program generating thousands of candidates, checking each candidate against the entire proteome is computationally prohibitive.

The next generation of medicines won't be magic bullets. They'll be precisely engineered multi-target drugs that modulate disease networks. If we're going to use AI to design them, we need to point it at the right problem: not atomic precision on single proteins, but system-level modulation of complex diseases. We would need new architectures that overcome these limitations and are more compatible with the complexity of biology.

‍