Drug discovery has always been biased toward what is accessible. Cell surface proteins and enzymes are tractable, they have structure, defined binding sites, and fit the tools we built. CTLA-4 and PD-1 are the canonical success story: immune checkpoints sitting on the cell surface, blocking them with antibodies was conceptually and technically straightforward. The biology was real, the target was reachable, and the results were transformative.
But most of the biology that drives disease, including resistance to those same checkpoint therapies, sits inside the cell.
That mismatch has shaped the field. Out of roughly 20,000 protein-coding genes, only around 15 to 20 percent has been considered druggable, and fewer than 1,000 proteins are targeted by approved drugs. The remaining majority includes many of the most important regulators of disease.
What has changed is not the biology. It is how we approach it.
The question is no longer whether a protein has a binding pocket. It is whether we can change what that protein does.
That shift underlies a convergent set of approaches. Small molecules can force new protein-protein interactions. Proteins can be brought into proximity to alter function or stability. Covalent chemistry opens entry points across the proteome by targeting ligandable cysteines. These approaches do not depend on classical inhibition. They create events. A small molecule does not block a target, it engineers a protein-protein interaction that has no biological precedent, recruiting a previously untouchable protein to a degradation complex or forcing a functional interaction that rewires cell behavior. The challenge has been that most discoveries of this kind have been serendipitous. The next frontier is moving from lucky hits toward systematic, rational design of molecular glues at scale. From serendipity to design. That transition is where the next decade of this space lives.
As a result, the proteome has opened up. But opening it creates a different problem.
We now have thousands of ligandable sites, an enormous number of possible interactions, and effectively unlimited chemical space. The limiting factor is no longer access. It is choosing what matters, which cysteine is functional, which interaction is worth inducing, which perturbation will shift a disease state.
In immuno-oncology, that question is increasingly pointed. Many of the mechanisms underlying checkpoint resistance involve cell fate decisions driven by epigenetic signaling. Cancer cells can silence immune recognition by adopting a neuroendocrine identity, a transdifferentiation program governed by chromatin remodeling, not mutation. T cells exposed to persistent antigen do not simply become less active; they undergo epigenetic reprogramming that locks exhaustion in place, making it refractory even after the original stimulus is removed. Tumor-associated macrophages are shaped by the same logic, their tumor-permissive phenotype is maintained epigenetically, not just by local signaling cues. These are not failures of target identification. They are failures of cell state. And cell state is controlled by the transcription factor networks and chromatin dynamics that classical drug discovery was never built to reach.
That is exactly where the new toolkit becomes relevant, and where the hardest problems now live.
What AI Actually Does Here
AI connects these layers by doing something pattern matching alone cannot. Tools like AlphaFold predict not just a protein's ground-state structure but the conformational landscape it occupies. That changes what is designable. BRD4, a transcriptional coactivator long considered undruggable due to its shallow, featureless interaction surface, becomes tractable when you can predict which conformational states are accessible and design compounds that stabilize or exploit them, whether to block a specific interaction or present the protein in a degradation-competent orientation for targeted destruction. The target did not change. The ability to design toward a specific conformation did.
Each layer of this toolkit produces useful but incomplete information in isolation. Chemoproteomics identifies where binding is possible. Induced proximity defines what interactions can be created. Functional assays measure what happens when they occur. AI integrates across all three, moving from ligandable site to predicted engagement, from interaction to functional outcome, and from experimental results back into improved models. Without that integration, these capabilities remain disconnected. With it, the space becomes navigable.
Clinical Proof Already Exists
The clinical proof of concept already exists. Lenalidomide, now standard of care in multiple myeloma, does not inhibit its targets. It recruits them to an E3 ligase and induces their degradation. At its peak it generated nearly $13 billion in annual sales, reflecting a global patient population treated over two decades with a mechanism that classical drug discovery frameworks would never have prioritized. In clinical trials, it nearly doubled progression-free survival after transplantation. That is not a footnote. It is a demonstration that proximity-induced pharmacology works in humans, at scale, against targets previously considered undruggable. The question the field is now asking is whether that logic can be extended systematically, rationally, and across the full range of disease-relevant biology that remains out of reach.
The Scale of Unmet Need
The scale of unmet need is significant. Checkpoint immunotherapy has transformed oncology. Today, roughly half of newly diagnosed cancer patients are eligible for PD-1 or CTLA-4 blockade. That is a genuine revolution. But it leaves the other half largely without a targeted immunotherapy option. And even among those who initially respond, many will relapse, driven by the same epigenetic resistance mechanisms that lock T cells into exhaustion, reprogram macrophages toward tumor permissiveness, or allow cancer cells to transdifferentiate entirely out of immune recognition. More than 53 million people are living with cancer globally. A substantial portion carry tumors driven by biology that checkpoint blockade was never designed to reach. That is not a failure of immunology. It is a target selection problem, and it is exactly the problem this toolkit is positioned to solve.
The most immediate opportunity is not replacement but combination. Checkpoint blockade works when the immune system can see the tumor and act on it. The new toolkit addresses both failure modes: epigenetic modulators can reverse the chromatin changes that lock T cells into exhaustion, restoring sensitivity in patients who have stopped responding; proximity-inducing compounds can remodel the microenvironment that shields tumors from immune recognition entirely. Neuroendocrine transdifferentiation, the escape mechanism that renders cancer cells invisible, becomes a candidate for intervention before it takes hold. The patients who need this most are not outside the immunotherapy era. They are stuck at its current boundary.
The remaining target space is large. The oncogenic drivers with no approved targeted therapy, MYC, mutant p53, the bromodomain proteins governing transcriptional addiction, the chromatin remodifiers that enforce tumor-permissive cell states, represent some of the highest-frequency alterations in human cancer. They have been validated by genetics, by functional studies, by decades of failed conventional drug discovery. The biology is not in question. What has been missing is the chemistry to reach it and the computational tools to navigate it rationally.
Where Capital Is Going
Capital is already moving in this direction. Bristol Myers Squibb's $14 billion acquisition of Karuna Therapeutics was not a bet on a conventional mechanism. It was a bet on the ability to design toward biology that classical approaches had failed to reach. That logic is becoming more common. The question investors are increasingly asking is not whether a company has a drug, but whether it has a platform capable of navigating complex target space. That is a different kind of value, and it requires a different kind of tool.
But the more important measure is not financial. It is the 53 million people currently living with cancer, the fraction whose tumors are driven by MYC, by mutant p53, by epigenetic programs that enforce exhaustion and permissiveness and escape. For a generation, those patients have had no targeted option not because the biology was unknown, but because the chemistry to reach it did not exist and the computational tools to navigate it were not available. Both of those conditions are changing simultaneously.
Chemical biology is providing the chemistry. AI is providing the navigation. The field is moving from access to control, and for the patients who have been waiting on the wrong side of that boundary, that transition is not academic. It is the difference between a diagnosis with options and one without.