The next breakthroughs in cancer will be in combinations, not single targets

Ask a pharma company what it is looking for and the answer is almost always the same: a novel target, ideally one with single-agent efficacy. That preference has ruled drug discovery since the beginning of rational drug design, and it still shapes how the industry allocates its money today.

It is easy to see why. A single target is tractable. One mechanism, one biomarker, one clean causal story to tell a regulator. And every so often it works spectacularly. Imatinib blocks the single BCR-ABL fusion protein that drives chronic myeloid leukaemia, and in the IRIS trial that defined the modern targeted era it produced a complete haematologic response in 95.3% of patients,1 turning a fatal leukaemia into a condition most people now survive for well over a decade.

The question is whether that model still serves us against the disease burden we actually carry now. An ageing population is accumulating conditions that are, almost by definition, not caused by any one thing.

The textbook criteria for a good target read like a checklist. It should sit in a pathway causally linked to the disease and to nothing else, ideally with human genetic evidence tying it to the condition. It should be expressed in diseased tissue and nowhere else. It should have a druggable surface, and it should be different enough from its relatives that you can hit it without hitting them. AstraZeneca formalised a version of this as its 5R framework after auditing its own pipeline (right target, right tissue, right safety, right patient, right commercial potential), and credited disciplined decisions across all five for reversing its clinical success rate.2

Biology does not organise itself to make that checklist easy to satisfy. Living systems are built on redundancy, feedback and failsafes, the same properties that make them robust to damage. Engineer a mouse to lack a gene and it often shows no defect at all, because a paralog quietly takes over the job. Delete Rpl22 and its paralog Rpl22l1 steps in;3 knock out one or even two of the three cyclin D genes and the remaining one covers for them.4 Tumours use the same trick against us: block one signalling route and the cell reroutes through another. There is no strong reason to expect a perfectly isolated, disease-only, non-redundant node to exist for most diseases, because that is not how the system was built.

The nodes that matter most are the ones evolution was most careful to back up.

Suppose you find one anyway. Diseases of ageing are rarely the output of a single broken part. Alzheimer’s is the clearest case: amyloid plaques, tau tangles, chronic neuroinflammation and vascular dysfunction interact and reinforce one another, and few researchers now believe any one of them fully explains progression.5 That is the backdrop to decades of amyloid-only programmes returning, at best, modest effects; even the anti-amyloid antibodies that finally cleared plaque slow decline by roughly a quarter rather than stopping it.6 The same shape, several interacting pathologies rather than one root cause, shows up across most of the diseases that accumulate with age.

So the more honest question may not be “where is the next great target?” We are largely standing on the shoulders of giants; much of the important target biology in a given disease has arguably already been found. What has not been solved is the expectation that any one of those targets, hit in isolation, should reverse a disease that never had a single cause. The work left to do lies in managing the relationships between the nodes we already have.

The industry has a name for its declining returns on research: Eroom’s Law, Moore’s Law spelled backwards. The number of new drugs approved per billion inflation-adjusted dollars of R&D has fallen by half roughly every nine years since 1950, even as the underlying science and computing power improved beyond recognition.7 A field that keeps insisting on single, clean, isolated mechanisms, in diseases that do not have them, is a plausible part of that gap between scientific progress and drug-approval productivity.

Pharma has not ignored any of this. It is visibly shifting toward dual- and multi-mechanism medicine. Bispecific antibodies went from a niche format to standard equipment, with fourteen approved by the end of 2023, and a first-in-class PD-1×VEGF bispecific, ivonescimab, has now beaten single-agent pembrolizumab head to head in lung cancer.8 The older proof points are already approved and have aged well. Combining two checkpoint inhibitors, nivolumab against PD-1 and ipilimumab against CTLA-4, extended median progression-free survival in advanced melanoma from 2.9 months on ipilimumab alone to 11.5 months together. Years on, that gap had widened into survival: a median of 72.1 months on the combination against 19.9 on ipilimumab, with roughly half the combination patients still alive at six and a half years.9 Anti-LAG-3 became the third checkpoint class to reach the clinic in combination, as relatlimab plus nivolumab, the first LAG-3 regimen approved and one that more than doubled progression-free survival over nivolumab alone.10

The same logic runs through the parts of oncology moving fastest right now. Antibody-drug conjugates carrying topoisomerase-1 payloads, the TROP2 conjugates datopotamab deruxtecan and sacituzumab govitecan and the HER2 conjugate trastuzumab deruxtecan, are being developed less as solo agents than as backbones. Sacituzumab govitecan with pembrolizumab lifted first-line progression-free survival in triple-negative breast cancer to 11.2 months against 7.8 for chemotherapy with pembrolizumab; datopotamab deruxtecan is in late-stage trials alongside the EGFR inhibitor osimertinib; trastuzumab deruxtecan is being paired with pertuzumab and with checkpoint blockade.11 KRAS G12C inhibitors needed partners too. Blocking the mutant protein alone triggers a feedback reactivation of EGFR, so sotorasib only worked in colorectal cancer once it was combined with the EGFR antibody panitumumab, roughly tripling progression-free survival over standard care.12 The direction of travel is clear. The hard part is no longer whether to combine, but which combination, in which order, for which patient. Regulators ask a version of the same question: each component of a fixed-dose combination has to earn its place under the contribution-of-components principle, and since 2013 the FDA has kept a narrow codevelopment pathway open for two novel agents that only work together.13

Combinations are harder to navigate than single agents precisely because of the complexity they are meant to address. The search space is enormous: a full pairwise screen of just 100 drugs at 100 doses each is on the order of 50 million experiments, and every extra drug, or every move to triplets and quadruplets, multiplies that again. Even the best physical screens have a deeper limit. BATCHIE, a Bayesian platform, recovered most of the synergies in a 206-drug panel while running only about 4% of its roughly 1.4 million wells,14 which is a genuine advance and still cells in a dish. A dish cannot see the immune compartment clearing or sparing a clone, the stroma that walls a drug out, how much of a dose actually reaches each metastatic site, or the order and timing in which agents are given. It will not tell you that a combination which clears the primary tumour reseeds from the marrow, or that an oncolytic fails in a liver metastasis because prior chemotherapy remodelled its vasculature. Those are the factors that decide a human outcome, and a physical screen leaves out almost all of them.

Add sequence, schedule and patient stratification on top, and the plausible regimens for a single cancer run into the billions. So physical screening was never going to be how we find these answers, and the enthusiasm for fully autonomous closed-loop labs makes us wary rather than hopeful: biology is not cheaply verifiable the way maths or software is, and automating a misleading assay mostly produces wrong answers faster.

The reason combinations matter is easiest to see by watching a tumour and its surroundings evolve. A single agent can shrink a cancer for a while, but it also selects: a resistant subclone that was a rounding error at baseline expands into the space the drug clears, and the disease returns. A well-chosen combination closes those exits, holding the resistant population down while the stroma normalises and the immune compartment does the durable work. Toggle between the two.

What the field needs is a model that can hold several mechanisms in mind at once and tie them to real human outcomes, so you know, before you go near a clinic, which combination will move the needle for which patient.

Most of the industry building trial-prediction models is chasing the newest architecture, foundation models and virtual cells today, classical machine learning a few years ago. Underneath, they all join the dots between past trials. That gives you a good model of how known drugs behave in familiar situations and very little ability to extrapolate to a novel drug, a new target, or a system pushed into a state it has not seen.

There is a comfortable tech-bio story that proprietary data solves this: the magic target is sitting somewhere in the pile, and a little more data plus scaling laws will make it appear. That is folly. What is scarce is not data but biological reasoning data, the kind that is separated in time, spans several sorts of interaction and ends in a verifiable outcome, and there is very little of it. More data earns its keep only when it is aimed at a specific question about how the system evolves.

The field today looks much like protein structure prediction did before AlphaFold. One camp pours in ever more data from static patient samples or highly artificial and isolated models (e.g. cancer culture cell lines) and waits for the pattern of the dynamic normalised tissue to fall out, which is the bet behind today’s biological foundation models. The other tries to reconstruct the whole system from the bottom up, the way physics-based methods once tried to compute a protein’s fold from the forces between every atom. The “virtual cell” programmes are the current version of that second camp.

AlphaFold did neither. It won by choosing the right representation of the problem. Solved structures were scarce and expensive, only a couple of hundred thousand in the Protein Data Bank, but sequences were everywhere, and every evolutionary homolog of a protein is a natural experiment: residues that sit close together in the fold tend to mutate in step, because a change in one is tolerated only if its partner changes too. AlphaFold read that evolutionary signal from sequence alignments and tied it to a three-dimensional, geometric representation of the chain.19 The cheap, abundant data constrained the space by analogy to what already works across biology, so the scarce, expensive data went a very long way.

Cancer is the same shape of problem. Our equivalent of AlphaFold’s abundant evolutionary homologs is the wide set of causal experiments, each one pinning down a single interaction in isolation, together with single-cell RNA measurements that fix a cell’s starting state. Clinical outcomes are scarce, but that mechanistic evidence is not. With the right mapping and alignment with clinical outcomes, those isolated results can be combined into a spatiotemporal view of how the interactions play out as one connected chain, in a given cell, across the cancer-stroma-immune interaction, across the body, from a given starting state, under a given intervention. So we never have to simulate every cell and every pathway. We model the state changes and knock-on effects that actually drive resistance, then optimise across that much smaller, decision-relevant set. One clean assay of bone penetration will tell you more about clearing prostate cancer than thirty thousand slides of immune access at the primary site.

Won’t frontier language models learn what is important eventually?

An off-the-shelf language model can already predict trial outcomes with close to 80% accuracy. Restrict it to the cases that actually matter, the surprising success, the true synergy, and that accuracy collapses toward zero. Those cases resist interpolation, precisely because, unlike verifiable domains such as maths, software engineering or even robotics, there is no training set or ‘gym’ that reasons through the underlying synergistic interactions to join the dots. A patient sample is usually a single point in time, and the rare temporal data, from blood, is a jumbled mix of everything happening at once. Very few papers work at the systems level; press coverage of a trial reports only whether it passed or failed at the target. Even buying and naively integrating pre-clinical data, which the frontier labs are now doing, would not necessarily surface the signal, because most experiments were never framed to show a synergistic interaction in the first place.

A language model has no framework for that kind of reasoning. Faced with a novel case it is pulled straight back into the reasoning that drove the last round of failures, because that is what sits in its training data. The harness we have spent the last year building does the opposite: it weighs every influential factor, from diffusion gradients inside a tumour to the variance in a receptor’s density across a population, scaled all the way through to the placebo rate at a particular trial site. That is what lets us predict exact endpoints to within weeks of overall survival.

There is a further reason existing outcomes are the wrong thing to learn from. Many of the most useful building blocks for a combination are the ones that failed on their own. The box is full of mechanisms of synergism and selectivity that never cleared pre-clinical because, in isolation, they do little, yet in the right pairing they beat any single target by a wide margin. A model trained on what has already read out in humans cannot see the worth of those pieces, so it quietly locks out the most desirable part of the search space.

Reasoning back to the driving node

Over the past year we built agents to unpick the earliest vulnerabilities of Alzheimer’s disease with the Allen Institute. Working from single-cell omics, they integrated across hundreds of smaller hypotheses, most of them far from the largest changes in the data, and converged on NMDA-receptor-expression-mediated hyperexcitability as a common early vulnerability.18 That subtlety is the whole point. Disease is driven by cell state, by network position and by how the system evolves, so the biggest change in a dataset is usually a downstream consequence, several steps from the lever that actually controls it. To intervene you have to reason back to the driving node. It is the lesson a decade of amyloid antibodies, aimed at a late and highly visible change, keeps teaching the hard way.

A world model, then a fast search over it

So we built the model differently. It plays the disease forward in time, step by step, across each metastatic site and the organs that matter, and across each group of patients in a trial, reasoning by analogy about how every pairwise interaction unfolds and bringing it all back together to ask whether a given trial design would actually hit its endpoint. This works because biology is largely conserved: a cell exposed to a given signal, absent competing factors, tends to respond in a knowable way, and the rules for scaling a dose from a dish to a human, or for why an oncolytic may fail after prior chemotherapy has remodelled the vasculature, can be reasoned about from first principles. It is the same reasoning that produced immunotherapy and the GLP-1 drugs, with some luck along the way; we taught a model to do it deliberately, at an enormous scale that can navigate the space of 16 billion combinations a human could never touch.

Prediction alone is not enough. The deep model is slow, about ten minutes a cycle, and there are billions of combinations to consider. So we distilled it into a fast approximation that runs in about 100 ms and paired that with a genetic optimiser that recombines the building blocks of a regimen, tests them against the fast model, learns, and tries again. Within minutes it returns the optimum combination for a given cancer, with or without stratification, the wider trade-off curve of survival against regimen complexity, and, most usefully, the two or three wet-lab experiments that would most reduce the remaining uncertainty. Those experiments sharpen the model’s grasp of this cancer, and of the systems biology underneath it.