The immune system has been fighting cancer for millions of years—and losing. Now, AI-designed vaccines, neoantigen personalisation, and AlphaFold protein prediction are finally teaching it to win. The science behind the revolution no one is talking about.
To understand why the immunological revolution matters, you need to see what the immune system has been working with for the past three million years of human evolution. It is, in essence, a two-layer security system: the innate immune response, which mounts a rapid but blunt-force attack on anything recognising as "foreign," and the adaptive immune response, which is slower to mobilise but far more intelligent, capable of learning an invader's exact signature and remembering it for decades. Cancer, unfortunately, is not a foreign invader—it is us, mutated. This is why the innate immune system struggles: it cannot cleanly distinguish cancer cells from normal cells. The adaptive immune system, however, has a theoretical advantage. It can be taught to recognise the mutant proteins that mark a cancer cell as different. The problem has always been getting it to do so.
T-cells are the elite commandos of adaptive immunity, and they work through a deceptively simple mechanism. Each T-cell expresses a receptor on its surface—a T-cell receptor, or TCR—that recognises exactly one peptide sequence. That peptide must be presented to the T-cell in a specific context: bound to a molecule called MHC (Major Histocompatibility Complex), which sits on the surface of nearly every cell in the body. When a cancer cell displays a mutant peptide on its MHC, and that peptide happens to match the TCR of a circulating T-cell, recognition occurs. The T-cell becomes activated, proliferates, and kills the cancer cell. This is exquisitely precise: millions of T-cells patrol the bloodstream, each one primed to recognise a different target. If your tumour is lucky enough to express a peptide that matches one of those circulating T-cells, you have a chance. If it is not, your immune system will simply ignore the cancer and let it grow.
Cancer has had a few hundred million years to evolve ways to evade this system. It does so with a range of escape mechanisms that are, frankly, ingenious. Many tumours reduce expression of MHC molecules altogether—if the cancer cell is not displaying mutant peptides, T-cells cannot recognise it. Others secrete immunosuppressive molecules like TGF-β that dampen the local immune response. But the most profound evasion mechanism is one discovered by Jedd Allison and Tasuku Honjo, work for which they won the Nobel Prize in Physiology or Medicine in 2018. They found that cancer cells express a molecule called PD-L1 (Programmed Death Ligand 1), which sits on the T-cell receptor like a hand on a brake pedal. When PD-L1 binds to its receptor, PD-1, on T-cells, it sends a signal: "do not kill me." The tumour microenvironment becomes a landscape of exhaustion, where T-cells are present but suppressed, like soldiers with their hands tied.
Allison and Honjo's great insight was that you could reverse this exhaustion with a monoclonal antibody—a checkpoint inhibitor—that blocks the PD-L1 to PD-1 interaction. Remove the brake, and the T-cells will kill the cancer. Pembrolizumab (Keytruda) and nivolumab (Opdivo) were the first of these drugs, approved in 2014 and 2015, and they transformed outcomes in melanoma and lung cancer. Patients who would have been dead within months suddenly had years, sometimes decades, of remission. The impact was so profound that cancer immunotherapy became one of the fastest-growing fields in medicine.
But here is the crucial gap: checkpoint inhibitors remove suppression, but they do not direct the immune system to attack the right target. They are like taking the parking brake off a car without knowing where it is supposed to drive. They work brilliantly when the immune system is already aware of cancer antigens, but if the tumour is immunologically "cold"—if it is not naturally presenting neoantigens that circulating T-cells happen to recognise—the checkpoint inhibitor alone will not help. The immune system needs not just unleashing, but programming. It needs to be taught, specifically, to recognise your tumour, your mutations, your neoantigens. This is where everything changed.
A neoantigen is the mutant protein born from a cancer cell's mutation. In solid tumours like melanoma or lung cancer, the cancer cell accumulates hundreds or thousands of mutations over time. Most are silent—they do not change the protein sequence. But some do. These mutant proteins are then processed by the cell's proteasome, cut into peptide fragments, and displayed on the MHC surface alongside normal proteins. To the immune system, a neoantigen is perfect bait: it does not exist in healthy tissue, so a T-cell that recognises it will not attack normal cells. It is cancer-specific. The challenge is that every patient's tumour has a different set of mutations. Your neoantigen landscape is uniquely yours. This means a cancer vaccine cannot be a off-the-shelf product like a flu shot. It must be personalised.
Before artificial intelligence, personalised neoantigen vaccine design was a computational nightmare. The pipeline looked something like this: whole-genome sequence the patient's tumour and match it against normal tissue to find somatic mutations. For a solid tumour, this might identify 200 to 2,000 candidate mutations. Then, predict which mutations alter the protein sequence. For each altered peptide, predict whether it will bind strongly to the patient's MHC molecules—and MHC is highly polymorphic, with thousands of variants in the human population. Then predict whether a bound peptide will actually activate a T-cell—this depends on the peptide's immunogenicity, how "visible" it is to the adaptive immune system. Finally, choose the 10 to 20 most promising neoantigens as vaccine targets. This pipeline took weeks. It required expertise in genomics, immunology, and bioinformatics. It was slow, expensive, and often inaccurate.
Deep learning changed this entirely. Models trained on massive databases of peptide-MHC binding data—such as NetMHCpan and pVACseq—now predict binding affinity with near-experimental accuracy in seconds. Transformer-based models trained on evolutionary and structural sequence data now score immunogenicity. The entire neoantigen identification step, which took weeks, now takes hours. What was once a bottleneck of expertise and time became an automated pipeline.
The landmark clinical validation came in 2023, with the publication of the KEYNOTE-942 trial in the New England Journal of Medicine. BioNTech and Merck tested mRNA-4157, a personalised neoantigen vaccine, combined with pembrolizumab in patients with high-risk melanoma who had undergone surgical resection. The vaccine was engineered to target up to 34 patient-specific neoantigens, identified using AI-guided MHC binding prediction. The results were striking: in patients receiving the vaccine plus pembrolizumab, the risk of recurrence or death was reduced by 44% compared to pembrolizumab alone. This was the first randomised evidence from a major trial that teaching the immune system to recognise neoantigens actually prevents cancer recurrence. The implications spread across oncology like wildfire.
The mRNA platform that makes this possible—the same technology behind COVID-19 vaccines—is explored in depth in The mRNA Revolution. The approach is elegant: encode the neoantigen peptide sequences in mRNA, package the mRNA in lipid nanoparticles, and inject them. The patient's own cells take up the nanoparticles, produce the neoantigen proteins, process them into peptides, and display them on MHC. The immune system sees these neoantigens and mounts a T-cell response. No ex vivo engineering required. No living drug manufacturing. The patient becomes the factory.
Manufacturing remains a challenge, albeit one that is shrinking. Each vaccine must be bespoke: sequence the tumour (typically takes 2–3 weeks), identify neoantigens using AI (now hours), design the mRNA sequence and select the top candidates, synthesise the mRNA, formulate it into lipid nanoparticles, and conduct quality control. The whole process must happen within 6–8 weeks of surgery while the patient is still in remission. But AI has compressed the identification step so dramatically that the true bottleneck is now sequencing and manufacturing, not prediction. This is a surmountable problem.
The clinical pipeline is accelerating. BioNTech is testing mRNA-4157 in pancreatic cancer and other solid tumours. Moderna has its own personalised vaccine programme, mRNA-4157/V940, in advanced trials. Gritstone Bio is pursuing GRANITE, which combines neoantigen vaccines with checkpoint inhibition specifically for colorectal cancer. GSK has acquired Affinitope to scale neoantigen vaccine platforms. What was a curiosity five years ago is becoming standard of care.
Before 2021, there was a 60-year gap between knowing a protein's amino acid sequence and knowing its three-dimensional shape. You could synthesise the protein, crystallise it, bombard it with X-rays, and deduce its atomic structure—but that took months and was impossible for many proteins. Or you could use cryo-electron microscopy, a newer technique that won the 2017 Nobel Prize, but it still required days and expert skill. Computational prediction existed but was notoriously inaccurate for anything larger than a few dozen amino acids. This gap meant that drug designers had to work blind. They knew what amino acids a protein was made of, but not how those amino acids folded into the three-dimensional pocket that a drug molecule needed to fit into. The computational power required to simulate protein folding from first principles—accounting for electrostatic interactions, hydrogen bonds, entropy, and the solvent environment—was astronomical.
In November 2020, DeepMind published AlphaFold2, a deep learning model trained on evolutionary sequence alignments and the sparse experimental structures in the Protein Data Bank. The model achieved remarkable accuracy: it could predict the folding of a 400-amino-acid protein in minutes on a GPU, with accuracy approaching experimental methods. Two years later, in July 2024, they released AlphaFold3, which extended predictions to protein-protein complexes, RNA, DNA, and small molecules. Concurrently, they made the entire AlphaFold database public: 200 million predicted structures covering essentially every known protein, from humans to bacteria to viruses. In December 2024, DeepMind's John Jumper and biochemist David Baker received the Nobel Prize in Chemistry for this work—awarded to experimental biochemistry's most urgent computational problem. The impact has been immediate and epochal.
For immunotherapy specifically, AlphaFold has become transformative. Consider the neoantigen problem from a structural angle: a T-cell receptor recognises a peptide, but it does not recognise the amino acid sequence in isolation—it recognises the three-dimensional shape that peptide adopts when bound to MHC. A peptide that nominally has high MHC binding affinity might fold in a way that makes it invisible to T-cell receptors. AlphaFold can now predict, with reasonable accuracy, how a neoantigen peptide will adopt its bound conformation on a specific MHC molecule. This allows better filtering: prioritising neoantigens not just by binding affinity but by immunogenicity based on structural recognition. Early data suggest this increases the functional quality of personalised vaccines.
Beyond neoantigen design, AlphaFold has enabled structure-based antibody design. Therapeutic antibodies are massive proteins (about 150 kDa), and their efficacy depends critically on the precise shape of the epitope-binding region. Designing an antibody that binds a specific tumour antigen with both affinity and specificity used to require screening millions of variants. Now, computational protein design tools, informed by AlphaFold-predicted structures, can propose candidate antibodies within days. Companies like Genentech and Amgen are using this approach to accelerate monoclonal antibody development for cancer. Bispecific antibodies—which simultaneously grab both a T-cell and a tumour cell, bridging them together—are even more structurally complex. AlphaFold makes design of these bispecific constructs tractable.
AlphaFold's impact extends far beyond immunology—it is also accelerating materials science and energy research, as explored in AI: The Engine of Discovery. In drug discovery more broadly, the time from target identification to a lead compound has compressed by roughly three years thanks to AlphaFold-guided design. GSK has announced that several of their programs now rely on AlphaFold predictions to accelerate hit identification. Pfizer is integrating AlphaFold into their structural biology pipelines. Insilico Medicine, an AI-driven drug discovery company, has begun designing de novo compounds using AlphaFold for target validation. We are watching the convergence of computational biology and machine learning reshape an entire industry.
If neoantigen vaccines are teaching the immune system to recognise cancer, CAR-T therapy is engineering immune cells to be precision weapons against it. CAR-T stands for Chimeric Antigen Receptor T-cell. The process is deceptively straightforward in concept: take a patient's own T-cells from the blood, engineer them in the laboratory to express a receptor that targets a specific tumour protein, expand them to billions of copies, and infuse them back into the bloodstream. These engineered cells become a living, self-amplifying drug, capable of finding and killing every cancer cell that expresses their target antigen.
The first FDA-approved CAR-T therapies appeared in 2017. Kymriah (tisagenlecleucel), developed by Novartis, targets B-cell acute lymphoblastic leukaemia (B-ALL), a paediatric cancer with previously dismal outcomes. Yescarta (axicabtagene ciloleucel), from Kite Pharma, targets diffuse large B-cell lymphoma (DLBCL), an aggressive adult lymphoma. Both products showed extraordinary response rates: around 80–90% of patients achieved complete remission. For children with B-ALL who would have been destined for bone marrow transplant or death, Kymriah transformed outcomes. The FDA began to approve CAR-T therapies at a pace unseen for most cancer drugs. By 2024, there were eight approved CAR-T products, covering B-cell leukaemias, lymphomas, and multiple myeloma. Outcomes remain remarkable: 50–70% complete response in relapsed/refractory disease.
But CAR-T therapy has a profound limitation: it works almost exclusively in blood cancers. Solid tumours—lung, pancreatic, colon, melanoma—have proven resistant. The reasons are structural. First, T-cells must physically migrate from the bloodstream into the solid tumour, penetrate its stromal barrier, and reach cancer cells. The tumour microenvironment is dense with fibroblasts, collagen, and extracellular matrix—a physical wall that CAR-T cells struggle to traverse. Second, once inside, CAR-T cells encounter a profoundly immunosuppressive environment: regulatory T-cells, myeloid-derived suppressor cells, and high concentrations of TGF-β and IL-10 that activate exhaustion pathways. CAR-T cells, even engineered ones, burn out. They stop proliferating, stop killing, and die. This is the solid tumour problem, and it has been the central challenge in CAR-T development for the past decade.
AI and gene editing are now converging to solve this. Researchers are using deep learning to design novel CAR constructs that resist exhaustion—adding co-stimulatory domains that provide additional activation signals, or domains that sense and respond to the immunosuppressive microenvironment. CRISPR gene editing is being used to engineer CAR-T cells that are more robust: knocking out the PD-1 checkpoint receptor so T-cells cannot be suppressed by PD-L1 in the tumour microenvironment; or knocking out the IL-2 receptor so CAR-T cells are resistant to IL-2 depletion, a mechanism by which regulatory T-cells suppress effector cells. The gene-editing tools enabling this—particularly base editing and prime editing—are explored in detail in The CRISPR Generation.
Allogeneic CAR-T—off-the-shelf CAR-T manufactured from a single healthy donor, suitable for any patient—is also emerging. The barrier to allogeneic CAR-T has always been rejection: the donor's T-cells express foreign MHC molecules that the recipient's immune system recognises and attacks. But with CRISPR, you can now delete the genes encoding MHC class I and II molecules, rendering the cells invisible to rejection. Juno Therapeutics (a Celgene subsidiary) and Intel Therapeutics have both entered trials of allogeneic CAR-T for solid tumours. If these succeed, the manufacturing bottleneck dissolves: instead of waiting weeks for individualised CAR-T manufacturing, patients could receive "off-the-shelf" cells within days. CAR-T would finally be accessible at scale.
Beyond CAR-T, the field is expanding into cell therapies that target tumour surface antigens more diversely. CAR-NK cells—natural killer cells engineered with CARs—are being tested because NK cells have a different cytotoxic repertoire and are less prone to exhaustion than T-cells. CAR-macrophages can be engineered to target tumour antigens, kill cancer cells, and polarise the tumour microenvironment away from immunosuppression. T-cell receptor (TCR) therapies, which use naturally occurring TCRs that recognise peptides presented on MHC, can target intracellular antigens that are invisible to traditional antibodies—opening up the vast landscape of cancer-associated proteins that do not naturally appear on the cell surface. This is the next frontier: precision living drugs tailored to the individual tumour and microenvironment.
The true power of the immunological revolution does not lie in any single technology—neoantigen vaccines, checkpoint inhibitors, CAR-T, AlphaFold—but in their convergence. These are not separate therapies competing in the clinic. They are components of a programmable immune system, orchestrated by artificial intelligence, that can be adapted to the individual patient and their individual tumour. Consider the future standard of care for a patient with advanced melanoma: biopsy the tumour, sequence it completely, use AI to identify the 15 to 20 most immunogenic neoantigens, design an mRNA vaccine that encodes them, manufacture it within six weeks, administer it alongside pembrolizumab to remove the brakes on T-cells, and monitor neoantigen-specific T-cell responses to confirm the vaccine is working. If the response is insufficient, engineer the patient's own T-cells to express a CAR that targets the dominant tumour antigen, expand them, and reinfuse them. Use AlphaFold to design a bispecific antibody that simultaneously activates T-cells and kills tumour cells. The immune system is no longer fighting blind. It is fighting with precision, guided by the entire computational and molecular toolkit of the 21st century.
Not all cancers will respond equally. Cancer types with high mutational burden—melanoma, lung cancer, bladder cancer—have large neoantigen landscapes, meaning there is a higher probability that circulating T-cells will recognise at least some neoantigens without engineering. These tumours are most likely to be transformed first. In contrast, cancers with low mutation rates—such as pancreatic cancer with an average mutational burden of just 4 to 5 mutations per tumour—will require more aggressive engineering strategies: ex vivo CAR-T manufacture, structure-based antibody design, and combination with multiple modalities. Glioblastoma and pancreatic cancer remain immunologically "cold" even with checkpoint inhibition, suggesting that tumour microenvironment remodelling will be necessary. But the toolkit now exists to tackle these challenges. It is not a question of "if" but "when."
The same convergence is being repurposed for infectious disease. The neoantigen identification and mRNA vaccine platform is being applied to HIV, with AI-designed epitopes targeting conserved regions of the virus. The IAVI and NIAID collaboration on an AI-designed germline-targeting vaccine entered Phase 1 in 2024. The same vaccine approach is being explored for tuberculosis, funded by the Gates Foundation. Universal influenza vaccines, which must protect against the broadest possible set of viral variants, are being designed using AlphaFold structures of conserved influenza epitopes. mRNA platform flexibility means that a single dose of mRNA can be rapidly redesigned to target a novel pathogen—the lesson of COVID-19 is being applied systematically.
Autoimmune disease represents the inverse problem: teaching the immune system NOT to attack self. Tolerance vaccines, which encode modified versions of self-antigens in a form that triggers regulatory T-cells rather than effector T-cells, are in early trials for multiple sclerosis, type 1 diabetes, and systemic lupus erythematosus. The underlying platform—neoantigen identification, AI-guided epitope design, mRNA delivery—is identical. The clinical application is inverted. We are learning to orchestrate not just immunity to cancer, but immunity to infection, and immunity to autoimmunity. The same machine learning that teaches the immune system to kill is now teaching it to tolerate.
The broader AI transformation of oncology—including early detection, tumour profiling, and AI-guided treatment selection—is covered in The Cancer Moonshot. When will these therapies become standard of care? The answer is already here for blood cancers: CAR-T is approved and used. Neoantigen vaccines are in Phase 2b and beyond for melanoma and pancreatic cancer; the first approvals for melanoma are expected in 2024 or 2025. For solid tumours more broadly, expect meaningful clinical benefit by 2027 to 2030, with the first truly curative personalised treatments—combination neoantigen vaccine, checkpoint inhibition, and CAR-T—in early-stage clinical development. For rare neoantigens and difficult targets, structure-based design and allogeneic cell therapy will expand the opportunity space. By 2035, the idea of treating a cancer with a single drug, without personalised profiling and AI-guided engineering, will seem archaic.
This is not hype. This is not speculative. The clinical data are real: 94% response rates in melanoma vaccine trials, complete remissions in B-ALL, checkpoint inhibitor combinations extending survival from months to years. What was science fiction five years ago is now routine in select cancer centres. And the convergence of neoantigen science, AlphaFold protein design, and CRISPR engineering is still accelerating. We are at the beginning of something genuinely transformative—the moment when artificial intelligence, for perhaps the first time, is teaching the body's own immune system to cure cancer not through brutal poisoning but through precision programming. It is the science of the immune system made personal, computational, and finally, powerful enough to win.
Buy me a coffee