Why Cancer Drugs Work—and Why Promising Biology Often Falls Short

A field guide from cancer genetics and immune recognition to drug design, clinical evidence, and biotech valuation.

An oncology presentation can be persuasive long before it is conclusive. A tumor carries a mutation. A drug binds a target. A biopsy shows more T cells. Each observation matters. None, on its own, establishes that patients benefit.

The useful question is what connects these observations. Does the mutation create a dependency? Does the drug reach the relevant cells at a tolerable exposure? Does an immune response reach the tumor, recognize it, and persist? Does the resulting biological activity improve the outcome the trial was designed to measure?

This essay follows that chain through 24 learning units and eight worked cases. The aim is to make oncology evidence easier to reason about, including what it can—and cannot—tell us about a drug asset's value. Numerical cases below are teaching hypotheticals, not reported results for named products.

1. Cancer is an evolving population

DNA is transcribed into RNA and translated into protein. Cancer can alter this system through point mutations, insertions and deletions, copy-number changes, rearrangements, gene fusions, and changes in gene regulation. A tumor is therefore more than a list of misspelled genes.

Somatic alterations arise in a subset of the body's cells. Germline alterations may be inherited and present across many tissues. Calling cancer a genetic disease does not mean that every cancer is inherited; finding an alteration in tumor sequencing does not establish its germline origin.

Nor are all cancer cells identical. Different clones can coexist within a lesion and across metastases. Treatment changes their relative fitness. A resistant population that was initially rare may expand after sensitive cells disappear. This evolutionary perspective explains why an effective drug can be followed by progression without having been biologically inactive at the outset. NCI: What Is Cancer?

2. Oncogenes and tumor suppressors change the rules of growth

An activated oncogene can increase a growth or survival signal. EGFR, KRAS, BRAF, and MYC provide familiar examples. Tumor suppressors constrain proliferation, coordinate damage responses, or maintain cellular control: TP53, RB1, PTEN, and APC illustrate different parts of that job.

The accelerator-and-brake analogy is useful, provided we remember its limits. Many tumor suppressors follow a two-hit pattern, but dominant-negative effects and haploinsufficiency complicate the picture. DNA repair genes such as BRCA1 and BRCA2 also connect tumor suppression to a potential therapeutic vulnerability. A mutation's functional effect matters more than its presence alone. American Cancer Society: genes and cancer

3. A driver is not the only useful kind of target

A driver alteration gives a clone a selective advantage in a particular context. A passenger accompanies expansion without an established driving role there. High frequency or high expression alone does not prove causality.

Some tumors become strongly dependent on an oncogenic pathway—oncogene addiction. A targeted inhibitor may exploit that dependence, although alternative survival programs can still limit durability.

Other modalities need something different. An antibody–drug conjugate may use a surface protein as a delivery address even if the tumor does not depend on its signaling. A vaccine may target a passenger-derived neoantigen if it is expressed, presented, and immunogenic. Target quality is therefore inseparable from the job the drug is supposed to perform.

4. A mutation must become a recognizable antigen

A mutation is only the beginning of the neoantigen pipeline. The altered gene must be expressed, its protein processed into peptides, an appropriate peptide loaded onto HLA, and the resulting complex recognized by a suitable T-cell receptor.

Predicted HLA binding is not proof of presentation. Presentation is not proof of immunogenicity. Central and peripheral tolerance shape the available TCR repertoire; similarity to self can matter, but counting amino-acid differences is not a reliable shortcut to immunogenicity.

Clonal neoantigens may cover more cancer cells than subclonal ones. Multi-antigen designs can reduce reliance on a single target. Yet sampling may miss heterogeneity, and tumors can lose antigen expression or presentation under selection. A good design follows the antigen through the entire process rather than stopping at a sequencing result.

5. HLA determines what can be shown; T cells determine what is recognized

MHC is the major histocompatibility complex; the human system is called HLA. Different HLA alleles present different peptide repertoires, which is why the same mutation may not be equally targetable across patients.

Class I MHC on most nucleated cells presents peptides for recognition by CD8 T cells. Class II MHC, principally on professional antigen-presenting cells, presents peptides for CD4 T cells. Dendritic cells can cross-present material acquired from tumors through class I, helping initiate CD8 responses.

CD8 cells can kill through mechanisms including perforin and granzymes. CD4 helper cells support dendritic-cell licensing, CD8 responses, and memory, including through CD40L–CD40 signaling. But CD4 is not synonymous with help: regulatory T cells are also part of the CD4 compartment.

6. TMB, dMMR, MSI-H, and B2M describe different things

Tumor mutational burden, or TMB, is commonly expressed as mutations per megabase. The number depends on the assay and counting rules. More mutations may create more candidate neoantigens; they do not guarantee a functional immune response.

Its interpretation as a biomarker depends on the disease, treatment, and testing context. NCI: TMB

Mismatch repair corrects replication errors. MLH1, MSH2, MSH6, and PMS2 are central proteins. Deficient mismatch repair, dMMR, describes a repair defect. Microsatellite instability-high, MSI-H, describes instability in repetitive DNA sequences. The two are closely related, but they are not interchangeable measurements, and testing can be discordant.

B2M, beta-2-microglobulin, helps stabilize classical MHC-I. Its loss can compromise tumor recognition by CD8 cells even when mutation burden is high. Patient research has linked acquired resistance to PD-1 blockade with defects in antigen presentation and interferon signaling, including B2M and JAK1/2 alterations. Zaretsky et al., 2016

This is not a claim that B2M loss eliminates every immune route. NK-cell biology differs, and conventional surface-targeting CAR-T cells do not require tumor peptide–HLA recognition. The relevant conclusion is modality-specific.

7. The cancer-immunity cycle has several points of failure

Antigen release must lead to uptake and presentation, T-cell priming, expansion, trafficking, tumor entry, recognition, and killing. Killing can release additional antigens and sometimes broaden the response through epitope spreading.

This is why blood-based T-cell expansion is informative but incomplete. We also want to know whether those cells reach cancer nests, recognize their intended target, and remain functional. Durable memory is attractive, particularly when residual disease is small, but even a persistent immune response is not a substitute for a clinical endpoint.

8. “Hot” and “cold” are starting descriptions

An inflamed or hot tumor contains an existing immune response, often with infiltrating T cells and inflammatory signaling. An immune desert lacks an effective response. An immune-excluded tumor can contain many T cells in surrounding stroma while keeping them out of cancer nests.

These are simplified phenotypes, not permanent categories. Different lesions in the same person may differ. A hot tumor can still resist killing through metabolic stress, dysfunctional T cells, or defective presentation.

PD-L1 also needs context. It may be driven by tumor-intrinsic signaling or induced by IFN-γ during immune attack. The latter is adaptive immune resistance: a response to immune pressure that can simultaneously indicate engagement and constrain its effectiveness.

9. The microenvironment is a community, not a single brake

Cancer-associated fibroblasts, or CAFs, can shape extracellular matrix, fibrosis, and signaling. Tumor-associated macrophages, TAMs, occupy diverse functional states. Myeloid-derived suppressor cells, MDSCs, can restrict T-cell activity through mechanisms including arginine depletion and reactive oxygen or nitrogen species. Regulatory T cells, Tregs, maintain tolerance and can suppress antitumor responses.

These populations should not be reduced to uniform villains. CAFs are heterogeneous, and the M1-good/M2-bad macrophage shorthand is too crude for drug development. High abundance may be prognostic without identifying a causal therapeutic bottleneck.

A study linking stromal TGF-β activity to T-cell exclusion and poor response to PD-L1 blockade illustrates how human associations and model experiments can build a hypothesis. It does not establish that every TGF-β inhibitor will improve patient outcomes. Mariathasan et al., 2018

10. Immune cells need access, oxygen, and fuel

VEGF-associated abnormal vasculature can impair perfusion and immune-cell access. Antiangiogenic treatment may create a normalization window; excessive vascular suppression can also worsen delivery. “Less VEGF” is not a complete pharmacological objective.

Chemokines are equally context-dependent. CXCL9/10–CXCR3 can support effector-cell recruitment. CXCL12–CXCR4 influences positioning and stromal interactions. CCL2–CCR2 recruits monocytes. CXCL is a family name, not a single biological instruction.

Hypoxia, lactate and acidity, competition for glucose, arginine depletion, and adenosine production associated with CD39/CD73 can all constrain immune function. Releasing PD-1 inhibition does not automatically repair these conditions.

11. Checkpoints differ in biology and clinical evidence

PD-1 is an inhibitory receptor found on immune cells. PD-L1 is one of its ligands; PD-L2 is another. Pembrolizumab targets PD-1. Blocking a receptor and blocking one ligand are related but distinct interventions. NCI: Pembrolizumab

CTLA-4 shares CD80/CD86 ligands with the costimulatory receptor CD28 and participates in regulatory T-cell biology. “CTLA-4 acts during priming; PD-1 acts during execution” is a helpful approximation with substantial overlap. Broader immune activation can increase immune-mediated toxicity. NCI: checkpoint inhibitors

LAG-3 and TIGIT regulate immune function through different molecular systems. TIGIT involves T/NK-cell signaling and a ligand network that includes competition with CD226. High expression during chronic stimulation does not, by itself, establish that blockade will restore useful function.

The randomized relatlimab–nivolumab study in advanced melanoma provides clinical evidence for combined LAG-3 and PD-1 inhibition in that setting. It does not give every emerging checkpoint the same evidentiary standing. RELATIVITY-047

12. An mRNA cancer vaccine is a complete production-and-immunity system

An individualized workflow can include tumor and normal sequencing, RNA analysis, HLA typing, antigen selection, manufacturing, and product release. Delivered mRNA supplies antigen instructions that must be translated into an effective immune response.

The design problem includes CD4 and CD8 responses, clonal coverage, persistence, and immune escape. The operational problem includes sample adequacy, turnaround time, batch consistency, and whether the patient can receive treatment within a useful window.

An adjuvant setting may provide lower tumor burden and time for immune control. It also includes patients already cured by local treatment and patients with biologically difficult residual clones. Lower burden is a rationale, not a guaranteed increase in success probability.

The randomized phase IIb KEYNOTE-942 melanoma trial and early personalized RNA-vaccine work in resected pancreatic cancer offer different levels and types of evidence. In the latter, comparing immune responders with nonresponders cannot isolate the vaccine's causal effect on survival. KEYNOTE-942, Rojas et al., 2023

13. A bispecific must earn the value of its architecture

Monoclonal antibodies can block receptors or ligands and recruit Fc-mediated functions. Bispecifics have two binding specificities, but that description covers very different jobs.

A conventional CD3 × surface-antigen engager connects T cells to target cells, generally without requiring tumor peptide–HLA presentation. Peptide–HLA-directed engagers are an important exception. PD-1 × VEGF or PD-1 × TGF-β designs instead seek to address immune inhibition alongside a microenvironmental constraint.

The question is whether the architecture improves localization, avidity, geometry, conditional activity, or exposure. Two arms in one molecule also make independent dose adjustment harder. Superiority to one monotherapy is not proof of superiority to an appropriate two-antibody combination. Stronger CD3 binding can increase systemic activation and cytokine-release syndrome rather than improve usable efficacy.

14. ADCs require more than a recognizable address

An antibody–drug conjugate combines an antibody, linker, and payload. Drug-to-antibody ratio, DAR, influences how much payload is attached, but higher DAR can also change aggregation, clearance, and toxicity.

For many ADCs, binding is followed by internalization, intracellular trafficking, and processing that releases an active payload. Endosomal recycling can produce a different result from effective lysosomal delivery. Some designs also use extracellular release; one trafficking model does not cover the entire class.

A membrane-permeable payload can produce bystander killing, potentially helping with antigen heterogeneity while widening exposure beyond the intended cell. Resistance can arise through antigen loss, altered uptake or lysosomal processing, drug efflux, and payload-related changes.

T-DXd's randomized HER2-low breast-cancer results show that delivery design can expand the relevant biological setting. The study also illustrates the importance of toxicity, including interstitial lung disease. Neither low target expression nor an ADC label guarantees a class-wide outcome. DESTINY-Breast04

15. CAR-T and TCR-T enter through different doors

CAR-T cells typically recognize a surface antigen through an engineered receptor containing intracellular activation and costimulatory elements. TCR-T cells recognize a peptide–HLA complex and can therefore target peptides derived from intracellular proteins, at the cost of HLA restriction and dependence on presentation.

Cell therapies also require manufacturing, release testing, preparative treatment where applicable, and management after infusion. In solid tumors, target overlap with normal tissues, heterogeneous expression, limited access, local suppression, and persistence can all limit performance. On-target normal-tissue injury, CRS, and neurotoxicity are distinct concerns. NCI: T-cell transfer therapy

A KRAS G12D-directed TCR case in pancreatic cancer illustrates feasibility and a mechanism of recognition, not a population response rate. Logic gating, dual targeting, local activation, and affinity tuning are development strategies whose benefits still need validation. Leidner et al., 2022

16. KRAS inhibition and the return of pathway signaling

The familiar pathway runs from receptor tyrosine kinases through RAS, RAF, MEK, and ERK. Phosphorylated ERK, p-ERK, is an important pharmacodynamic readout. Its interpretation depends on sampling time, assay, and cellular composition.

Suppression suggests pathway modulation. Rebound can suggest reactivation, but does not identify the cause. Target mutations, KRAS amplification, upstream feedback, bypass pathways, downstream alterations, and cell-state or histological changes can contribute. Patient studies document this diversity after KRAS G12C inhibition. Awad et al., 2021

Other networks matter too. PI3K–AKT–mTOR coordinates survival and metabolism, with PTEN acting as a restraint. WNT/β-catenin can connect to immune exclusion in particular contexts. VHL–HIF dysregulation in clear-cell renal-cell carcinoma links genetics to angiogenic biology.

Targeted therapy may improve immune engagement, but broad pathway suppression can also affect T cells. The deepest p-ERK reduction is not necessarily the best combination profile.

17. DDR, HRD, and PARP reveal a conditional vulnerability

DNA damage response, DDR, is a network. Homologous recombination, HR, is one repair process; HRD denotes its impairment. BRCA1/2 participate in HR, but HRD is not synonymous with any BRCA variant.

PARP inhibition involves catalytic inhibition, PARP trapping, and replication-associated damage. HR-deficient cells may be less able to tolerate that damage, creating synthetic lethality: two defects that are tolerable separately become lethal together.

A genomic scar can record past HR deficiency without proving that repair remains deficient now. BRCA reversion, restored HR, enhanced replication-fork protection, efflux, and PARP-related changes can produce resistance. A strategy aimed at fork protection should therefore be matched to evidence of that mechanism, rather than applied to all PARP-resistant disease.

18. DNA damage can engage innate immunity, but the arrows are conditional

Cytosolic DNA can activate cGAS, which produces cGAMP and engages STING. Downstream signaling through pathways including TBK1/IRF3 can induce type I interferons such as IFN-α and IFN-β. These should be distinguished from type II interferon, IFN-γ, commonly produced by T and NK cells.

Dendritic-cell activation can connect innate sensing to adaptive immunity. A BRCA1-deficient ovarian-cancer mouse study supports a PARP–STING link; it does not prove broad clinical benefit from PARP–checkpoint combinations. Ding et al., 2018

STING function, dendritic-cell competence, antigen presentation, and local suppression can each interrupt translation. Tumor-cell and host-cell signaling are not interchangeable. Acute and chronic interferon states can also have different consequences.

19. Myeloid remodeling is about function as well as abundance

CD47–SIRPα signaling can restrain macrophage phagocytosis. Removing that restraint may help when appropriate pro-phagocytic signals are present. Normal red cells also express CD47, making anemia relevant to some designs. Molecules with different binding, Fc properties, and dosing should not be treated as identical. Early CD47-blockade clinical study

CSF1R influences monocyte/macrophage survival and differentiation; CCL2–CCR2 contributes to monocyte recruitment. Blocking one route may permit compensatory myeloid recruitment through others.

Depletion reduces cells. Reprogramming changes their state. Neither is automatically superior. We need to examine phagocytosis, antigen presentation, cytokines such as IL-12, CD8 function, and eventually clinical outcomes. An impressive reduction in TAM count is not itself evidence of patient benefit.

20. PK/PD and the therapeutic window determine whether biology is usable

Pharmacokinetics asks what the body does to the drug. Pharmacodynamics asks what the drug does to biology. Plasma exposure does not guarantee sufficient free drug in the tumor, and a single biopsy does not establish sustained inhibition across a dosing interval.

The therapeutic window is the range between effective and unacceptably toxic exposure. It is not response rate divided by adverse-event rate. Dose interruptions, reductions, discontinuations, serious events, and quality of life all affect usable treatment.

Likewise, a combination outperforming either component does not establish more-than-additive synergy. Different patients may benefit from different components. A claim about interaction requires an appropriate model and comparison; a clinical development claim requires evidence that the incremental benefit justifies the incremental burden.

21. Biomarkers must be named by the question they answer

A prognostic biomarker identifies differences in outcome risk. A predictive biomarker identifies variation in a treatment's effect relative to its comparator. A pharmacodynamic biomarker tracks a biological response to intervention.

Even a familiar label such as PD-L1 requires an assay definition. TPS measures the proportion of positive tumor cells; CPS also includes specified positive immune cells in its numerator. Scoring method, assay, cutoff, and indication must travel with the reported number.

Circulating tumor DNA is the tumor-derived component of cell-free DNA. MRD describes small amounts of residual disease after treatment; in solid-tumor molecular testing it is often called molecular residual disease. A negative assay does not prove zero disease: shedding, analytical sensitivity, and timing matter. Clonal hematopoiesis and technical artifacts can complicate positive calls.

If ctDNA-positive and ctDNA-negative groups both have a treatment HR of 0.70, higher baseline risk in the positive group does not establish a stronger relative treatment effect. It may increase absolute benefit and event accrual. One subgroup reaching significance while another does not is also not proof of an interaction.

Prespecification, multiplicity control, reproducible assays, independent replication, and prospective validation strengthen a biomarker strategy. Enrichment can improve efficiency while increasing screening burden and narrowing applicability. ctDNA clearance is not a universally validated substitute for a patient outcome. FDA guidance on ctDNA in early-stage solid-tumor development

22. Endpoints measure different dimensions of benefit

These are not a universal hierarchy. Their value and regulatory interpretation depend on context. FDA: oncology trial endpoints

A hazard ratio compares instantaneous event hazards. Under an appropriate proportional-hazards interpretation, HR 0.70 corresponds to an approximately 30% lower hazard—not 30% longer life or 30% fewer events at every fixed time. Delayed separation and crossing curves may require additional summaries, such as milestone estimates or an appropriate restricted-mean analysis.

A confidence interval communicates uncertainty. A p-value is not the probability that the drug is ineffective. Power is the probability of detecting an assumed effect under a specified design, not the probability that the effect exists.

Censoring means observation ended without a recorded target event; it does not mean the patient will never have one. Informative loss to follow-up can bias estimates. Event count often matters more than enrollment alone, and “median not reached” can reflect short follow-up. Prespecified endpoints, multiplicity, analysis population, missing data, and clinically meaningful effect sizes belong in every interpretation.

23. Phase III tests whether the evidence survives a harder question

Single-arm phase II comparisons with historical controls can be distorted by selection, treatment era, assessment, and supportive care. Randomization helps, but small randomized trials remain vulnerable to noisy estimates and subgroup overfitting.

Phase-II-to-phase-III effect shrinkage can reflect winner's curse, regression to the mean, changing populations, or a stronger comparator. A move from HR 0.48 to 0.63 is not a definition of failure. Whether the prespecified phase III test succeeds, and whether the benefit–risk balance is useful, are the relevant questions. There is no universal shrinkage factor.

An independent data monitoring committee, DMC, operates under a charter. A recommendation to continue unchanged does not reveal the hazard ratio and may not follow an unblinded efficacy review at all. Early stopping, futility rules, conditional power, and sample-size adaptation depend on the actual design. Conditional power also requires assumptions about future data. FDA: DMC establishment and operation

Long-term immune control may involve memory and self-renewing T-cell populations; terminal exhaustion may not be reversible through another checkpoint alone. A survival plateau is worth examining, but sparse tails and censoring can mislead. Complete response, short follow-up, or a persistent immune signal alone cannot establish functional cure.

24. From clinical evidence to biotech valuation

A drug asset is worth the future net cash flows its owner can obtain, adjusted for uncertainty and timing. The model begins with eligible patients, uptake, treatment duration, net pricing, competition, patent life, and costs—not a mechanism label.

Risk-adjusted net present value, rNPV, combines discounted cash flows across probability-weighted scenarios. Development spending must be included at the times and probabilities at which it occurs. Peak sales are not asset value. If success probabilities or economic rights are already embedded in cash flows, multiplying them in again double-counts the adjustment.

A simplified bridge is:

Equity value ≈ asset portfolio value + cash − debt − corporate costs and obligations not already included.

Enterprise value ≈ market capitalization + debt − cash, with further adjustments when the capital structure requires them.

Cash supports development but may be consumed before the next milestone. Financing changes both cash and share count. Regional rights, royalties, milestones, shared costs, and profit splits cannot always be compressed into one ownership percentage.

Probability of success must also be defined: a primary endpoint, approval, and commercial adoption are different events. Mechanistic scores can organize uncertainty, but cannot mechanically produce a calibrated probability. Historical reference classes, relevant human evidence, trial design, and sensitivity analysis are more useful than false precision.

Eight cases to practice the chain

Case 1 — The vaccine with the bigger blood signal. In a hypothetical comparison, one program expands blood T cells twelvefold without downstream changes. Another expands them fivefold with tumor infiltration and falling ctDNA. The second provides a more connected biological story, but neither observation replaces randomized clinical evidence. Cross-cancer extrapolation still requires a new assessment of presentation, burden, and TME.

Case 2 — The PD-1 × VEGF construct. An abnormal vascular environment gives the combination a rationale. A higher ORR with short responses and more vascular toxicity would leave net benefit unresolved. Demonstrating architectural value requires more than beating one component alone.

Case 3 — The ADC with a convincing stain. Similar antigen staining can conceal different internalization and trafficking. At resistance, losing the address suggests a different problem from becoming insensitive to the payload. Switching antibodies while retaining a shared payload may leave cross-resistance intact.

Case 4 — The broadly expressed CAR target. A hypothetical target covers 95% of tumors but is also present in normal tissue; a peptide–HLA target fits 20% of screened patients. Coverage alone cannot choose the better program. The comparison needs normal-tissue risk, presentation, manufacturing, persistence, and clinically usable exposure.

Case 5 — The p-ERK rebound. Initial pathway suppression followed by rebound suggests renewed signaling. Tissue and ctDNA can help distinguish target changes, bypass activation, and other mechanisms. Escalating dose without locating the cause may add toxicity without solving resistance.

Case 6 — PARP activity with a broken recognition step. DNA damage and STING signaling may rise while B2M loss limits CD8 recognition. The immune-combination rationale weakens, but direct PARP-mediated cytotoxic activity may remain. Different mechanisms must be evaluated separately.

Case 7 — The macrophage count that falls by 80%. If ctDNA, tumor burden, and T-cell function do not change, the intervention has altered biology without yet demonstrating patient benefit. Functional reprogramming offers different evidence, not an automatic victory over depletion.

Case 8 — A phase III readout, then the company. Assume a hypothetical randomized trial in a prespecified TGF-β-high population has PFS as its sole primary endpoint, with the planned statistical testing properly controlled:

The appropriate conclusion is a successful primary PFS test, with supportive efficacy measures, unestablished OS benefit, and additional toxicity to assess. A stronger phase II estimate does not invalidate this outcome. Approval scope and commercial performance remain separate questions.

Now consider two debt-free hypothetical companies, excluding other assets. X has a $300 million market cap and $180 million in cash: simplified EV is $120 million. Y has a $150 million market cap and $35 million in cash: EV is $115 million.

If X also has randomized evidence, prespecified biomarkers, full economic rights, and funded phase III development, while Y has single-arm data, post-hoc selection, half the economics, and a near-term financing need, X has more favorable evidence and funding attributes at a similar simplified EV. We still cannot calculate fair value without the market, costs, competition, and cash-flow model.

The habit to carry forward is precise questioning. What does this result establish? What remains untested? Which uncertainty will the next experiment reduce? Cancer drug development becomes more understandable when every attractive number is placed back into the chain that connects biology to a durable, tolerable benefit for patients.