📄 medicine

Antiangiogenic Class and Trial Population Are Confounded in First- Line Immunotherapy for Hepatocellular Carcinoma: A Bayesian Network Meta-Analysis of Nine Randomised Trials

This Bayesian network meta-analysis of nine trials reveals that the apparent superiority of anti-VEGF antibody-based regimens over TKI-based regimens in first-line hepatocellular carcinoma treatment is unidentifiable as a true drug-class effect due to severe confounding by trial population characteristics (specifically hepatitis B prevalence and geographic enrollment), rendering class-level rankings unreliable without individual patient data or head-to-head comparisons.

Original authors: Burkhoniddin Tillashaykhov, Iroda Ikromova, Dilnavoz Berdiyeva

Published 2026-10-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Burkhoniddin Tillashaykhov, Iroda Ikromova, Dilnavoz Berdiyeva

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Liver cancer remains one of the most formidable challenges in modern medicine, ranking as the third leading cause of cancer death worldwide. For patients whose disease has spread beyond the liver or cannot be removed by surgery, the outlook has historically been grim. For over a decade, the standard treatment was a single drug that slowed tumor growth but offered only modest survival benefits. In recent years, a new approach has emerged, combining two different types of medicines to attack the cancer more effectively. One type of drug acts as a key, unlocking the body's own immune system to recognize and fight the tumor. The other type acts as a blocker, cutting off the blood supply that the tumor needs to grow. Doctors now routinely pair these two strategies, but they face a difficult choice: which blocker works best? One group of blockers targets a specific protein on blood vessels, while another group acts as a broader inhibitor of several growth signals. Because no large clinical trial has ever directly compared these two groups against each other, researchers have relied on complex statistical methods to piece together evidence from separate studies. This is where the question of what actually works becomes murky, and where a new analysis by researchers at Central Asian University offers a crucial, if sobering, clarification.

The researchers set out to examine nine major clinical trials involving more than 5,300 patients with advanced liver cancer. These trials tested various combinations of the immune-activating drug and one of the two types of blood-supply blockers. The goal was to determine if one type of blocker was inherently superior to the other when paired with the immune drug. The team used a sophisticated method called a network meta-analysis, which allows scientists to compare treatments that have never been tested head-to-head by linking them through a common reference point, in this case, the older standard drug. They looked at whether patients lived longer, whether their tumors shrank, and how long it took for the disease to progress. They also carefully tracked severe side effects to see if one approach was safer than the other.

The analysis confirmed that several of the new combination treatments were indeed better than the old standard. Specifically, three regimens showed a clear survival advantage, with patients living longer than those on the older drug. When the researchers looked at the two groups of blockers as a whole, the data suggested a slight edge for the group that uses a specific antibody to block blood vessel growth. This group appeared to help patients live longer and achieve better tumor shrinkage more often than the group using the broader inhibitors. However, the researchers stopped short of declaring this a victory for one drug class over the other. Instead, they discovered a fundamental flaw in the way the evidence was arranged.

The core problem was that the choice of drug was not independent of the patients who received it. The trials were not a random mix of patients from around the world; they were heavily clustered by geography and by the underlying cause of the liver disease. The trials that used the antibody-based blockers were conducted almost exclusively in China, where the vast majority of patients had liver cancer caused by the hepatitis B virus. In contrast, the trials using the broader inhibitors included more patients from other regions and with different causes of liver disease. Because the hepatitis B virus is known to make the immune system respond differently to treatment, the researchers found that the apparent success of the antibody group could just as easily be explained by the fact that they were treating a specific population of patients rather than by the drug itself. The two factors—the type of drug and the type of patient—were so tightly intertwined that the statistical models could not separate them.

This means that the rankings often cited in medical guidelines, which place one drug class above the other, are actually describing the specific mix of patients in the available studies rather than a universal truth about the drugs. The researchers found that if you try to adjust for these differences, the advantage disappears or becomes impossible to measure with certainty. The data simply does not contain enough information to tell whether the antibody is better, or if the patients in the Chinese trials just happened to respond better to any treatment. The only comparisons that remain clear and trustworthy are those made within a single trial, where patients are randomly assigned to different arms. For example, the researchers could confidently say that the broader inhibitors caused a specific type of skin and foot reaction much more frequently than the antibody-based drugs, a difference that held true regardless of the patient population.

Ultimately, the study concludes that the current evidence base cannot definitively answer which class of drugs is superior. The rankings that currently guide treatment decisions reflect the geography and viral history of the trial participants, not an intrinsic property of the medicines. To truly solve this question, the medical community would need either a massive trial that directly compares the two approaches in the same group of patients, or a pooled analysis of individual patient data from all existing trials. Until that happens, the slight advantage seen in the antibody group should be viewed with caution. The findings serve as a reminder that in complex medical research, the way data is collected can sometimes tell a story that looks like a drug's success but is actually a story about the people who took it.

Technical Summary: Antiangiogenic Class and Trial Population Are Confounded in First-Line Immunotherapy for Hepatocellular Carcinoma

Problem Statement
First-line systemic therapy for advanced hepatocellular carcinoma (HCC) currently relies on combining an immune checkpoint inhibitor (ICI) with an antiangiogenic agent. These agents fall into two mechanistic classes: anti-VEGF monoclonal antibodies (e.g., bevacizumab) and VEGFR-targeted tyrosine kinase inhibitors (TKIs). While no head-to-head randomized trials exist comparing these two backbones, previous network meta-analyses (NMAs) have ranked them, often favoring the anti-VEGF antibody class. These rankings are frequently interpreted as intrinsic properties of the drug classes. However, the validity of such indirect comparisons relies on the assumption of transitivity—that trials contributing to each class are exchangeable regarding effect modifiers. The authors posit that in the current HCC evidence base, treatment class is inextricably confounded with trial population characteristics, specifically viral etiology (hepatitis B virus prevalence) and geographic enrollment (China-only vs. global). Consequently, aggregate data may be unable to distinguish whether observed survival advantages stem from the drug mechanism or the specific patient populations in which those drugs were tested.

Methodology
The study is a Bayesian random-effects network meta-analysis (NMA) of nine randomized controlled trials involving 5,333 patients, registered with PROSPERO (CRD420261353528).

Network Geometry: The evidence network is a "star" design anchored on sorafenib. Eight trials compared an ICI combination directly to sorafenib. One trial (LEAP-002) used lenvatinib as a control; this was connected to the sorafenib anchor via the REFLECT trial (lenvatinib vs. sorafenib), forming a single two-link indirect bridge.
Statistical Framework: Analyses were conducted using PyMC with Hamiltonian Monte Carlo. Because each active treatment appears in only one trial, between-study heterogeneity ( $\tau$ ) is not identifiable from the data. Therefore, heterogeneity was specified using external empirical predictive distributions (Turner et al.) rather than estimated from the network.
Effect Modifiers: Five pre-specified effect modifiers were examined: hepatitis B virus (HBV) prevalence, AFP $\ge$ 400 ng/mL, macrovascular invasion/extrahepatic spread (MVI/EHS), ECOG performance status, and Child–Pugh class.
Transitivity Assessment: The authors explicitly tested for collinearity between treatment class and these modifiers. They performed exploratory meta-regressions and sensitivity analyses (including fixed-effect models, alternative priors, and leave-one-trial-out re-estimations) to determine if the class-level probabilities were robust to model specification or if they remained driven by the confounding structure.
Outcomes: Overall survival (OS), progression-free survival (PFS), objective response rate (ORR), and grade $\ge$ 3 adverse events.

Key Results

Efficacy Estimates: Three regimens demonstrated a credible survival advantage over sorafenib (95% credible intervals excluding unity): sintilimab plus bevacizumab (HR 0.57), finotonlimab plus bevacizumab (HR 0.60), and camrelizumab plus rivoceranib (HR 0.64).
Class-Level Probabilities: The posterior probability favoring the anti-VEGF antibody class over the TKI class was 75.0% for OS, 74.8% for ORR, and 45.3% for PFS.
Confounding and Identifiability: A critical finding was the strong collinearity between treatment class, HBV prevalence, and geographic region. The four trials with the highest HBV prevalence (84–94%) were all conducted exclusively in China and utilized specific regimens. The pairwise correlation between HBV prevalence and China-only enrollment was $r = 0.87$ .
Meta-Regression: Exploratory meta-regression suggested greater survival benefit in trials with higher HBV prevalence (slope -0.068 per 10%). However, due to the high intercorrelation of modifiers ( $r = 0.81–0.91$ ) and the small number of trials (N=9), the model was underpowered to disentangle the independent contributions of viral etiology and treatment class.
Sensitivity: The class probability remained between 73.9% and 85.5% across various sensitivity analyses (different priors, fixed-effect models, removing the lenvatinib bridge). The authors argue this stability reflects the stability of the confounding, not the robustness of a causal class effect.
Safety: Grade $\ge$ 3 hand-foot syndrome occurred in every TKI-based regimen (2–12%) but in none of the anti-VEGF antibody regimens, a contrast observable within the same trials. Overall, TKI regimens were approximately twice as toxic as antibody regimens (median OR 2.10 vs. 1.09).

Significance and Claims
The paper's primary contribution is methodological and interpretive rather than a definitive ranking of drug efficacy. The authors claim that:

Class Rankings Describe the Evidence Base, Not the Drugs: The current aggregate data cannot separate the effect of the drug class from the effect of the trial population. Therefore, the reported "advantage" of anti-VEGF antibodies is likely a reflection of the specific populations (e.g., high HBV prevalence in China) in which those trials were conducted, rather than an intrinsic superiority of the antibody class.
Unidentifiability: The confounding is structural. No amount of modeling (reweighting, prior specification, or meta-regression) can resolve this with the current aggregate data. The quantity reported by previous NMAs (class-level probability) does not possess the causal meaning routinely attributed to it.
GRADE Assessment: On the GRADE-NMA scale, the class-level indirect comparison is rated as "very low certainty" due to unresolvable confounding by trial population characteristics.
Defensible Claims: The only defensible class-level claims are those derived from within-trial contrasts, such as the distinct toxicity profiles (e.g., hand-foot syndrome in TKI arms vs. antibody arms).
Future Requirements: Resolving the efficacy question requires either a head-to-head randomized trial (which the authors note is commercially unlikely) or a pooled individual-patient-data (IPD) analysis that allows for the estimation of aetiology-by-treatment interactions within trials rather than across them.

In conclusion, the study argues that while specific regimens show efficacy, the current evidence base does not support a causal inference that one antiangiogenic class is superior to the other for first-line HCC therapy. The observed rankings describe the distribution of the available trials rather than a property of the drugs themselves.

More like this