Symptom Networks Map How PD-1/PD-L1 Immune Toxicity Clusters Across Organs
Checkpoint blockade has reshaped oncology because PD-1/PD-L1 inhibitors remove a brake on antitumor immunity. The same immune activation that benefits patients can also spill into healthy tissue, producing immune-related adverse events that are often diffuse, overlapping, and difficult to interpret when they appear one by one in spontaneous reports. This study approached that problem differently: instead of treating each symptom as an isolated signal, it examined how adverse events travel together across a network. The result was a structured view of suspected toxicity patterns rather than a flat list of preferred terms.
In both FAERS and JADER, the investigators focused on six agents introduced before 2019: pembrolizumab, nivolumab, cemiplimab, atezolizumab, durvalumab, and avelumab. The analysis drew from a very large spontaneous reporting base, then filtered reports using three established pharmacovigilance methods: reporting odds ratio, proportional reporting ratio, and multivariate gamma-Poisson shrinkage. Only signals meeting all three thresholds were carried forward into the final network work. This conservative design reduced the risk of overcalling weak associations in a dataset where reporting noise, duplicate cases, and uneven clinical suspicion can distort apparent safety patterns.
Nivolumab and pembrolizumab accounted for the largest FAERS report volumes, with 48,330 and 37,224 reports, respectively. That volume alone does not prove greater toxicity; it reflects extensive use and the realities of spontaneous reporting. Still, it provided enough signal density to estimate symptom co-occurrence structures with greater confidence than sparse datasets usually allow.
Fatigue, Appetite Loss, Diarrhea, and Nausea Form the Core PD-1 Pattern
Across the FAERS-derived PD-1 inhibitor networks, the center of gravity consistently sat within a systemic-gastrointestinal cluster. Decreased appetite, fatigue, diarrhea, and nausea formed the most prominent core symptoms. In clinical terms, this combination reads like a familiar bedside pattern: a patient becomes anorectic, drained, and gastrointestinally unstable long before a single organ-specific diagnosis is pinned down. The network result suggests these complaints are not random companions in the database; they move together often enough to define the architecture of the suspected irAE profile.
One sentence captures the practical implication. These symptoms cluster before they fully declare a target organ.
That matters because checkpoint toxicities rarely announce themselves with a single, neatly isolated complaint. Fatigue may accompany endocrine dysfunction, diarrhea may represent inflammatory colitis, and anorexia can accompany either systemic inflammation or evolving organ-specific injury. By placing the symptoms into a shared network, the analysis gives clinicians a way to think about multi-symptom presentations as a connected syndrome rather than as disconnected adverse event labels.
Hematologic Signals Become More Visible in PD-L1 Networks
The PD-L1 inhibitor networks showed a broader profile. Alongside systemic and gastrointestinal symptoms, hematologic abnormalities were more prominent than in the PD-1 networks. The article does not claim a single mechanistic explanation for that difference, and it should not be overread as proof of a class effect, but the pattern is clinically relevant. Hematologic toxicity can be harder to spot in routine practice because it may emerge silently in laboratory testing before the patient describes a specific symptom. A network that brings those abnormalities into the core structure can help alert clinicians to a wider toxicity footprint than symptom review alone might suggest.
The methodological advantage here is plain: co-occurrence analysis can expose whether certain abnormalities repeatedly appear together across reports, which may mirror the multi-organ immune dysregulation seen with checkpoint inhibition. In practice, that can sharpen surveillance for patients who are already vulnerable because of disease burden, prior therapy, or limited physiologic reserve.
JADER Points Toward Pulmonary and Endocrine Centrality
A different pattern emerged in JADER. Interstitial lung disease and adrenal insufficiency became central nodes in the Japanese database, creating a profile that differed from the FAERS-derived networks. This divergence is not a minor statistical footnote; it shows how pharmacovigilance output can shift across countries because of differences in reporting behavior, clinical attention, and physician-initiated reporting practices. The same drug class can therefore appear to have a different toxicity center depending on the surveillance system used to capture it.
The pulmonary signal deserves particular attention because checkpoint inhibitor lung injury can escalate quickly and may be difficult to distinguish from infection, tumor progression, or treatment-related deconditioning. Adrenal insufficiency carries another layer of risk because it can present nonspecifically, with fatigue, anorexia, weakness, or hypotension—symptoms that overlap with the general syndrome already visible in the PD-1 networks. When endocrine collapse sits near the center of the network, it reinforces the need to measure more than what the patient spontaneously reports.
Centrality Metrics Held Up Under Bootstrap Testing
The investigators did not stop at network visualization. They calculated Strength, Closeness, Betweenness, and Expected Influence, then used 1,000 bootstrap iterations to test edge stability and centrality reliability. Strength and Expected Influence showed good to excellent stability in most networks, with CS values at or above 0.672. In practical terms, the main symptom clusters were not brittle artifacts of one particular resampling run; they persisted when the model was stress-tested.
That stability supports the idea that the core symptoms identified here are not random byproducts of database structure alone. The network is still built from spontaneous reports, and those reports remain subject to underreporting, reporting bias, and variation in coding practices. Even so, the internal consistency of the centrality estimates argues that the symptom clusters are robust enough to merit clinical attention.
What the Network Approach Adds to Pharmacovigilance
Traditional disproportionality analysis answers a narrow question: which drug-event pairs are reported more often than expected? Symptom network analysis asks a different one: which adverse events appear to anchor a larger pattern of co-occurring toxicity? For checkpoint inhibitors, that shift is useful because immune toxicity is rarely monosymptomatic. A patient may enter the clinic with fatigue and appetite loss, then disclose diarrhea, nausea, cough, or endocrine symptoms only after more directed questioning. A network view aligns more closely with that bedside reality than a single-signal table does.
The study also makes a careful distinction between suspected and clinically confirmed irAE. The events came from spontaneous reporting systems, not adjudicated case series, so the findings should be read as a signal-mapping exercise rather than a definitive incidence study. Even with that limitation, the work provides a practical framework for surveillance: FAERS points toward a systemic-gastrointestinal core, while JADER gives greater weight to pulmonary and endocrine toxicity. Those differences are useful because they remind clinicians that immune toxicity does not always announce itself with the same leading symptoms in every population or database.