Everyday chemicals that quietly interfere with hormones—from the bisphenols lining food cans to the “forever chemicals” lingering in drinking water—may be steering thyroid tumors toward a more aggressive state by acting on a single, druggable cell-surface molecule. That is the case advanced by a new study published in the journal Molecular Diversity on 30 August 2026, in which researchers at Zhongnan Hospital of Wuhan University, working with a collaborator at The Chinese University of Hong Kong, Shenzhen, welded computational toxicology, machine learning, large-scale genomics, molecular simulation and laboratory experiments into a single investigative pipeline. At the far end of that pipeline stood an unexpected suspect: CD44, a renowned cancer stem cell marker that had never before been positioned so centrally in the story of pollution-linked thyroid cancer. The work delivers both a diagnostic lead of near-clinical accuracy and a potential therapeutic target for a disease whose global incidence has climbed for decades.
Endocrine-disrupting chemicals, or EDCs, are synthetic compounds that can mimic, block or scramble hormonal signaling. They are woven into modern life almost invisibly: bisphenol A (BPA) leaches from polycarbonate plastics, epoxy-can linings and thermal receipt paper; perfluorooctanoic acid (PFOA) belongs to the sprawling family of per- and polyfluoroalkyl substances, or PFAS, long used in nonstick cookware, water-repellent textiles and firefighting foams; di(2-ethylhexyl) phthalate (DEHP) softens vinyl products and medical tubing; the flame retardant BDE-209 sheds from consumer goods into household dust; the pesticide DDT, banned in most countries decades ago, persists in soils and fatty tissue; and the dioxin TCDD, an industrial byproduct, ranks among the most potent synthetic toxins ever characterized. Because the thyroid gland depends on precisely tuned hormonal feedback to regulate metabolism, growth and development, it is considered acutely vulnerable to such exposures. Global burden analyses have documented a steep, decades-long rise in thyroid cancer incidence, and while improved detection explains part of the trend, environmental contributors remain a live scientific concern. Epidemiological studies have linked several of these chemicals to thyroid dysfunction, nodules and cancer risk, yet the molecular steps that turn exposure into tumor progression have remained stubbornly opaque—precisely the gap the new study set out to close.
The Wuhan-led team began not at the laboratory bench but at the computer. They first ran all six chemicals—BPA, PFOA, DDT, BDE-209, TCDD and DEHP—through ADMETlab 3.0, a machine-learning web platform that predicts a compound’s absorption, distribution, metabolism, excretion and toxicity directly from its molecular structure, providing a standardized read on how hazardous each molecule is likely to be. They then mined the Comparative Toxicogenomics Database, a curated public repository that logs which genes have been experimentally shown to respond to which chemicals. In parallel, the researchers compiled lists of genes implicated in thyroid cancer from multiple disease databases. Overlaying the chemical-response gene sets with the cancer gene sets produced a shared space of 1,113 EDC–thyroid cancer targets: genes that both respond to endocrine-disrupting chemicals and participate in malignant thyroid disease.
To give that list biological meaning, the team performed pathway enrichment analysis, a statistical method that tests whether a set of genes clusters within particular signaling networks more densely than chance would predict. The 1,113 shared targets concentrated strikingly in three circuits. The PI3K–Akt pathway, a core growth-control cascade promoting cell survival, proliferation and metabolism, is among the most frequently dysregulated networks in human cancer. The FoxO transcription factor family acts downstream of Akt and governs cell-cycle arrest, DNA repair, apoptosis and oxidative-stress resistance—functions of special relevance in the thyroid, where hormone synthesis inherently generates reactive oxygen species. The AGE–RAGE axis, which couples advanced glycation end products to their cell-surface receptor RAGE, sustains chronic inflammatory signaling and has previously been implicated in the migratory behavior of thyroid cancer cells. Convergence of EDC-responsive genes on these pro-growth, pro-survival, pro-inflammatory circuits handed the researchers their first mechanistic hypothesis: chemical exposure may be rewiring the very pathways that decide whether a thyroid tumor grows, spreads or dies.
The next question was which of those 1,113 genes actually separate tumor tissue from healthy thyroid. Differential expression analysis of transcriptomic datasets filtered the candidates down to genes consistently dysregulated in cancer, and the shortlist then went to machine learning. The team applied least absolute shrinkage and selection operator (Lasso) regression, an algorithm that penalizes model complexity and drives the coefficients of uninformative genes to exactly zero, compressing hundreds of features into a minimal signature. The surviving genes were classified with linear discriminant analysis (LDA), a supervised method that separates patient groups along an optimal linear boundary. The resulting six-gene diagnostic model—FN1, which encodes the extracellular-matrix protein fibronectin; BCL2, a canonical anti-apoptotic gene; CD44; CDKN1A, which encodes the cell-cycle brake p21; CTNNB1, the gene for β-catenin at the core of WNT signaling; and JUN, a pillar of the AP-1 transcription factor—achieved an average area under the receiver operating characteristic curve (AUC) of 0.976 across a training cohort and three independent validation cohorts. An AUC of 1.0 signifies perfect discrimination and 0.5 mere coin-flipping, so a value nearing 0.98 represents diagnostic performance close to clinical grade.
Within that six-gene panel, one name kept rising to the top. Evaluated alone, CD44 achieved AUC values of 0.950 in the training set, 0.801 in the external dataset GSE27155, 0.878 in GSE29265 and 0.938 in GSE153659—performance that persisted across cohorts generated by different laboratories on different platforms, a robustness that matters because datasets built independently are far less likely to share hidden technical biases. To understand why the algorithm leaned so heavily on this gene, the researchers deployed SHAP analysis, short for SHapley Additive exPlanations, a game-theoretic framework borrowed from economics that distributes the credit for every prediction among the features that produced it. In the SHAP ranking, CD44 made the largest single contribution to the model’s output, evidence that the machine had not latched onto a statistical artifact but onto the gene that best captured the boundary between tumor and healthy tissue.
CD44 is no obscure molecule. It encodes a transmembrane glycoprotein that serves as the principal cell-surface receptor for hyaluronic acid, the gel-like polymer filling the space between cells, and through that interaction it governs adhesion, migration and invasion. It is a defining marker of cancer stem cells—the self-renewing subpopulation thought to seed relapse and shrug off therapy—and has been tied to progression and metastasis across many tumor types, including papillary thyroid carcinoma. The new study layered further dimensions onto that profile. Immune infiltration analysis indicated that CD44 expression covaries with the makeup of the tumor immune microenvironment, the mix of macrophages, T cells and other immune players surrounding a growing tumor. Survival analysis of data from The Cancer Genome Atlas linked CD44 levels to patient prognosis, and single-cell RNA sequencing positioned the gene as a marker of shifting cellular states within malignant cells. Taken together, these analyses cast CD44 as sitting at the junction where environmental stress, tumor identity and immune context meet.
The most provocative question was whether the chemicals themselves can physically engage CD44. To probe it, the team used molecular docking, a computational technique that fits flexible small molecules into the three-dimensional structure of a protein’s binding region and scores how well each one lodges there. Docking produced plausible binding poses for BPA, DEHP and PFOA on CD44, with PFOA—the eight-carbon fluorinated compound infamous for its nearly unbreakable carbon–fluorine backbone—returning the most favorable predicted docking score. The researchers then stress-tested each protein–ligand complex with 200-nanosecond molecular dynamics simulations, which track every atom of the pair in a simulated water environment and reveal whether an interaction holds together or falls apart over realistic molecular timescales. The complexes remained plausible across the simulated run, supporting—though not yet proving—direct physical contact between these environmental chemicals and the CD44 protein. The authors are careful with language here: docking scores and simulated stability are computational hypotheses that will need confirmation by direct biophysical measurements such as surface plasmon resonance or calorimetry.
Computational hypotheses were not the endpoint. In the laboratory, the team confirmed that CD44 is expressed at higher levels in thyroid cancer tissues and thyroid cancer cell lines than in normal counterparts. When the cells were exposed to endocrine-disrupting chemicals, CD44 expression climbed further. The decisive experiment followed: using molecular tools to knock down CD44—silencing the gene so its protein is no longer made—the researchers showed that the gains in proliferation, colony formation and migration that ordinarily follow EDC exposure were substantially attenuated. Colony-forming assays, which measure how many single cells can found entire colonies, and migration assays, which track how quickly cells close a wound-like gap or invade through a membrane, are standard readouts of malignant potential. In effect, the experiment closed a loop: chemical exposure raises CD44, elevated CD44 licenses aggressive behavior, and removing CD44 strips much of that behavior away.
The authors frame CD44 as a candidate target associated with EDC-responsive malignant phenotypes in thyroid cancer—a deliberately measured formulation that reflects both the strength and the limits of the evidence. The study does not claim that endocrine-disrupting chemicals initiate thyroid cancer, and a docking score is not a demonstrated drug-like interaction. What it does establish is a coherent, multi-layered chain of evidence: computational toxicity profiling, toxicogenomic mining, pathway convergence, a near-clinically accurate six-gene diagnostic model, single-gene robustness across independent cohorts, microenvironmental and single-cell corroboration, structural simulation, and finally laboratory intervention. If future work confirms direct CD44 binding by these chemicals in living systems and validates the diagnostic panel in prospective patient cohorts, the implications are considerable, because CD44 is already pursued as an oncology target through antibodies and hyaluronan-based drug delivery strategies, offering a plausible road from biomarker to intervention. In the meantime, the study adds molecular weight to a public-health argument that has been building for years: curbing exposure to endocrine-disrupting chemicals is not merely an endocrine issue—it may be an oncological one.
Subject of Research: Identifying CD44 as a candidate molecular target linking endocrine-disrupting chemical exposure to thyroid cancer progression through integrated multi-omics, machine learning, molecular simulation and experimental validation.
Subject of Research: Medicine
Article Title: Multi-omics, machine learning, and molecular simulation identify CD44 as a candidate target in endocrine-disrupting chemical–associated thyroid cancer progression
Article References: Hu, Y., Liu, K., Chen, T., He, Z., Li, S., Hu, W., Fu, Q., & Chen, X. (2026). Multi-omics, machine learning, and molecular simulation identify CD44 as a candidate target in endocrine-disrupting chemical–associated thyroid cancer progression. Molecular Diversity. https://doi.org/10.1007/s11030-026-11721-0
Image Credits: AI Generated
DOI: 10.1007/s11030-026-11721-0
Keywords: Thyroid cancer, Endocrine-disrupting chemicals, CD44, Toxicology, Carcinogenicity, Single-cell analysis, Multi-omics, Machine learning, Molecular docking, Molecular dynamics simulation, Bisphenol A, PFOA
Cite Scienmag News
APA MLA Chicago
Nathaniel Bowman. (August 31, 2026). AI and multi-omics reveal CD44 as target in chemical-linked thyroid cancer. Scienmag. https://scienmag.com/ai-and-multi-omics-reveal-cd44-as-target-in-chemical-linked-thyroid-cancer/
Nathaniel Bowman. “AI and multi-omics reveal CD44 as target in chemical-linked thyroid cancer.” Scienmag, 31 August 2026, https://scienmag.com/ai-and-multi-omics-reveal-cd44-as-target-in-chemical-linked-thyroid-cancer/. Accessed 31 August 2026.
Nathaniel Bowman. “AI and multi-omics reveal CD44 as target in chemical-linked thyroid cancer.” Scienmag. August 31, 2026. https://scienmag.com/ai-and-multi-omics-reveal-cd44-as-target-in-chemical-linked-thyroid-cancer/
Copy citation Download RIS
Tags: advances in diagnostic and treatment strategies for thyroid cancerAI in cancer researchCD44 as cancer stem cell markerCD44 as therapeutic targetchemical interference with hormonal signalingchemical-linked carcinogenesiscomputational toxicologycomputational toxicology and machine learningdiagnostic biomarkers for thyroid tumorsdruggable cell surface molecules in cancerEDCs and cancer progressionEndocrine disrupting chemicalsenvironmental chemicals and cancer progressionenvironmental factors in thyroid tumor aggressivenessgenomics and molecular simulation in oncologymachine learning in oncologymulti-omics analysismulti-omics in cancer researchpollution impact on hormonal signalingpollution-linked cancer biomarkerstherapeutic targets in thyroid cancerThyroid cancer
