Endometrial cancer (EC) is increasingly affecting younger women who have not yet had children. Up to 14% of EC cases occur in premenopausal women, and 70% of women under 40 diagnosed with EC have not yet given birth. As women increasingly delay childbearing, a growing number of patients face the difficult situation of being diagnosed with a cancer that typically requires removal of the uterus.
Endometrial atypical hyperplasia (EAH) is the precancerous condition that precedes many ECs. It has at least a 32.6% chance of progressing to invasive cancer if untreated - but it also represents an opportunity for intervention before cancer fully develops. Standard treatment for both early EC and EAH is hysterectomy, but this permanently ends a woman's ability to carry a pregnancy.
Fertility-sparing treatment offers an alternative for carefully selected patients. The main approach is hormone therapy using progestogens (progesterone-like drugs such as megestrol acetate or medroxyprogesterone acetate, often delivered via an intrauterine device). Additional options include hysteroscopic surgery to remove the tumor while preserving the uterus, and adjunct medications like GnRH agonists, letrozole, and metformin. Most patients are assessed every 3-6 months with endometrial biopsies to monitor response.
The problem is that fertility-sparing treatment does not work equally well for all patients. Complete response (CR) rates range from 79.7% to 84.6% across studies, recurrence after initial response occurs in 23-35% of patients, and some patients experience disease progression. Predicting in advance which patients will respond well or poorly is the central challenge - and the focus of this review article.
The Cancer Genome Atlas (TCGA) project identified four distinct molecular subtypes of endometrial cancer, each with different biological behavior, prognoses, and treatment implications. These are now incorporated into the 2023 FIGO (International Federation of Gynecology and Obstetrics) staging system and are becoming standard in clinical practice.
POLEmut (POLE ultramutated): caused by mutations in the DNA repair gene POLE, leading to an extremely high mutation burden. Despite this, the prognosis is excellent - the immune system appears to recognize and attack these highly mutated tumors, and recurrence is rare. Patients tend to be younger with lower BMI and typically present at Stage I.
MMRd (mismatch repair deficient): caused by loss of mismatch repair proteins (MLH1, MSH2, MSH6, or PMS2), resulting in microsatellite instability. Prognosis is intermediate, but these patients respond poorly to hormone therapy due to low progesterone receptor (PR) expression, and they have high recurrence rates after initial response. MMRd is also associated with Lynch syndrome, a hereditary cancer predisposition.
NSMP (no specific molecular profile): the most common subtype in younger women (64% of women under 50), typically estrogen-dependent (Type I) endometrioid cancers. These patients have the highest PR expression and the best response to hormone therapy, but there is significant heterogeneity within this group. p53 abnormal (p53abn): carries TP53 mutations, is biologically the most aggressive, has the lowest PR expression, the lowest complete response rates, and the worst prognosis. Fertility-sparing treatment is generally not recommended for this group.
POLEmut patients generally respond well to fertility-sparing treatment, with most achieving complete response across multiple studies, regardless of histological grade. However, numbers are small (POLE mutations are less common in the low-grade ECs typical of fertility-sparing candidates), and a few studies report late recurrence and eventual hysterectomy. There is also early evidence that GnRH agonist combined with letrozole may outperform standard progestogen therapy for this subtype.
MMRd patients show consistently poor outcomes with hormone therapy. Their low PR expression means progesterone cannot exert its cancer-suppressing effects as effectively. Studies show low CR rates and significantly higher recurrence rates compared to NSMP patients. Women with Lynch syndrome (germline MMR mutations) can sometimes achieve CR and even conceive, but relapse rates are high and rapid progression after relapse is documented. Immune checkpoint inhibitors (which are effective in MMRd cancers in other settings) represent a promising but still experimental avenue for this subgroup.
NSMP patients respond best overall to hormone therapy, with CR rates ranging from 63% to 96.2% across different studies. However, this wide range reflects meaningful biological heterogeneity within the NSMP category. Gene mutations within NSMP tumors (particularly PTEN, KRAS, PIK3CA, and CTNNB1) may explain why some NSMP tumors do not respond well, making further molecular testing within this group valuable for personalizing treatment. p53abn patients have the lowest CR rates and highest progression risk; the molecular profile and clinical behavior of this subtype do not support conservative management.
Within and across molecular subtypes, specific gene mutations help predict which patients will respond to progestogen-based therapy. PTEN mutations (present in 69-80% of endometrioid EC) are associated with resistance to progesterone therapy: patients with PTEN mutations show lower CR rates, longer time to achieve CR, and higher recurrence rates one year after CR, likely because PTEN regulates the PI3K/AKT/mTOR pathway that is also involved in progesterone receptor signaling.
KRAS mutations are linked to disease progression and the development of invasive cancer in patients initially treated for EAH, though their effect on CR rates is less clear-cut. PIK3CA mutations delay the time to achieve CR and may independently predict lower cumulative CR rates, likely through similar PI3K pathway disruption. These findings suggest that the PI3K/AKT/mTOR signaling axis is a key resistance mechanism against progestogen therapy.
CTNNB1 mutations (present in up to 50% of G1-G2 EC, mainly in NSMP) activate the WNT pathway and are associated with a higher recurrence risk in early-stage low-grade EC. ARID1A mutations reduce PR expression and are associated with microsatellite instability; EC with these mutations may not respond well to hormone therapy. While direct evidence in fertility-sparing contexts is still limited for CTNNB1 and ARID1A, they represent important biomarkers for further stratifying the heterogeneous NSMP group.
Traditional biomarkers measured with tissue staining (immunohistochemistry) also carry predictive value. Progesterone receptor (PR) expression is the most important: high PR expression before treatment is associated with better response to fertility-sparing therapy, while low PR levels during follow-up suggest ongoing good response. The specific isoform matters too - increased progesterone receptor B (PRB) and decreased progesterone receptor A (PRA) during treatment are linked to relapse.
Estrogen receptor (ER) expression is less predictive at baseline, but low ER levels during follow-up indicate a higher risk of relapse after achieving CR. Importantly, ER and PR are predictive mainly when intrauterine progesterone (via an IUD) is used, and may not reliably predict the outcome of oral progesterone therapy - a distinction with direct clinical relevance.
Ki-67, a marker of how rapidly cells are dividing, shows that elevated expression after treatment - but not necessarily before - predicts poor response and higher recurrence risk. The L1CAM protein, typically overexpressed in p53abn tumors, may identify a subgroup of NSMP patients whose tumors behave more aggressively (similar to p53abn) and who may not respond well to hormone therapy, providing a way to identify higher-risk patients within the otherwise favorable NSMP category.
Artificial intelligence (AI) is beginning to play a role in predicting treatment outcomes from routine pathology slides. Current research focuses on using whole slide images (WSIs) - high-resolution digital scans of stained tissue - to extract patterns that predict molecular subtype, gene mutations, and potentially treatment response. Deep learning models can classify endometrial cancer molecular subtypes and predict individual gene mutations from H&E stained images, tasks that traditional pathologists cannot do reliably from morphology alone.
An unsupervised deep learning model trained on morphological features of endometrial tissue has already achieved an AUC of 0.79 for predicting hormone therapy response - demonstrating that the tissue appearance carries encoded information about how the tumor will behave. AI tools could make molecular information accessible without requiring expensive genetic sequencing, by inferring it from routinely performed staining.
The authors envision a future integrated prediction framework combining molecular classification, specific gene mutation markers, immunohistochemical protein expression, and AI-based image analysis. This multi-layered approach would allow clinicians to identify at the time of initial biopsy which patients are good candidates for fertility-sparing treatment - maximizing chances of success while avoiding prolonged ineffective treatment in women likely to progress.
Key gaps remain: most studies have small sample sizes (especially for rare molecular subtypes like POLEmut and p53abn), fertility outcomes after treatment are rarely tracked long-term, and molecular testing is not yet standardized across fertility-sparing guidelines. Cost-effectiveness analyses suggest that molecular testing is worthwhile, particularly for POLEmut identification, and costs are falling as next-generation sequencing becomes more widespread.