AUTHORS
Kelly Sharon Chien, MD, Ling Zhang, MD2, Ian Bouligny, M.D.3, Najla Al Ali, MS4, Len Brandes, PhD5, Ruoyu Wang, PhD6, Xuezi Hu, MS7, Vasco Ferreira Pinto8, Alan Aberdeen, MS9, Pezhman Hosseinpoor10, Korsuk Sirinukunwattana, PhD11, Rami Komrokji, MD12, Guillermo Garcia-Manero, MD13, Sanam Loghavi, MD14, David Sallman, MD15
AFFILIATIONS
1 The University of Texas MD Anderson Cancer Center, Houston, United States
2 Moffitt Cancer Center, Tampa, United States
3 The University of Texas MD Anderson Cancer Center, Houston, United States
4 Moffitt Cancer Center, Tampa, United States
5 Ground Truth Labs, Oxford, United Kingdom
6 Ground Truth Labs, Oxford, United Kingdom
7 Ground Truth Labs, Oxford, United Kingdom
8 Ground Truth Labs, Oxford, United Kingdom
9 Ground Truth Labs, Oxford, United Kingdom
10 The University of Texas MD Anderson Cancer Center, Houston, United States
11 Ground Truth Labs, Oxford, United Kingdom
12 Moffitt Cancer Center, Tampa, United States
13 The University of Texas MD Anderson Cancer Center, Houston, United States
14 The University of Texas MD Anderson Cancer Center, Houston, United States
15 Moffitt Cancer Center, Tampa, United States
ABSTRACT
Background. Treatment decisions in myelodysplastic syndromes (MDS) are guided by risk stratification. Current tools, such as IPSS-R and IPSS-M, estimate risk using blood counts, bone marrow blast percentage, cytogenetics and, for IPSS-M, mutations across 31 genes. Bone marrow core biopsies are routinely obtained at diagnosis and during treatment, yet beyond blast counts, their rich microenvironmental information is excluded from risk assessment. We investigated histomorphologic features (“morphomics”), applying attention-based deep learning to whole-slide images. This “morphomics” approach derives a high-dimensional morphological representation of the bone marrow microenvironment without restricting analysis to predefined pathological features and evaluates its prognostic value alongside established risk scores.
Methods. We analysed 466 Hematoxylin and Eosin (H&E)-stained bone marrow biopsies from 384 MDS patients across two centres. Patch-level features from the Virchow-1 pathology foundation model were aggregated using attention-based multiple-instance learning (CLAM) within a Cox proportional hazards framework to predict overall survival (OS) directly from the whole-slide image. We combined the resulting morphomic risk score with IPSS-M using a linear Cox model. Discrimination was measured using Harrell's C-index under patient-level 5-fold cross-validation (three folds used for training, one for validation and one for testing), while independent contribution was assessed using likelihood-ratio tests (LRTs) on the test folds. Additionally, we trained a proof-of-principle model for hypomethylating agent (HMA) treatment response on a subpopulation of 147 patients, reporting results on validation folds in 3-fold cross-validation.
Results. Morphomics alone predicted OS (C-index mean ± standard deviation; 0.627 ± 0.037) similar to IPSS-R (0.666 ± 0.040) and below IPSS-M (0.698 ± 0.044). LRTs indicated that bone marrow biopsies add independent prognostic information on top of both IPSS-R and IPSS-M (all p ≤ 1e-4). Combining morphomics with IPSS-M increased discrimination to 0.710 ± 0.051. For response to HMA therapy, morphomics and IPSS-M alone moderately predicted time to treatment failure (C-index 0.643 ± 0.043 and 0.657 ± 0.011, respectively), while combining them improved discrimination to 0.695 ± 0.048.
Conclusions. H&E-stained slides from routine bone marrow biopsies carry additional prognostic information that has historically not been accounted for in prognostic models. This information can be extracted with deep learning, and the signal is statistically independent of, and complementary to, current clinical and genetic risk scores in MDS. Morphomics offers an accessible route to improve prognostic discrimination in combination with IPSS-M, with early evidence that the same biopsy may predict outcomes to hypomethylating agents, pending prospective validation.
DISCLOSURES
Kelly Sharon Chien, MD: Rigel - Research Funding; Geron - Consultancy (Includes expert testimony)