Alzheimer's subtyping no longer needs every modality: ADSMoE finds four subtypes in ADNI and reproduces them in NACC and UKB with a frozen model
Synopsis
In real cohorts, imaging, genetic, protein and clinical data are often only partly measured, and earlier subtyping methods either rely on predefined biomarker trajectories or anchor subtypes to a single dominant modality. ADSMoE places mixture-of-experts modules, in which separate subnetworks specialize in different patterns, inside a transformer autoencoder: the encoder accepts any combination of modalities, and the decoder reconstructs inputs through four subtype-specific experts, so subtypes are learned as latent variables together with the representation. Trained on ADNI, the model identified four subtypes differing in neuroimaging, genetic, protein and clinical features; applied frozen, without retraining, to NACC and UKB, subtype-associated characteristics remained consistent.
Interpretation
Alzheimer's subtypes can be learned directly from incomplete multimodal data, without first filling in modalities or picking one dominant modality. Earlier progression-based frameworks rely on predefined biomarker trajectories and typically use only a few consistently acquired markers; modality-centered methods anchor subtyping to a dominant modality such as imaging and treat the rest as auxiliary; autoencoder and transformer multimodal models are built mainly for prediction or classification, with subtyping as downstream analysis, and assume a fixed input modality set during training. On ADNI, against GeneSGAN, CDEC, EDESC, Kmeans and spectral clustering, ADSMoE reached 189±16 subtype-associated SNPs (N-Asso-SNPs) with a Davies-Bouldin index of 4.24±0.74 under five-fold cross-validation; GeneSGAN gave 100±9 and 22.57±0.46, and CDEC reached a DBI as low as 0.46±0.00 but only 59±18 N-Asso-SNPs.
The model routes each patient to one of four subtype experts, and the experts specialize in distinct disease patterns while reconstructing the inputs. The encoder is a Fleximodal MoE encoder that fills missing modalities with learnable missing-modality tokens and fuses top-K routed experts; the decoder replaces the first feed-forward network with four subtype-specific experts and uses patient-level soft routing so all experts contribute to reconstruction. Training runs in three stages: normative modeling against cognitively normal participants regresses out age, sex and head size to leave standardized residuals; the encoder is then trained; finally, after expectation-maximization initializes subtype gate weights, encoder, decoder and routing network are jointly optimized with an expert-balancing regularizer. The model is trained on ADNI, and NACC and UKB serve only as external validation without retraining or fine-tuning.
Perspective
The demonstrated scope is a three-cohort setting with training on ADNI and frozen external validation in NACC and UKB, including all ADNI participants and NACC and UKB cases with at least two modalities, with 876 analyzed participants per cohort; the number of subtypes is fixed at K=4, consistent with prior studies identifying four neuroimaging-based subtypes.
The authors note that cross-cohort harmonization did not fully account for platform- and biospecimen-specific differences, particularly between CSF and plasma proteomic measurements, which may affect comparability of molecular associations; subtype discovery relied mainly on cross-sectional multimodal measurements; and although the genetic and proteomic associations support biological relevance, experimental validation is needed to establish mechanisms and potential causal contributions.
