68. CLINICAL TRAJECTORIES AND GENETIC ARCHITECTURE ACROSS THE NEUROLOGICAL–PSYCHIATRIC BOUNDARY

Background Neurological and psychiatric disorders are clinically and institutionally treated as distinct diagnostic categories, despite shared genetic risk factors and overlapping clinical presentation. Recent genome-wide analyses have revealed substantial pleiotropy crossing this divide (Smeland et al., Nat. Neurosci. 2025), but a comparable quantitative measure of clinical trajectory similarity across many disorders has not been available until recently. Whether clinical similarity mirrors genetic similarity across neurological and psychiatric disorders has therefore not been tested. Here we bring disease trajectory embeddings from a recent foundation model and genetic architecture into a single comparative framework, and ask whether the two views yield concordant or divergent maps of the neurological–psychiatric boundary. Methods We used pairwise genetic similarity estimates from Smeland et al., derived using LD score regression and bivariate MiXeR on GWAS summary statistics. We combined these with disease trajectory embeddings from Delphi-2M (Shmatko et al., Nature 2025), a generative transformer trained on electronic health records from 400,000 UK Biobank participants. Delphi-2M embeds more than 1,000 ICD-10 diagnoses in a space that captures their timing and co-occurrence across the life course. For each of 19 neurological and psychiatric disorders, we computed pairwise cosine similarity between disease trajectory embeddings, giving a clinical similarity matrix on the same disorder set as the genetic similarity matrix. We tested the global correspondence between the two matrices using a Mantel test (Spearman; 10,000 permutations). For each disorder, we then computed boundary strength as the difference between its average within-category and cross-category similarity, separately for genetic and clinical similarity, with leave-one-disorder-out sensitivity. Results Across 171 disorder pairs, clinical and genetic similarity showed shared structure (Mantel r = 0.33, p = 1.15 × 10⁻⁵). Both boundaries held for most disorders. The two measures agreed on the exceptions: multiple sclerosis, migraine, and essential tremor crossed the boundary in clinical trajectory similarity and in genetic similarity. An unsupervised model with no diagnostic-category supervision thus independently recovered the genetic boundary structure, with the consistent crossings highlighting where the institutional divide tracks biology imperfectly. Discussion The findings show that clinical trajectory similarity mirrors genetic pleiotropy across the neurological–psychiatric boundary, including in identifying which specific disorders cross it. The three boundary-crossers share features that distinguish them from neurological disorders that respect the boundary: earlier age of onset, substantial psychiatric comorbidity, and positive genetic correlations with internalizing psychiatric disorders. These results support a more integrated approach to brain disease across the traditional neurological–psychiatric divide, with implications for how cross-category disorders are screened and managed in clinical practice. The same joint-mapping approach extends naturally to other complementary views of brain disease, including imaging, transcriptomics, or molecular phenotyping.