One protein, many brain cell types: SEMPER aligns sequences with interaction networks and recovers clinical targets for five neurodegenerative diseases zero-shot
Synopsis
SEMPER aligns protein language model sequence representations with protein interaction networks from 442 cell types of the adult human brain, giving each protein a representation in every cell type where its gene is expressed that keeps sequence information while varying with its interaction environment. In zero-shot retrieval it recovers clinically pursued targets for Parkinson's disease, Alzheimer's disease, multiple sclerosis, Huntington's disease and amyotrophic lateral sclerosis at higher rates than cell type-specific interaction models; combined with a large language model, mean recovery of targets among the top 25 candidates across the five diseases rises from 8.3 to 9.2.
Interpretation
The same protein can now carry different representations across brain cell types while each representation stays anchored to its sequence. Protein language models previously gave one cell type-agnostic representation per sequence, and network-based models used interaction information alone, restricted to proteins expressed in a given cell type. Across 30 CNS cell-type superclusters, mean cosine similarity of protein pairs within a cell type exceeded that of cross-cell-type pairs in every supercluster, with Benjamini-Hochberg-corrected Mann-Whitney U adjusted P < 0.001 for all comparisons.
The representation space organizes proteins by cell type and biological function and supports prediction without task-specific training. Proteins sharing a Gene Ontology term lie closer in the embedding space, with mean distances lower by 0.08 for biological process, 0.13 for cellular component and 0.15 for molecular function than randomly sampled proteins. In a permutation test that reassigns embeddings across proteins, same-term proteins remained more compact than average at least 49 SD above the permutation mean (2,000 permutations, P < 5×10⁻⁴ for all three domains).
In zero-shot retrieval, SEMPER recovers clinically pursued targets for five neurodegenerative diseases at higher rates than all cell type-specific baselines. Cell type-specific baselines PINNACLE, PPI-GT and ORCA-CT recovered mean Hits@10 of at most 0.3 in every disease, whereas SEMPER reached a mean Hits@10 of 7.0 of ten in Huntington's disease and recovered at least six held-out targets in 81.4% of its 129 cell types. Paired one-sided Wilcoxon signed-rank tests against the three cell type-specific baselines gave P < 10⁻²¹ in every disease; against the cell type-agnostic ProTrek-Seq, however, SEMPER was lower in Parkinson's disease (2.2±1.0 versus 3.2±1.6), Alzheimer's disease (1.4±1.0 versus 2.0±1.1) and multiple sclerosis (1.1±0.8 versus 1.5±0.9).
Handing SEMPER's candidate genes to a large language model for reordering supplies what experimental data and literature knowledge each lack. Across 50 disease-cell type pairs (five diseases, ten high-recovery cell types each), mean Hits@25 was 4.7 for the large language model alone, 8.3 for SEMPER, and 9.2 after reordering SEMPER's candidates. On a genome-wide CRISPR screen in human dopaminergic neurons, SEMPER exceeded the large language model at Hits@25, Hits@50 and Hits@75 (1.8±1.5 versus 1.1±1.2, 3.4±1.7 versus 2.7±1.5, 4.9±2.2 versus 4.1±1.9; Wilcoxon signed-rank P ≤ 2.3×10⁻³).
Perspective
The demonstrated scope is 442 cell types of the adult central nervous system and cell type-specific interaction networks built by intersecting a single-nucleus RNA sequencing atlas with a cell type-agnostic interactome; evaluation covers protein complex prediction, zero-shot target retrieval for five neurodegenerative diseases and essential-gene prediction from a dopaminergic neuron CRISPR screen.
The zero-shot retrieval benchmark measures whether targets already carrying clinical evidence and ChEMBL mechanism annotations place near one another, and leaves open the prioritization of targets not yet pursued; clinical pursuit is also distinct from therapeutic efficacy. Embedding proximity is associative, so predictions are hypotheses requiring experimental testing. Many disease-modifying multiple sclerosis targets act on peripheral immune cells outside the brain cell types SEMPER represents and therefore fall outside its candidate genes.
