GeneSpeak-FP: Target and Compound Retrieval from Observed Cell-Level Perturbation Signatures

arXiv:2607.17671v2 Announce Type: replace Abstract: Single-cell perturbation profiles may help identify known compounds across cellular contexts. We present GeneSpeak-FP, a gene-token Transformer that jointly ranks recorded compounds and their target annotations from a treated cell relative to a cell-line-specific DMSO reference, together with a coarse organ token. On Tahoe-100M, we split 11,673 drug--cell-line groups 90/10 within each compound. No complete group is shared across partitions, although compound and cell-line identities may recur. In a 38,400-query Monte Carlo pass, the correct compound appears first for 12.9% of cells and within the top 10 for 34.3%, among 379 fixed candidates. The model retrieves 40.8% of annotated targets within the top 10 and 54.4% within the top 20, from a vocabulary of 278 genes. Top-10 compound retrieval and target recall are approximately 13 and 11 times their expected values under uniformly random rankings; top-1 compound retrieval is nearly 49 times chance. Top-10 compound retrieval also exceeds the strongest evaluated post-hoc bag-of-genes control by more than nine times. A separate stochastic pass from the same validation pool gives nearly identical aggregate results. These findings support retrospective candidate-space narrowing in held-out drug--cell-line contexts. Because compound identities may recur and target labels are non-exhaustive metadata, they do not establish unseen-compound discovery, causal target identification, or clinical utility.