GenFish: Generative Target Retrieval from Partially Observed Interactions

We investigate retrieval of additional protein targets for molecules with one known training target. GenFish represents each catalogue protein by a short semantic identifier: a dense dual encoder learns interaction-aware protein vectors, an RQ-VAE converts them into residual-quantization codes, and a molecule-conditioned autoregressive predictor ranks the resulting identifiers. Shared code prefixes let interactions with different targets supervise common prediction steps. On a ChEMBL pair-level development holdout of 2,303 queries and 5,820 candidate targets, GenFish improves mean reciprocal rank from 0.6526 to 0.7127 over a dense baseline. The study explores structured target prediction as an alterna-tive to continuous similarity scoring for completing partially observed interaction records.