Due to low enrichment and high sequence diversity, SELEX-based aptamer discovery is time-consuming and often ineffective at identifying high-affinity binders in early selection rounds. Existing computational methods typically rely on late-round sequence data, limiting their ability to guide refinement during early stages of selection. In this work, we present a ligand- and structure-aware framework for refining early-round aptamer sequences using biochemical and structural features of the target molecule. Our approach leverages TxGemma, a generative model conditioned on the ligand’s molecular descriptors and the aptamer’s predicted secondary structure. Rather than generating sequences de novo, TxGemma selectively edits loop and flanking regions of early-round candidates to improve compatibility with key ligand features. Using Theophylline as a test case, we show that sequences generated from an early round (Round 10) align closely with aptamers enriched in later SELEX rounds (Round 16–20), demonstrating the model’s ability to emulate evolutionary convergence and accelerate aptamer discovery with reduced experimental effort.