There are more amino acid permutations within a 40-residue sequence than atoms on Earth. This vast chemical search space hinders the use of human learning to design functional polymers. Here we couple supervised and unsupervised deep learning with highthroughput experimentation to drive the design of high-activity, novel sequences reaching 10 kDa that deliver antisense oligonucleotides to the nucleus of cells. The models, in which natural and unnatural residues are represented as topological fingerprints, decipher and visualize sequenceactivity predictions. The new variants boost antisense activity by 50-fold, are effective in animals, are nontoxic, and can also deliver proteins into the cytosol. Machine learning can discover functional polymers that enhance cellular uptake of biotherapeutics, with significant implications toward developing therapies for currently untreatable diseases.One sentence summary: Deep learning generates de novo large functional abiotic polymers that deliver antisense oligonucleotides to the nucleus. typically occurs by energy-dependent uptake, meaning that endosomal escape presents an additional challenge. While some peptides can efficiently escape the endosome, designing a novel CPP sequence for this task is nearly impossible. In addition to the diversity of physicochemical properties of CPPs, variation in experimental design has resulted in inconsistent and sometimes contradictory datasets.(22) These inconsistencies preclude establishing sequence-activity relationships to guide the design of next-generation CPPs and can be remedied by testing PMO-CPP conjugates in a nuclear delivery-based assay that provides quantitative activity data and selects for sequences that can escape the endosome. In order to uncover CPP design principles for PMO delivery, it is necessary to have a standardized, biologically relevant dataset with which to train machine learning models.