Chemical similarity-based approaches employed to repurpose or develop new treatments for emerging diseases, such as COVID-19, correlates molecular structure-based descriptors of drugs with those of a physiological counterpart or clinical phenotype. We propose novel descriptors based on a COSMO-RS (short for conductor-like screening model for real solvents) σ-profiles for enhanced drug screening enabled by machine learning (ML). The descriptors' performance is hereby illustrated for nucleotide analogue drugs that inhibit the ribonucleic acid-dependent ribonucleic acid polymerase, key to viral transcription and genome replication. The COSMO-RS-based descriptors account for both chemical reactivity and structure, and are more effective for ML-based screening than fingerprints based on molecular structure and simple physical/chemical properties. The descriptors are evaluated using principal component analysis, an unsupervised ML technique. Our results correlate with the active monophosphate forms of the leading drug remdesivir and the prospective drug EIDD-2801 with nucleotides, followed by other promising drugs, and are superior to those from molecular structure-based descriptors and molecular docking. The COSMO-RS-based descriptors could help accelerate drug discovery for the treatment of emerging diseases.