Presentation Information
[3FMBS-04]RaptScore: a large language model-based algorithm for versatile aptamer evaluation
○Akira Kimura-Yamazaki1, Tatsuo Adachi2, Shigetaka Nakamura2, Yoshikazu Nakamura2, Michiaki Hamada1,3,4 (1. Graduate School of Advanced Science and Engineering, Waseda University (Japan), 2. RIBOMIC (Japan), 3. National Institute of Advanced Industrial Science and Technology (Japan), 4. Graduate School of Medicine, Nippon Medical School (Japan))
Keywords:
RNA aptamer,drug discovery,large language model,AI,SELEX
[Purpose]RNA aptamers are a high-potency tool in the life sciences, offering promising applications in drug discovery and beyond. They are typically obtained through systematic evolution of ligands by exponential enrichment (SELEX), which imposes constraints on sequence length and diversity. Several metrics, such as frequency and enrichment, have been developed to identify high-activity aptamers from SELEX. However, existing evaluation metrics are limited to sequences that appear within SELEX and cannot assess sequences of varying lengths, limiting their utility in optimizing aptamer design. To overcome these limitations, we developed RaptScore, a novel binding activity evaluation metric leveraging large language models.
[Method]RaptScore is a modified pseudo-log-likelihood (PLL) score computed using a DNABERT model continually pretrained on SELEX-derived sequence data. For calibration of the scoring settings, representative sequences selected from each SELEX round based on frequency and enrichment were experimentally evaluated by surface plasmon resonance, and the setting showing the highest correlation with binding activity was selected. The optimized score was then applied to downstream tasks such as sequence ranking, truncation analysis, genetic algorithm-based maturation, and integration with other deep learning-based aptamer discovery tools.
[Results]Across three SELEX datasets, the best Pearson correlations between RaptScore and experimentally measured binding activity were 0.65, 0.78, and 0.65. In truncation analyses, high-scoring shortened variants frequently retained activity comparable to, or greater than, that of the original aptamers. Furthermore, integration with genetic algorithm-based maturation enabled the identification of shortened aptamer variants that maintained binding activity despite an approximately 30% reduction in the length of the random region. When combined with RaptGen for candidate prioritization, RaptScore also retained approximately 80% of the top five binders within the top half of the ranked generated sequences.
[Consideration]These results suggest that RaptScore can serve as a practical surrogate for aptamer binding activity beyond conventional SELEX-derived metrics such as frequency and enrichment. However, its performance is influenced by parameter selection, and its sequence-based design does not explicitly account for structural features important for aptamer-target interactions. Therefore, careful calibration and future incorporation of structural information will be important for further improving its utility.
[Conclusion]RaptScore provides a versatile framework for evaluating, optimizing, and truncating RNA aptamers, addressing key limitations of conventional aptamer binding metrics. Its ability to assess sequences beyond SELEX data and integrate into computational optimization pipelines positions it as a valuable tool in RNA-based therapeutic and diagnostic development.
[Method]RaptScore is a modified pseudo-log-likelihood (PLL) score computed using a DNABERT model continually pretrained on SELEX-derived sequence data. For calibration of the scoring settings, representative sequences selected from each SELEX round based on frequency and enrichment were experimentally evaluated by surface plasmon resonance, and the setting showing the highest correlation with binding activity was selected. The optimized score was then applied to downstream tasks such as sequence ranking, truncation analysis, genetic algorithm-based maturation, and integration with other deep learning-based aptamer discovery tools.
[Results]Across three SELEX datasets, the best Pearson correlations between RaptScore and experimentally measured binding activity were 0.65, 0.78, and 0.65. In truncation analyses, high-scoring shortened variants frequently retained activity comparable to, or greater than, that of the original aptamers. Furthermore, integration with genetic algorithm-based maturation enabled the identification of shortened aptamer variants that maintained binding activity despite an approximately 30% reduction in the length of the random region. When combined with RaptGen for candidate prioritization, RaptScore also retained approximately 80% of the top five binders within the top half of the ranked generated sequences.
[Consideration]These results suggest that RaptScore can serve as a practical surrogate for aptamer binding activity beyond conventional SELEX-derived metrics such as frequency and enrichment. However, its performance is influenced by parameter selection, and its sequence-based design does not explicitly account for structural features important for aptamer-target interactions. Therefore, careful calibration and future incorporation of structural information will be important for further improving its utility.
[Conclusion]RaptScore provides a versatile framework for evaluating, optimizing, and truncating RNA aptamers, addressing key limitations of conventional aptamer binding metrics. Its ability to assess sequences beyond SELEX data and integrate into computational optimization pipelines positions it as a valuable tool in RNA-based therapeutic and diagnostic development.
Comment
To browse or post comments, you must log in.Log in