ASOkai: A Modular Open-Source Framework for Systematic ASO Design and Evaluation

Poster presented on June 12, 2026 by Seyedarash Ayatollahi at 5th Dutch Antisense Therapeutics Symposium, Utrecht, Netherlands.

Abstract: The design of effective antisense oligonucleotide (ASO) therapeutics is challenging: the combinatorial space of candidate binding sites, chemical modification patterns, and backbone configurations grows rapidly with gene and oligonucleotide length, making exhaustive experimental screening impractical. Furthermore, prioritizing these candidates requires evaluating them against complex sequence, thermodynamic, and specificity constraints. In silico approaches have therefore become increasingly important, with tools such as PFRED[1] and ASOptimizer[2] addressing distinct aspects of ASO design. However, no single tool offers a unified, reproducible, extensible, and automation-ready framework. Here we present ASOkai, an open-source Python workflow framework for systematic ASO design and evaluation. ASOkai is built to support end-to-end ASO evaluation while remaining flexible to new methods and use cases. The platform organizes candidate analysis into complementary categories, including intrinsic sequence features, off-target risk profiling, target accessibility, and kinetic modeling of target knockdown, informed by sequence, thermodynamic, and chemical features relevant to ASO efficacy. Specialized for modularity and interoperability, ASOkai supports both standalone use and integration into existing computational biology workflows. In a typical ASOkai workflow for RNase-H1 mediated knockdown, the user provides a target gene from which ASOkai extracts and filters candidate binding sites based on criteria such as GC content and genome-wide uniqueness. Filtered candidates undergo screening for target accessibility and availability of repeated sites prone to RNase-H1 cleavage, followed by kinetic modeling of target knockdown using unspecific off-target profiles, and unintended gene knockdown arising from specific off-target interactions. A machine learning model trained on empirical antisense data then combines these multidimensional evaluations to identify optimal binding sites and their corresponding oligonucleotides.

Download PDF of Poster.