ASO predictions.
Molecular context.

Beyond a ranked list

For an antisense oligonucleotide (ASO) researcher, the useful output is more than which candidate ranks first. It is a comparison: what differs between two designs, and what evidence makes that difference worth examining?

Our earlier article on evidence-led oligonucleotide design introduced the broader KumiChem workflow. Here we focus on one component: connecting toxicity-related predictions to chemical and sequence differences that researchers can recognize.

A score alone compresses those differences into one number. A comparison preserves them. Two ASOs may share a base sequence but use different sugars, or share the same chemical pattern while placing bases in different regions. A useful system needs to retain both kinds of information and connect them to the experiment the researcher cares about.

Chemical comparisons

The molecular representation records bases, sugar modifications, backbone linkages and their positions. It also describes combinations, such as neighbouring modified sugars or a short sequence pattern within the central DNA region. These numerical features, or descriptors, let the model distinguish molecules that an ordinary base sequence would treat as identical.

A HeLa cell experiment makes the distinction concrete. Two ASOs share the sequence 5′–TACTGCCTTCTTA–3′ and the same phosphorothioate (PS) backbone. Replacing the locked nucleic acid (LNA) sugar at residue 12 with DNA sugar changes the terminal three-sugar pattern from LNA–LNA–LNA to LNA–DNA–LNA. At 100 nM transfection and a 24-hour measurement, the recorded caspase response falls from 0.256 to 0.078. Models trained without either molecule or its base sequence predict the same direction, 0.285 to 0.145. [1]

The edit does two things together: it removes one LNA sugar and breaks a continuous LNA stretch. The measured contrast captures amount and arrangement together, and the predicting models use features describing both. For design review, this is a more informative comparison than “LNA versus DNA”: it preserves the amount and arrangement of the modification alongside the measured response.

More specifically, the LNA count falls from five of 13 sugars to four of 13, adjacent LNA pairs fall from two to zero, and the longest continuous modified stretch falls from three residues to one. These are linked descriptions of the same edit. They explain why a system should retain local sugar arrangement as well as total modification content when comparing designs with the same bases.

A TLR8 reporter panel illustrates the complementary case. Its molecules all use the same sugar and backbone design. Three selected molecules each contain two occurrences of 5′–TCC–3′, yet their measured reporter responses differ. Their central DNA regions contain one, two and four thymine bases, accompanying responses of 1.00, 1.70 and 3.95. Regional composition adds information that motif count alone misses in this comparison. The assay measures enhancement of the response to agonist R848, with R848 alone set to 1.0. [2]

Together, these examples show why chemistry and sequence belong in the same molecular description. A fixed sugar pattern provides the background for interpreting sequence variation. A sugar change within a shared sequence provides a chemical contrast. The three TLR8 records illustrate a local association, not a general rule that adding thymine increases response.

Assay-aware priorities

A useful comparison also needs the right measurement. Lower caspase response in HeLa, lower liver-enzyme activity and lower TLR8 reporter enhancement answer different questions. Their scores should stay attached to their assay, response direction and experimental context rather than become one universal “toxicity” label.

Our retrospective benchmark contains 13,304 assay records across nine endpoints. These cover HeLa cytotoxicity, neuronal calcium response, mouse behavioral response to 2′-O-methoxyethyl (2′-MOE)-modified ASOs, TLR7 inhibition, TLR8 potentiation, mouse and rat alanine aminotransferase (ALT) measurements, and mouse and rat behavioral outcomes.

Against the original studies’ models rerun on the same test records, mean Spearman rank correlation increased from 0.509 to 0.603 (+0.094), with improvement in all nine endpoints. Spearman measures agreement between predicted and measured ordering. Evaluation retained source-defined sequence-cluster or patent groups, with five partitions per endpoint. Assigned test groups were excluded from fitting, although some groups share base sequences. These results support ranking within the retrospective assay settings, rather than prospective or external validation.

The largest gains are in mouse ALT (+0.183) and rat ALT (+0.163). Their final correlations are 0.544 and 0.321, respectively. This distinction matters: a substantial improvement over a weak baseline does not make the resulting predictor equally strong in both species. Rat ALT remains the weakest ranking result in this benchmark.

HeLa cytotoxicity rises from 0.588 to 0.732 (+0.144), supporting better ordering in the same cellular assay used for the sugar-pattern example. The LNA neuronal-calcium endpoint improves from 0.381 to 0.445 (+0.063), while the 2′-MOE mouse behavioral assay rises from 0.627 to 0.674 (+0.047). These different readouts should remain separate when a researcher weighs cellular and neurological evidence.

Mouse and rat behavioral endpoints reach 0.756 and 0.643, with gains of +0.074 and +0.086. TLR8 potentiation improves from 0.657 to 0.709 (+0.052), and TLR7 inhibition from 0.570 to 0.602 (+0.032). The receptor assays describe different response directions, so a shared increase in ranking performance is not a shared recommendation about whether to increase or decrease a biological response.

For system design, the implication is practical: keep the assay identity and meaning of its score alongside every candidate. A researcher can then consider that assessment with activity, sequence specificity and project constraints without treating different biological readouts as interchangeable.

Evidence an agent can use

To support that comparison, the output needs more than a score and a list of feature names. Our research workflow assembles the specified molecule, assay context, predicting-model identity, feature contributions, measured comparators and source references into a structured evidence record.

Feature contributions describe which inputs raise or lower the model’s prediction. For the HeLa pair, that record connects the residue-12 edit to LNA amount and the interrupted terminal stretch, then places the predicted difference beside the measured caspase difference. Programs package these records into explanation templates that an AI assistant can use to draft a molecular explanation for scientific review.

For an agent-assisted design pipeline, this gives subsequent steps something concrete to work with. A comparison can be framed as keeping the base sequence fixed while examining a sugar pattern, or keeping chemistry fixed while comparing regional sequence. The model features explain the estimate, the measured compounds establish the observed contrast, and literature provides biological context. Each contributes a different part of the design assessment.

Value for design workflows

The main design lesson is to preserve the molecular comparison through the workflow. Chemical identity is retained at input, the prediction stays attached to an assay, and the explanation points back to recognizable structural differences and measured compounds.

This makes candidate review more specific: which sugar arrangement distinguishes these molecules, which regional sequence information matters in this panel, and what measurement supports the comparison? Those questions connect prediction to design without reducing all the evidence to a single ranking.

Most benchmark molecules are gapmers, with a central DNA region and modified wings. The nine endpoints describe research coverage, not a list of commercially available tests. Code and model cards document the evaluated workflows, assay scope and reuse terms.

The aim is a prediction that helps researchers understand a molecular choice—not just a score that orders a list.