Design choices in candidate selection
Every oligonucleotide design reflects choices about the target site, sequence, and chemical modifications. Each choice should have a clear scientific rationale.
In an antisense oligonucleotide (ASO) or small interfering RNA (siRNA) program, a candidate list opens a series of decisions. Which sequences should move into the first screen? What evidence supports them? Which uncertainties should that screen resolve?
DLeader grew out of work in nucleic acid chemistry and machine learning. We are building KumiChem to help researchers answer those questions by connecting candidate design to its evidence, assumptions, and experimental priorities.
Integrating evidence and uncertainty
A research team may consult literature, transcript databases, design tools, structural analyses, and prediction models before choosing a candidate. Each output answers part of the question. The decision depends on how those answers fit together.
An activity prediction may favor a candidate while a sequence-based off-target screen flags potential unintended interactions. These flags require interpretation and experimental follow-up. Changing the target transcript may also change which sites are relevant. Candidate selection therefore requires explicit consideration of uncertainty and the intended experimental context.
Experimental studies have demonstrated siRNA seed-mediated off-target transcript silencing [1] and chemistry-dependent differences in splice modulation for ASOs with the same sequence in retinal models [2]. These findings underline the importance of a defined chemical and biological context when evaluating candidates.
Each design should show what supports it, what remains uncertain, and what to test next.
Connecting candidate design and evidence
KumiChem is our beta AI agent platform for oligonucleotide drug discovery, initially focused on ASO and siRNA. It connects candidate generation and computational evaluation within a shared workflow. Our development goal is a prioritized candidate list that researchers can inspect and use to plan experimental testing.
Candidate design defines the target, sequence constraints, and chemistry under consideration. Evidence evaluation examines how predictions, published findings, and experimental observations support those choices. Our aim is to keep these activities connected throughout the workflow.
“Which candidates will give us an informative first screen?”
A reviewable shortlist should let the team inspect three things:
- Design definition. The target transcript, candidate-site assumptions, and chemistry under consideration.
- Supporting evidence. The analyses and observations behind each candidate, including conflicting results and unresolved risks.
- Experimental question. What the first experiment should establish, and how its results could change the design.
Principles for transparent decision support
Traceability of inputs and methods
Our aim is to retain sequence identifiers and versions, tool and model versions, key parameters, and source references. Researchers should be able to inspect the analyses underlying a recommendation and reproduce them where feasible.
Separation of prediction and experimental evidence
A computational ranking supports a prioritization hypothesis. Candidate activity and off-target effects require experimental evaluation under specified conditions. We aim to distinguish model predictions, literature-derived evidence, and experimental measurements throughout the workflow.
Researcher-defined objectives and criteria
The team defines the target, constraints, and evaluation criteria. Scientists assess the trade-offs and choose the experiments. KumiChem should give them a stronger basis for those choices.
Refining designs through experimental evidence
Our next priority is evaluation with research partners. The target, candidate chemistry, assay conditions, endpoints, and comparison baselines should be agreed in advance. Where models are evaluated, test data should be kept separate from model development. Prospective experiments can then assess candidate performance and identify assumptions that need revision.
In this blog, we will share design decisions, evaluation results, and lessons from that process.
Our goal is a design–evidence loop: use evidence to guide candidate design, then use experimental findings to refine the next design.
Selected background literature
- Jackson AL et al. (2006). Widespread siRNA “off-target” transcript silencing mediated by seed region sequence complementarity. RNA 12(7):1179–1187.
- Vázquez-Domínguez I et al. (2024). Efficacy, biodistribution and safety comparison of chemically modified antisense oligonucleotides in the retina. Nucleic Acids Research 52(17):10447–10463.