Oligonucleotides

Oligonucleotide Therapeutics

The modality that sidesteps the protein entirely. If a target has no druggable pocket, no extracellular epitope, and no ligandable cysteine, an siRNA or antisense oligonucleotide can still silence its transcript — and the design is sequence arithmetic rather than chemistry intuition.

No installation, no network, no key. Sequence tiling, nearest-neighbour thermodynamics, and seed scanning are implemented in the standard library. Transcriptome-wide off-target scanning needs a FASTA that you supply. Thermodynamics: SantaLucia (1998) unified nearest-neighbour parameters.

Read [references/sirna-and-aso-design.md](references/sirna-and-aso-design.md) before choosing a site, [references/chemical-modifications.md](references/chemical-modifications.md) before drawing a pattern, and [references/delivery-and-safety.md](references/delivery-and-safety.md) before committing to the modality — that one is judgement, not syntax, and it is where programmes fail.

The three scripts

| Script | Answers | |---|---| | oligo_design.py | Which sites, and are their thermodynamics right? | | offtarget_scan.py | What else will this silence? | | chemistry_plan.py | What modifications, and where? |

Two mechanisms, two incompatible rule sets

siRNA loads into Argonaute-2 and is cleaved by RISC in the cytoplasm — it needs an RNA-like duplex throughout. Gapmer ASO recruits RNase H1, works in the nucleus, and needs an unmodified DNA core.

Two consequences. ASOs can target introns and pre-mRNA; siRNA cannot, because RISC only sees mature mRNA. And the chemistry is not interchangeable: a DNA gap in an siRNA breaks Argonaute loading, while fully modifying an ASO silently removes RNase H recruitment — the molecule binds its target beautifully and does nothing.

Duplex asymmetry decides which strand is loaded

The siRNA rule that matters most. RISC keeps the strand whose 5' end is less thermodynamically stable. Get it backwards and RISC loads the sense strand, silences something else, and your molecule looks simply inactive — sending you to hunt for delivery problems that do not exist.

python skills/oligonucleotides/scripts/oligo_design.py tile --sequence ACGT... --modality sirna

position  sense                  antisense              gc     tm_c  asymmetry  antisense_loaded  seed     passes  flags
12        CGTCCAGATCGGATCCAAGTT  AACTTGGATCCGATCTGGACG  0.524  73.5  3.2        true              ACTTGGA  true
10        TACGTCCAGATCGGATCCAAG  CTTGGATCCGATCTGGACGTA  0.524  72.6  2.4        true              TTGGATC  false   as_pos1_not_au

The thermodynamics are the SantaLucia 1998 unified parameters and reproduce the paper's worked example exactly — CGTTGA gives ΔH = −41.2 kcal/mol and ΔS = −115.4 cal/mol/K.

GC content is a window, not a direction. Below ~30% the duplex is too weak to hybridise; above ~60% it is too stable for RISC to unwind. Optimising GC upward is a common silent error.

Zero off-targets is not achievable

Antisense positions 2–8 are the seed, and seed pairing with a 3' UTR gives microRNA-like repression with no full-length complementarity at all. A full-length aligner scores that as a non-hit, which is why BLAST is the wrong tool here.

A 7-mer occurs often enough to hit hundreds of transcripts in any real transcriptome. The useful question is comparative:

python skills/oligonucleotides/scripts/offtarget_scan.py seeds --antisense AACTTGG... --fasta tx.fa
python skills/oligonucleotides/scripts/offtarget_scan.py contig --antisense AACTTGG... --fasta tx.fa

contig searches the other risk: RNase H cleaves on partial complementarity, so a contiguous 12–14 nt match elsewhere is a real gapmer hepatotoxicity liability.

The gap must be at least eight DNA residues

python skills/oligonucleotides/scripts/chemistry_plan.py gapmer --sequence GCTAGCTACGTAGCTAGCTA \
    --wing moe --wing-length 5

# 5-10-5 gapmer, MOE wings
# pattern: WWWWWddddddddddWWWWW
# 10 nt DNA gap -- RNase H needs at least ~8 to cleave the heteroduplex
# 1 CpG site(s) marked for 5-methylcytosine. Unmethylated CpG is a TLR9 agonist; this is not optional.

Every 2' modification blocks RNase H, which is the entire reason gapmers have an unmodified core. The script refuses to emit a short gap, because that failure is silent.

Phosphorothioate is the central trade-off. It gives nuclease resistance and the plasma protein binding that drives hepatic uptake — and that same protein binding causes complement activation, thrombocytopenia, and injection-site reactions. The delivery and the toxicity are one mechanism.

Delivery is the whole problem

Every approved siRNA targets a hepatic gene. That is a fact about delivery, not about biology. GalNAc conjugation gives 10–30× potency into hepatocytes via ASGPR and nothing anywhere else.

| Tissue | Status | |---|---| | Liver | solved — GalNAc, subcutaneous, multiple approvals | | CNS | works, intrathecal | | Eye | works, intravitreal | | Muscle, lung, tumour, elsewhere | unsolved |

If the target tissue is not liver, CNS, or eye, say so before designing anything.

Four ways this misleads

1. Accessibility dominates and is not modelled here. mRNA is folded and protein-coated; a

thermodynamically perfect site inside stable secondary structure is inaccessible. Use ViennaRNA or SHAPE data, or tile densely and screen.

1. The rules are necessary, nowhere near sufficient. Published hit rates for rule-compliant

designs run one in three to one in ten.

1. A single designed molecule is not a deliverable. Synthesise and screen 20–50.
2. The essential control is a panel with different seeds. If five sequences produce the

phenotype it is on-target; if one does, it probably is not. Worth more than any prediction here.

Composing with the rest of the bundle

• binding-site-analysis → here: when a target has no druggable pocket, this is one of the

remaining routes.

• target-safety → before: knocking down a constrained gene carries the same warning as

inhibiting one.

• degraders → alongside: the other way to act on an "undruggable" target, at the protein level

rather than the transcript.

• pkpd-translation → after: oligonucleotide PK is unusual — tissue half-lives of weeks decouple

plasma exposure from effect.

Reporting results honestly

Say which modality and why. Give the rules applied and note they are necessary, not sufficient. State that accessibility is not modelled. Report seed off-target counts comparatively, never as an absolute. Name the tissue and route, and if it is not liver, CNS, or eye, say plainly that delivery is unsolved. Recommend a panel and a seed-mismatch control — not a molecule.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: K-Dense-AI
• Source: K-Dense-AI/drug-discovery-agent-skills
• License: MIT
• Homepage: www.k-dense.ai

Install and usage instructions live in the source repository linked above.