The fatty acid has stopped being a passive greasy tail and become one of the most actively engineered parts of a modern drug. Across 2025 and 2026, three things happened at once: lipid chemistries that no organism makes (brominated, fluorinated, branched, sterol-fused) outperformed clinical benchmarks; AI models and self-driving labs began proposing those chemistries instead of chemists; and lipid conjugation moved oligonucleotides and antibodies into tissues that receptor-targeted ligands could not reach.
Lipidation: the fatty acid as a pharmacokinetic switch
Attaching a fatty acid to a peptide buys it days of circulation, because the conjugate rides on human serum albumin instead of being filtered by the kidney. Unmodified peptides below roughly 5 kDa are cleared within minutes; a fatty acid promotes reversible albumin binding and creates a protected circulating drug-albumin pool. Albumin carries nine fatty-acid binding sites, and under normal physiology only two are occupied, so a lipidated drug is not competing with the body's own cargo.
The design levers are chemical rather than biological, and they are now well mapped. Chain length and saturation set the strength of the albumin grip: insulin detemir used myristic acid, a plain C14, while semaglutide moved to a C18 diacid that binds harder and stretches plasma half-life to around 46 hours in mini-pigs. Between the peptide and that acyl tail sits a linker, usually a γ-glutamate with AEEA or ADO spacers, whose job is to preserve receptor potency and solubility while letting the tail reach albumin. Where the tail attaches matters as much as what it is. Random acylation costs activity, whereas site-specific conjugation through click chemistry at a permissive residue took one small protein binder from a twenty-minute half-life to 10.7 hours, with a matching gain in tumour suppression.
The trick does not scale indefinitely, which is the part most often glossed over. Large proteins carrying a fatty acid start competing with FcRn for albumin, because the primary fatty-acid sites overlap the FcRn interface; proteins around 28 kDa escape that competition, and for them even linker length barely moves the result.
The commercial proof is the metabolic class: liraglutide, semaglutide and tirzepatide are all products of carefully engineered acylation, not of a better peptide sequence.
Lipid conjugates as the address label for nucleic acids
GalNAc solved the liver. Everything outside the liver is currently being solved with fatty acids, because there is no comparable receptor-targeted ligand for heart, muscle, brain or eye. Hydrophobicity, valency and spatial arrangement of the lipid do the routing instead.
The workhorse is C16, a saturated sixteen-carbon ether hung off the siRNA, which reaches the central nervous system, eye and lung and keeps silencing for at least three months in non-human primates. Docosahexaenoic acid, the polyunsaturated fatty acid the brain is already full of, drives conjugates into striatum and cortex and gets them taken up by neurons and astrocytes. Docosanoic acid, a saturated C22, favours skeletal and cardiac muscle. Divalent C18 structures and dendritic amphiphiles bearing four C12 chains lean on high-affinity albumin binding to reach tumour, heart, fat and lung.
The clinical anchor is mivelsiran (formerly ALN-APP), the first C16-siRNA in humans. Intrathecal dosing at 75 mg produced mean reductions of about 61% in soluble APPα and 74% in soluble APPβ at one month in a 36-patient cohort, with reductions sustained for around six months. Alnylam reported in July 2025 that single and multiple doses were generally well tolerated with no signal of increased amyloid-related imaging abnormalities, and completed enrolment in the Phase 2 cAPPricorn-1 trial in July 2026.
A 2026 preprint sharpened the design rules with a modular click-chemistry platform: across seven lipid ligands tested at different valencies and positions, 3',5'-bis-conjugates consistently beat 5'-mono conjugates, and C16 was the strongest ligand for heart and skeletal muscle when presented as a bis-conjugate or as a branched 5' conjugate with a PEG4 arm. Head-group chemistry matters independently of hydrophobicity: lengthening a phosphatidylcholine linker on a docosanoic-acid conjugate raised kidney, pancreas and heart accumulation several-fold over the parent.
Tails no organism makes
Inside a lipid nanoparticle, the ionizable lipid's hydrophobic tail is the part most open to invention. It sets molecular geometry, packing, apparent pKa and endosomal escape, and unlike the head group it has almost no biological constraint on its structure. The 2025–2026 literature reads as a systematic march away from natural fatty acids.
Branching came first. Adding alkyl side chains along the main chain changes the lipid's cone geometry and its fusogenicity, and the α-branched fatty acids behind it are made by alkylating dimethyl malonate or a linear fatty acid with iodoalkanes, then esterifying the product onto an ionizable head. A 2026 platform went further and fused a cholesterol-derived tail and a branched dialkyl tail onto the same molecule, tuning packing and protonation together; the lead, L1-aCho-e3, out-transfected SM-102, held apparent pKa between 6.2 and 6.8 with encapsulation above 90%, and swung biodistribution from liver to lung or spleen when DOTAP or DOPA was added as a fifth component.
Halogens were the next departure from anything metabolic. Fluorination both promotes cytosolic delivery and stabilises the particle, and fluorinated ionizable lipids produced spleen-targeted mRNA delivery for cancer immunotherapy; a fluorinated system published in PNAS in January 2026 doubles as a 19F NMR and MRI tracer, so biodistribution and expression can be watched without killing the animal. Brominated tails arrived by a stranger route, described in the next section. Running alongside all of this, natural polyunsaturated acids — EPA, DHA, ALA, GLA — are being bolted onto branched ester scaffolds by Michael addition, where a metabolite is treated purely as an engineering part.
One caution runs through this work: potency and tolerability are separable. A 2025 study varying tail architecture alone, including disulfide-containing biodegradable tails, cut cytotoxicity and cytokine induction and balanced Th1/Th2 activation in non-human primates while keeping translation high.
AI-designed lipids, and what simulation adds
The field crossed a line in 2025–2026. Models stopped ranking chemist-proposed libraries and started proposing chemistries humans had not considered.
MIT's LiON, published in Nature Biotechnology in 2025, trained a directed message-passing neural network on more than 9,000 LNP activity measurements, screened 1.6 million virtual lipids, and surfaced FO-32 and FO-35, which delivered mRNA to mouse and ferret lung and to nasal mucosa; FO-35 turned out to have a terpene-based structure no one had proposed. Toronto's LUMI-lab, in Cell in 2026, coupled a transformer foundation model pretrained on 28 million molecules to a robotic active-learning loop, and over ten cycles synthesised and screened more than 1,700 LNPs without human intervention. Two data-poor workarounds appeared alongside them: LipidAI tripled a sparse training set by shifting methyl groups on tail chains, and an NCNST group converted molecular dynamics trajectories into 2D density images before training on them, finding that lipid P1's stable three-tail cone drives IgM corona formation and spleen targeting.
Two results deserve to be pulled out of that paragraph. LUMI-lab's brominated tails were, by the authors' account, unreported in the prior literature: the system found a structural feature rather than an optimum within a known one. Five of its top six lipids matched or beat SM-102, the ionizable lipid in Moderna's COVID-19 vaccine, and inhaled LNPs built on the best of them, LUMI-6, achieved 20.3% CRISPR editing in mouse lung epithelium. The NCNST work makes a quieter but broader point: the representation mattered more than the model. Switching from a lipid's 2D chemical structure to its dynamic 3D conformation under physiological conditions is what made organ targeting predictable at all.
Simulation is the quiet half of this. Coarse-grained molecular dynamics now reaches the length and time scales of actual LNP self-assembly. Martini-based and OPLS4-based coarse-grained models reproduce pH-driven reorganisation — bilayers at the aqueous interface with an amorphous core at pH 4, bulk amorphous structure at pH 7.4 — and have spontaneously reproduced the bleb structures seen experimentally. For permeation enhancers, coarse-grained simulation of medium-chain fatty acids in bile-containing media explains membrane incorporation that cell assays alone cannot. These trajectories are increasingly the training data for the screening models above, which is why the two halves are converging.
Antibodies and antibody derivatives meet the lipid
Three distinct convergences are running at once, and they are often confused with each other.
Antibodies on the particle. Ab-LNPs give specificity that liver-tropic lipids cannot. The 2026 constraint was reproducibility, not concept: a Purdue group published a post-conjugation process that tunes antibody surface density on finished LNPs and quantifies it reliably, which is what a rational Ab-LNP design framework needs. An alternative route avoids conjugation entirely by capturing unmodified antibodies in their correct orientation on the particle surface, preserving affinity. Note the geometry cost: an IgG is 10–15 nm, so conjugation typically adds 20–30 nm to particle diameter, which itself changes immune-cell uptake.
Antibodies as the payload. Rather than delivering to a cell, the LNP makes the antibody in the patient. mRNA-LNP-encoded monoclonals conferred protection against SARS-CoV-2 and influenza in a study collected in June 2026, potentially bypassing recombinant manufacturing. The fatty-acid chemistry of the carrier, not the antibody sequence, is what limits this route.
Antibodies designed by diffusion models. In November 2025, a fine-tuned RFdiffusion network produced VHHs, scFvs and full antibodies binding user-specified epitopes entirely in silico, with cryo-EM confirming the designed CDR conformations for an influenza haemagglutinin binder. The interesting prospect is the stack: a de novo binder, a lipid tail for albumin-mediated protraction, and an AI-designed carrier — all three components generated computationally, none of them found in nature.
Fatty acids that open the gut wall
A separate use of the same molecules: medium-chain fatty acids as transient permeation enhancers for oral biologics. Sodium caprate (C10), sodium caprylate (C8) and salcaprozate sodium (SNAC, a caprylate derivative) are the three with clinical validation — SNAC in oral semaglutide, C8 in oral octreotide.
The honest number is sobering. Three decades of work put all three enhancers at roughly 1% oral bioavailability with high variability; oral semaglutide lands around 0.4–1%, meaning over 99% of the dose never reaches circulation. Rybelsus works commercially because semaglutide's half-life and therapeutic index absorb that inefficiency, not because the enhancer is good.
The 2025 move was combination. Niu et al. formulated SNAC and C10 in one tablet for a GLP-1/GIP co-agonist, on the logic that SNAC opens a gastric absorption window and C10 an intestinal one. Dose-corrected AUC exceeded either enhancer alone in preclinical models, with supportive early clinical PK; the two also improve each other's solubility in simulated gastric fluid. The authors are careful that this is not a universal formulation — species and API differences dominate. For nucleic acids specifically, oral delivery via these enhancers remains substantially harder than for peptides.
What is still hard
Data scarcity remains the binding constraint on the AI side. LUMI-lab and LipidAI exist in their particular forms because lipid datasets are small — one answers that with foundation-model pretraining on 28 million unrelated molecules, the other with an augmentation trick — and neither is a substitute for measurement. The measurements themselves are also suspect: most published libraries are ranked in HeLa or bronchial epithelial cells, and the tissue that matters is rarely the tissue screened. In vitro potency does not predict organ targeting in any case; the protein corona does, and the spleen tropism in the 2026 conformation study came from IgM adsorption driven by lipid shape, which no cell-based assay reports.
Synthesisability is the recurring embarrassment of generative chemistry. Models happily propose lipids no one can make, which is why synthesis-constrained generation and combinatorial-chemistry-plus-ML approaches keep outperforming free generation. And tolerability scales with novelty. Brominated and fluorinated tails have no metabolic precedent; their degradation products, accumulation and immunogenicity are largely uncharacterised, and the safety work published so far is mostly rodent and short-term. The chemistry has run ahead of the toxicology, and the next two years will be spent finding out by how much.
This article is written for general scientific and educational information only. It is not medical advice, a treatment recommendation, or professional guidance of any kind, and it must not be used to diagnose, treat, or make decisions about any health condition. Speak to a qualified clinician before acting on anything here.
Several technologies described are investigational. Mivelsiran, the AI-designed lipids FO-32, FO-35, LUMI-6, P1 and L1-aCho-e3, the fluorinated and brominated tail chemistries, the antibody-conjugated and antibody-encoding LNPs, and the SNAC/C10 combination tablet are all preclinical or in clinical trials as of 26 September 2026. None of them is an approved medicine, and results in mice, ferrets or non-human primates routinely fail to translate to humans. Approved products named here (insulin detemir, liraglutide, semaglutide, tirzepatide, patisiran, givosiran, Rybelsus, Mycapssa) are mentioned as engineering examples, not as endorsements.
Figures, doses and percentages are reproduced as reported by the cited sources and have not been independently verified. Links point to the original papers, which were consulted on 26 September 2026; findings may have been superseded, corrected or retracted since. Preprints cited here have not completed peer review. Where a company filing is cited, it reflects that company's own disclosure.
No commercial relationship, sponsorship or funding influenced this piece, and no confidential or proprietary information was used. Some cited authors have filed patents on the work described; those declarations sit in the original publications. Views expressed are the author's own.
Sources
1. Xu Y, Cui H, Pang K, et al. LUMI-lab: a foundation model-driven autonomous platform enabling discovery of ionizable lipid designs for mRNA delivery. Cell 2026;189:1620–1635.e25. doi:10.1016/j.cell.2026.01.012
2. Witten J, Raji I, Manan RS, et al. Artificial intelligence-guided design of lipid nanoparticles for pulmonary gene therapy. Nat Biotechnol 2025;43:1790–1799. doi:10.1038/s41587-024-02490-y
3. Artificial intelligence-guided design of LNPs for in vivo targeted mRNA delivery via analysis of the spatial conformation of ionizable lipids. Nat Biomed Eng 2026. doi:10.1038/s41551-026-01640-8
4. Duan X, Chen J, Jiang S, et al. A novel machine learning framework-based rapid screening of ionizable lipids in LNPs for highly-efficient mRNA expression. Acta Pharm Sin B 2026;16:5053–5069. doi:10.1016/j.apsb.2025.10.040
5. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 2025. doi:10.1038/s41586-025-09721-5
6. Qu F, Fosu K, Wagner AA, et al. Post-conjugation process for antibody-conjugated lipid nanoparticles enabling tunable antibody surface density for targeted RNA delivery. ACS Nano 2026;20:15092–15102. doi:10.1021/acsnano.5c21719
7. A versatile antibody capture system drives specific in vivo delivery of mRNA-loaded lipid nanoparticles. Nat Nanotechnol 2025. doi:10.1038/s41565-025-01954-9
8. Fluorinated lipid nanoparticles enable real-time tracking of mRNA delivery and uncover spatiotemporal mechanisms of immune activation. PNAS 2026. doi:10.1073/pnas.2519823123
9. Spatial tail design in ionizable lipids enhances the safety and efficacy of mRNA delivery. 2026. PMC13397292
10. Titze-de-Almeida R, Gomes GMO, dos Santos TC, Titze-de-Almeida SS. C16-siRNAs in focus: development of ALN-APP. Pharmaceuticals 2026;19:26. doi:10.3390/ph19010026
11. Brown KM, et al. Expanding RNAi therapeutics to extrahepatic tissues with lipophilic conjugates. Nat Biotechnol 2022;40:1500–1508.
12. Click chemistry-based strategy for modular ligand attachment to siRNAs: toward extrahepatic RNAi. bioRxiv 2026.
13. Bioconjugates for improved delivery of oligonucleotide therapeutics to the central nervous system. 2026. PMC12990044
14. Niu Z, La Zara D, Blaabjerg L, et al. Combining SNAC and C10 in oral tablet formulations for gastric peptide delivery. J Control Release 2025;378:92–102.
15. Bohley M, et al. Gastrointestinal permeation enhancers beyond sodium caprate and SNAC. Adv Sci 2024.
16. Structuring of lipid nanoparticle mRNA formulations at acidic and neutral pH: X-ray scattering and molecular dynamics studies. Mol Ther Methods Clin Dev 2025.
17. A review of lipidation in the development of advanced protein and peptide therapeutics. PMC7520907
18. Alnylam Pharmaceuticals. Form 10-Q, FY2026 (quarter ended 30 June 2026). US SEC.
","deck":"","published_at":"2026-09-26T12:25:25.875Z","title":"Fatty acids as a design variable","url":"https://kandulamahesh.com/insights/fatty-acids-as-a-design-variable"}