ADC Design: DAR, Linker and Payload Liabilities

You have a target-validated antibody and a payload, and now someone has to decide the drug-to-antibody ratio, the conjugation chemistry and the linker. Get it wrong and you find out six months later in a HIC trace, an aggregation shoulder on SEC, or free payload in plasma. Most of the early decisions can be pressure-tested from sequence and SMILES before anyone books conjugation lab time.

This walks through what you can and cannot get from a sequence-level ADC assessment, using real outputs from CovaSyn's covabio_adc on the trastuzumab heavy chain with MMAE as payload.

The four decisions that define an ADC

1. Payload class. Potency, mechanism, membrane permeability (bystander effect) and hydrophobicity. Auristatins, maytansinoids and camptothecin derivatives behave very differently in all four. 2. Linker. Cleavable (protease, acid, disulfide) versus non-cleavable. Drives systemic stability and whether you get bystander killing. 3. Conjugation site. Interchain cysteine, engineered cysteine, lysine, or enzymatic/glycan site-specific. 4. DAR. Average payloads per antibody, and just as importantly the *distribution* around that average.

They are coupled. A hydrophobic payload plus a high DAR plus stochastic lysine conjugation is the classic route to an aggregation-prone, fast-cleared molecule.

Worked example: trastuzumab heavy chain + MMAE

Input to covabio_adc: the trastuzumab IgG1 heavy chain (450 aa, a public therapeutic sequence used here as a realistic test case) and the MMAE SMILES resolved from PubChem CID 11542188. Verbatim outputs, three target DAR values:

Site counts, constant across the DAR sweep:

• conjugatable_cysteines: 11 total cysteines, 6 estimated conjugatable, method interchain_disulfide_estimate
• conjugatable_lysines: 32

Run the light chain (214 aa) separately and you get antibody_mw_da 23442.81, 5 cysteines, 13 conjugatable lysines. The tool scores one sequence at a time, so an intact IgG1 means running heavy and light and adding the chains yourself. The estimated_adc_mw is a mass balance of antibody plus n payloads; it does not include linker mass, so treat it as a floor rather than an expected deconvoluted mass.

What the numbers actually tell you

The useful signal is the 32 lysines against 6 conjugatable cysteines on the heavy chain. That is the whole argument for cysteine over lysine conjugation in one line. Stochastic lysine conjugation at DAR 4 draws from a pool of dozens of reactive amines, which is why lysine-conjugated ADCs show broad, heterogeneous species distributions. Reduced-interchain-cysteine conjugation draws from a much smaller, better-defined pool and gives the even-numbered DAR 0/2/4/6/8 ladder you see on HIC.

The second signal is the DAR-dependent aggregation call flipping from low at DAR 2 and 4 to moderate at DAR 8, with the recommendation to monitor the HIC profile. That matches what the field has published: brentuximab vedotin and ado-trastuzumab emtansine both sit at an average DAR in the 3.5 to 4 range, and the DAR 8 subspecies of vc-MMAE conjugates were reported to clear faster and be less tolerable than DAR 2 to 4 species in preclinical work. Higher DAR is only viable when the payload and linker are engineered for it, which is what the DAR 8 topoisomerase-I ADCs did with a much more polar linker-payload.

The payload-class caveat, stated plainly

We ran the same antibody with camptothecin (payload_mw_da 348.36) at DAR 8. The output was identical on the risk fields: payload_hydrophobicity "high", dar_contribution 1.2, risk_level moderate. In this version of the model the DAR term dominates and the payload term is a coarse class label, not a computed logP-driven delta. So covabio_adc will tell you that DAR 8 is where you start watching HIC. It will not currently rank an auristatin against a camptothecin derivative for you. If payload-to-payload discrimination is the decision, you need the small-molecule descriptors alongside it.

Conjugation site selection is a liability problem

The site you conjugate is also a site with its own chemistry. On the same trastuzumab heavy chain, covabio_developability returned a composite score of 29/100 ("poor") against 72/100 ("good") for the light chain, with 5 aggregation flags, 3 critical NG deamidation motifs, 4 exposed-Met oxidation sites and an unpaired cysteine. covabio_antibody (Kabat numbering) placed an NG motif in CDR-H2 at position 55 and a DG motif in CDR-H3 at position 102. Fc methionine oxidation was flagged at 255, 361 and 431.

Two practical consequences.

Formulation is coupled too. covabio_viscosity on the same heavy chain returned 18.97 cP at 150 mg/mL, pH 6 ("moderate"), driven by charge asymmetry 0.33 and a hydrophobic patch score of 6, with arginine or proline suggested as viscosity reducers. Adding a hydrophobic payload does not make that better. If you are already at moderate viscosity as a naked antibody, high-DAR hydrophobic conjugation is a formulation risk you should cost in early.

Linker stability: what you have to measure

Linker choice is the one part of this that sequence-level tools cannot settle. Practically:

• Cleavable (valine-citrulline, glucuronide, acid-labile hydrazone, disulfide): enables bystander killing, but premature cleavage in circulation is the failure mode. Hydrazones are the least stable of the common set.
• Non-cleavable (thioether, as in ado-trastuzumab emtansine): the released species is payload plus linker plus lysine, which is charged and poorly membrane-permeable. Better plasma stability, no bystander effect.
• Maleimide-thiosuccinimide adducts are subject to retro-Michael exchange with plasma albumin. Hydrolysing the succinimide ring, or using self-hydrolysing maleimides, is the standard mitigation.

You establish this with forced degradation and plasma-stability studies, tracking free payload by LC-MS and DAR drift over time. That is wet-lab work, and no in-silico triage substitutes for it.

Honest limits

• covabio_adc scores one sequence at a time. Intact IgG numbers require running both chains and summing.
• estimated_adc_mw excludes linker mass and assumes a discrete integer payload count. It is not a predicted deconvoluted mass.
• Conjugatable cysteine count uses an interchain_disulfide_estimate heuristic, not a structure-based solvent-accessibility calculation.
• The aggregation risk call is driven primarily by DAR, with a coarse payload hydrophobicity class. It does not discriminate between payload chemotypes at equal DAR, as our camptothecin control showed.
• Nothing here predicts DAR distribution width, HIC retention, plasma linker stability, potency, or ADCC/target-binding retention after conjugation. Those are measurements.
• All of it is triage to rank options and prioritise experiments, not a substitute for validated analytics or anything you would put in a filing.

Frequently asked questions

Related reading

• pLDDT explained: when to trust a predicted structure
• Forced degradation study design under ICH Q1A(R2)
• Mass balance in forced degradation studies

Run your own sequence and payload through covabio_adc on the free tier before you commit a conjugation campaign.