ADC toxicity is not random. A decade of clinical experience across approved ADCs shows that the major toxicity domains track with design choices, and payload class is the strongest single predictor:

MMAE (vedotin ADCs): peripheral neuropathy and hematologic toxicity, driven by free payload from cleavable linkers.
MMAF and DM4: ocular toxicity, with corneal findings that require ophthalmologic monitoring.
DXd and related topoisomerase I payloads: interstitial lung disease and pneumonitis, the boxed-warning risk that reshaped monitoring practice after trastuzumab deruxtecan.
Calicheamicin: hepatotoxicity including veno-occlusive disease.

DAR and dosing schedule then modulate the base rate: higher DAR raises free payload burden, and repeat dosing accumulates exposure in slow-clearing tissues like the cornea. None of this is controversial. The problem is that this knowledge lives in scattered labels, review articles, and the heads of experienced tox scientists, and it rarely gets applied systematically when a new design is on the table.

A worked example: first-in-class payload, known design levers

Consider a program with a novel auristatin-class, membrane-permeable payload against a new solid-tumor antigen, DAR 3-4, repeat dosing. First-in-class means there is no FDA precedent for the exact payload: no validated MTD, no established DLTs, no tissue-expression data in any curated source.

Run this design through the ADC Consultant tox module in the Pylot demo and you get an honest structure for exactly that situation:

Inherited risk profile: UNKNOWN, stated as such. The engine does not pretend a novel payload has a known profile. It labels the evidence strength as indirect and precedent-based, and says so on screen.
Design-driven risk signals, quantified. Independent of payload precedent, the design itself carries signals. In this example, repeat dosing contributes the largest single risk-score increment, which matches clinical experience that cumulative exposure drives several ADC toxicity domains.
Reference ADCs with similar features. The engine surfaces approved ADCs that share design features with yours, along with the management strategies that worked: for ocular risk, the precedent pattern is to continue at the same dose while increasing ophthalmologic monitoring frequency, rather than reflexive dose reduction.
Domain-by-domain monitoring implications. ILD, ocular, neuropathy, and hematologic domains each get their own precedent-based assessment, so the output maps directly onto the monitoring sections of an IND-enabling plan.

Why precedent-based beats intuition-based

The alternative to systematic precedent analysis is asking whoever in the room has seen the most ADC programs. That works until it does not: people over-index on their own programs, class effects get missed when the payload name is new, and the reasoning is not documented anywhere a regulator or partner can review.

A precedent engine makes the same kind of inference an experienced tox scientist makes, but it does it exhaustively, it shows the reference set it used, and it writes down its assumptions. Every risk conclusion in Pylot traces to the design inputs and the curated precedent catalog, and the built-in AI guide will explain any score in those terms. It is constrained to the knowledge base and the calculations on screen; it does not invent findings.

What to do with the output

For a discovery or early-development team, the practical deliverables are:

A domain-specific monitoring plan draft (ILD, ocular, neuropathy, hematologic) grounded in class precedent.
Dose interruption and reduction criteria per domain, aligned with what regulators have accepted for comparable designs.
A documented, reviewable rationale for why the program's tox strategy looks the way it does.