Antibody developability — flag classes and ROC-AUC against measured outcomes

Abstract

Developability is the unromantic but expensive part of biologics discovery. A binder with great affinity is useless if it aggregates in the formulation tank, cannot be expressed at scale, or generates an immune response. We characterise Aether Bio ^[1] developability-flag performance on a curated set of 412 late-stage antibody candidates with measured downstream outcomes, across five flag classes (aggregation propensity, viscosity at formulation concentration, glycosylation hot-spots, deamidation/oxidation liabilities, chemical-modification hot-spots). Aggregate ROC-AUC reaches 0.83; per-class numbers span 0.71 (viscosity) to 0.89 (aggregation propensity). We describe the calibration interface that drops customer-specific false-positive rates by 15–30 percentage points, why viscosity remains the weakest endpoint, and how the flags integrate into the broader discovery loop.

1. Why developability matters

Developability is the second-most expensive failure mode in biologics drug discovery. The most expensive is clinical attrition; the second is committing engineering and CMC resources to a candidate that ultimately cannot be manufactured at therapeutic concentrations, cannot be stored stably, or fails immunogenicity screening. The classes of developability liability are well-characterised in the literature — aggregation, viscosity, glycosylation, post-translational modifications, chemical liabilities — and the question is how reliably they can be predicted before committing to expression and characterisation.

The economic stakes are concrete. A typical late-stage antibody candidate consumes single-digit-to-low-double-digit millions of dollars in expression, formulation development, and stability characterisation before the team is confident in its developability profile. A flag that reliably identifies the worst candidates early — before this spend — pays for itself many times over. Conversely, a flag that produces too many false positives slows the team down by triggering re-engineering passes on candidates that would have been fine, which is its own failure mode.

2. Setup

We curated 412 antibody candidates from late-stage development programmes — drawn from public patent filings, clinical-trial disclosures, and previously-published developability datasets — with downstream developability outcomes documented in the literature. Each candidate has at least one measured developability endpoint and a documented programmatic outcome (advanced, redesigned, terminated for developability reasons).

We split developability into five flag classes.

• Aggregation propensity — propensity to form non-functional or immunogenic aggregates at formulation concentration.
• Viscosity at formulation concentration — dynamic viscosity at therapeutic dose (typically 100–200 mg/mL).
• Glycosylation hot-spots — predicted glycosylation sites that affect activity or pharmacokinetics.
• Deamidation / oxidation liabilities — Asn deamidation, Met oxidation, Trp oxidation hot-spots.
• Chemical-modification hot-spots — Asp isomerisation, succinimide formation, lysine modification.

Each flag is evaluated on its own ROC-AUC against the corresponding measured outcome (binary: developability failure vs no failure on that flag). Aggregate ROC-AUC is the macro-average across the five flags. We hold out a 20% test set from candidates not represented in any public training corpus we are aware of, to control for distribution-shift effects ^[3].

3. Headline numbers

Aggregate ROC-AUC across the five-flag panel is 0.83 on the held-out external set. Per-flag numbers span 0.71 (viscosity at formulation concentration) to 0.89 (aggregation propensity). The pattern is consistent across model sizes — the 280B sparse-MoE configuration improves aggregate ROC-AUC to 0.86 but the per-flag ordering remains the same.

4. Per-flag breakdown

• Aggregation propensity — 0.89 ROC-AUC. Strong, particularly on the high-aggregation tail. The model learns from sequence + structure + computed surface-property features; the worst-aggregating candidates are reliably identified.
• Glycosylation hot-spots — 0.87 ROC-AUC. Mostly a sequence-pattern problem with structural context; the well-characterised consensus motifs are easy, and the model captures the context-dependent cases that determine whether a site is occupied in expression.
• Deamidation / oxidation — 0.82 ROC-AUC. Local sequence + solvent accessibility is the right feature set; the model uses both. Some context-sensitive Met oxidation cases (where buried Met has unexpected access) remain harder than the average.
• Chemical liabilities — 0.79 ROC-AUC. Asp isomerisation is the harder sub-class because context-dependence is broader; succinimide formation is more reliably predicted. Performance on this flag is meaningfully better when the customer provides their formulation buffer for calibration.
• Viscosity at formulation — 0.71 ROC-AUC. The weakest endpoint. Solvent and pH context matter more than the model currently uses; see Section 5.

5. Why viscosity is harder

Viscosity at therapeutic-formulation concentrations (typically 100–200 mg/mL) depends on net-charge distribution, secondary-structure stability, and solvent conditions in ways that are well-understood individually but interact in domain-specific ways. The model currently under-uses solvent and pH context — it tends to predict from sequence and structure alone, which gives a respectable but not great ROC-AUC.

Concretely, the weakest cases are antibodies with comparable sequence-level properties that nonetheless exhibit very different viscosities at the same concentration because of subtle differences in how their charge distributions interact at high concentration. The model's structural representation captures the static charge layout but not the concentration-dependent collective behaviour that dominates the rheology.

We are training a small auxiliary head on solvent + pH context. Preview results are encouraging — ROC-AUC rises to roughly 0.79 on the same test set — but we are not yet at parity with the other flags and the engineering ahead is non-trivial. Customers running viscosity-sensitive programmes today should expect to confirm borderline predictions with bench measurement ^[4].

6. Calibration to customer bench history

Each customer has a slightly different definition of "failed developability" — driven by their downstream processing, their formulation constraints, and the willingness of their development team to engineer around early flags. A candidate that one customer would terminate, another might re-engineer; the same model-output probability needs to map to different threshold decisions across customers.

The model ships with a calibration interface ^[3] that fits thresholds and weights against the customer's bench history. The customer provides their measured-outcome dataset (typically 50–200 candidates with bench-confirmed developability outcomes); the runtime fits per-flag thresholds plus an aggregate-decision rule that reflects the customer's risk posture.

Most customers see their false-positive rate drop by 15–30 percentage points after calibration. The exact gain depends on how different their bench history is from the public developability literature; customers with substantial in-house measured data see the larger end of the range.

7. How this fits the loop

Developability flagging happens inside the same model that does generative design and FEP ^[2]. A flagged candidate triggers a re-design pass against the offending feature; a borderline flag triggers an experimental confirmation slot in the autonomous-lab planning queue ^[4]. The information flows in both directions automatically.

Concretely: the generative-chemistry agent has access to developability-flag predictions during candidate generation. A candidate generated against a binding-affinity objective is filtered in real time against the developability flags; candidates that fail the configured threshold are re-generated with the offending feature avoided. This is meaningfully cheaper than the conventional workflow of generating freely and triaging downstream.

When developability flags are uncertain, the active-learner ^[5] preferentially picks the candidate for experimental confirmation. The flag prediction enters as a prior on the assay outcome; the measured assay outcome updates the prior and refines future predictions. The integration is per-customer and uses only the customer's data; the calibration does not flow across tenants.

8. Limitations

Three meaningful limitations. First, viscosity at formulation concentration remains weaker than the other flags and we have flagged the dependency on solvent and pH context as the gap to close. Second, the external evaluation set is drawn from public sources and is biased toward late-stage candidates that survived earlier filters; performance on earlier-stage candidates with broader liability profiles may differ. Third, we cover the five most commonly-cited developability classes but the field uses additional finer-grained classifications (hydrophobicity, polyspecificity, self-association beyond aggregation) that we treat as derived flags rather than as first-class outputs.

Reproducibility: the public portion of the curated 412-candidate set is open under the Research tier. The customer-calibrated configurations are per-tenant and not shareable.

9. Conclusion

Developability flags in Aether Bio reach aggregate ROC-AUC 0.83 on a curated external set of 412 antibody candidates, with per-flag performance spanning 0.71 (viscosity) to 0.89 (aggregation). Customer-specific calibration drops false-positive rates by 15–30 percentage points; the predictions flow back into the generative-design and active-learning loops automatically ^{[2, 5]}. Viscosity at formulation concentration remains the open problem; the rest of the panel is at production-grade accuracy for biologics-discovery customers running candidate triage at scale.