Chapter 5
How the immune system answers a therapeutic protein, how a protein ages in the vial, and how formulation ties the two together, read through WIA-BIO-016 §6–7, its Appendix A.3 and the SDK functions that implement them.
Every therapeutic protein is, at some level, a foreign or altered protein presented to an immune system that is very good at noticing such things. The response that matters for drug development is the formation of anti-drug antibodies (ADA). Chapter 4 treated a biologic as a molecule with a dissociation constant, a clearance and a volume of distribution. This chapter adds the complication that the patient can change those numbers by making antibodies against the drug.
ADA are not a single thing. WIA-BIO-016 §6.1.2 distinguishes two functional classes, and the distinction drives almost every downstream decision:
The most serious outcome is cross-reactivity with an endogenous protein. If a recombinant version of a natural protein provokes antibodies that also neutralise the body's own molecule, the patient can lose a function they had before treatment. The best-known example is the cluster of pure red cell aplasia cases reported with one epoetin alfa product in the early 2000s, which investigators linked to a formulation change and to the container closure rather than to the protein sequence. That case is the reason this chapter treats immunogenicity, stability and formulation together: a change that looks like a pharmaceutical detail can become an immunological event.
Immunogenicity is multifactorial, and it is useful to sort the factors into three groups, because each group is controlled by a different team:
| Group | Examples | Who controls it |
|---|---|---|
| Product-related | Non-human sequence, T-cell epitopes, aggregates and sub-visible particles, non-human glycans, host-cell protein impurities, oxidation and deamidation products | Discovery, process development, CMC analytics |
| Treatment-related | Route (subcutaneous exposure is often considered more immunogenic than intravenous), dose, dosing interval, treatment duration, intermittent vs continuous exposure, co-medication with immunosuppressants | Clinical development |
| Patient-related | Underlying disease and immune status, genetic background (for example HLA type), pre-existing antibodies, prior exposure to related products | Clinical development, pharmacovigilance |
A fully human sequence removes one product-related factor but not the others. Antibodies discovered from human libraries or transgenic animals still elicit ADA in some patients, partly because the variable regions are unique and partly because aggregates and formulation can break tolerance. No single parameter predicts immunogenicity, which is worth remembering when we meet a single score in section 5.3.
Regulators expect ADA to be measured with a validated, multi-tier bioanalytical strategy. The logic is described in the FDA guidance on developing and validating assays for anti-drug antibody detection and in the EMA guideline on immunogenicity assessment of therapeutic proteins; WIA-BIO-016 §6.1.1 records only the two numbers most people remember, the screening cut point formula and an assay sensitivity target. The full flow looks like this:
Patient serum sample
|
v
+--------------------+ below screening cut point +------------------+
| Tier 1: SCREEN | ----------------------------> | Reported ADA-neg |
| (e.g. bridging | +------------------+
| immunoassay) |
+--------------------+
| above cut point ("potentially positive")
v
+--------------------+ signal NOT inhibited by excess drug
| Tier 2: CONFIRM | ----------------------------> ADA-negative
| spike with excess |
| unlabelled drug |
+--------------------+
| signal inhibited beyond confirmatory cut point
v
+--------------------+
| Tier 3: TITER | serial dilution until signal < cut point
| (magnitude) | titer = highest dilution still positive
+--------------------+
|
v
+--------------------+
| Tier 4: NAb | cell-based or competitive ligand-binding assay
| (function) | answers: does the ADA block activity?
+--------------------+
|
v
Correlate with PK (trough, clearance), PD, efficacy, safetyThe screening cut point is set so that a known fraction of drug-naïve samples screen positive. The specification writes it as
Cut point = Mean_negative + 1.645 × SD_negativeThe constant 1.645 is the one-sided 95th percentile of the standard normal distribution. If the negative-population signals are approximately normal (often after a log transform and removal of outliers), about 5% of truly negative samples will exceed the cut point. That false-positive rate is deliberate: a screening tier is meant to over-call so that the confirmatory tier can sort true from false positives, and a low false-negative rate is worth more than a clean screening count. The confirmatory tier usually uses a stricter cut point based on the percentage inhibition caused by excess drug.
Three practical cautions follow. First, the negative population must resemble the study population: serum from the target disease can have a different baseline from healthy donors, so a cut point derived only from healthy volunteers can misclassify patients. Second, the normality assumption must be checked; if the distribution is skewed, a non-parametric percentile is the honest alternative. Third, the cut point is a property of the assay and lot of critical reagents, so it has to be re-verified when those change.
§6.1.1 also states an assay sensitivity target of better than 100 ng/mL, which matches the value commonly cited in FDA recommendations for screening and confirmatory assays. Sensitivity is measured with a surrogate positive control, usually an affinity-purified or monoclonal anti-drug antibody, so it is only an approximation of how the assay treats real, polyclonal patient ADA. The companion parameter is drug tolerance: circulating drug binds ADA and hides them from the assay. For a monoclonal antibody with a long half-life and high trough concentrations, drug tolerance often limits detection more than nominal sensitivity does, which is why sampling at trough, acid-dissociation pre-treatment and post-treatment washout samples are standard design choices.
WIA-BIO-016 stores clinical immunogenicity results in the ADAData interface of types.ts:
| Field | Type | Meaning and practical note |
|---|---|---|
incidenceRate | number, 0–1 | Fraction of evaluable subjects who are ADA-positive. Record the denominator separately; the type does not. |
titer | number, optional | Magnitude of the response; meaningful only together with the dilution scheme and the assay. |
neutralizing | boolean | Whether NAb were detected. A study-level boolean cannot express “NAb in some patients”; a per-subject record is needed for that. |
timeToADA | number (days), optional | Onset. The type does not distinguish transient from persistent ADA. |
clinicalImpact | 'none' | 'minimal' | 'moderate' | 'severe' | Sponsor's qualitative judgement of consequences for PK, efficacy and safety. |
The interface is a summary record, useful for a candidate dossier or a portfolio dashboard. It is not a bioanalytical data format: it has no sample identifiers, visit times, assay versions, cut-point values or treatment-emergent flags. A system that manages actual ADA testing should keep those raw results elsewhere and derive ADAData from them.
WIA-BIO-016 §6.2 introduces a single number meant to summarise immunogenicity risk for early triage:
IRS = (ADA_rate × Severity × Duration) / Tolerance
ADA_rate incidence of ADA, 0–1
Severity clinical impact, 1–10
Duration treatment duration factor (no scale defined)
Tolerance patient tolerance factor (no scale defined)The specification then bins the score: below 10 is low, 10–30 moderate, 30–60 high and above 60 very high. It is important to be clear about the status of this construct. The IRS is a WIA triage heuristic. It does not appear in ICH, FDA or EMA guidance, it has not been calibrated against clinical outcomes, and the action column of the specification's table (“contraindicated” above 60) should not be read as clinical guidance. What the score can do is force a team to write down, in a consistent format, how often ADA occur, how bad the consequences are, and how long patients will be exposed. That is a reasonable purpose for a data standard, as long as the number is not mistaken for a validated prediction.
ADA_rate = 0.4, Severity = 8, Duration = 6 and Tolerance = 0.5, the score is 38.4; lower the tolerance factor to 0.1 and it becomes 192.Appendix A.3 of the specification uses an ADA incidence of 0.15, severity 5, duration factor 2 for chronic use and tolerance factor 0.5. The arithmetic is (0.15 × 5 × 2) / 0.5 = 1.5 / 0.5 = 3.0, which falls in the low band. The SDK reproduces this exactly:
import { assessImmunogenicity } from './src/index';
const risk = assessImmunogenicity({
adaRate: 0.15, // 15 % of subjects ADA-positive
severity: 5, // 1–10 scale, "moderate"
durationFactor: 2, // sponsor convention: chronic use
toleranceFactor: 0.5, // sponsor convention: low tolerance
});
console.log(JSON.stringify(risk, null, 2));The function returns:
{
"riskScore": 3,
"riskLevel": "low",
"predictedADA": 0.1545,
"monitoringFrequency": "standard",
"mitigationStrategies": [],
"recommendations": [
"Monitor ADA levels standardly during treatment"
]
}
Three things in this output deserve comment. The value predictedADA = 0.1545 is computed as adaRate × (1 + riskScore / 100), capped at 1. Because the risk score is itself built from adaRate, this “prediction” is only a monotone rescaling of the input; it contains no information that was not already supplied, and it should not be shown to users as a forecast. Second, the recommendation text is assembled by appending “ly” to the monitoring frequency, which is why the output reads “standardly”; applications should map the enum to their own wording rather than display the string. Third, the mitigation list is driven by fixed thresholds (score above 20 adds sequence optimisation and T-cell epitope removal; above 40 adds an alternative format and an immunosuppression protocol; ADA rate above 0.3 adds combination therapy). These are generic prompts for discussion, not recommendations a tool should present to clinicians.
assessImmunogenicity() throws BIO007 (INVALID_PARAMETERS) when adaRate is outside 0–1 or severity is outside 1–10. It does not check the other two factors, and the consequences are easy to demonstrate:
| Input | SDK behaviour | What a conformant wrapper should do |
|---|---|---|
toleranceFactor: 0, adaRate > 0 | Score is Infinity, level very-high, predictedADA 1 | Reject with BIO007 |
toleranceFactor: 0, adaRate: 0 | Score is NaN; every comparison fails, so level falls through to very-high and predictedADA is NaN | Reject with BIO007 |
Negative durationFactor | Negative score classified as low | Reject with BIO007 |
| Score above 100 | Accepted silently | Accept, but flag as outside the documented range |
The shell tool behaves slightly differently. wia-bio-016 immunogenicity --ada-rate 0.15 --severity 5 uses the same defaults as Appendix A.3 (duration 2, tolerance 0.5) and prints IRS 3.00 and a predicted incidence of 15.4% (the SDK's 0.1545 is rounded by printf from a binary value just below 15.45). It does not perform the range checks at all, it does not emit the combination-therapy prompt for high ADA rates, and it labels the very-high band “Intensive + Consider alternatives”. With --tolerance 0 the underlying bc call fails with a divide-by-zero error, the empty result is then printed as “IRS 0.00”, and the following comparisons emit syntax errors, so a script that only reads the headline number would record a reassuring score for invalid input. When the CLI and SDK disagree, treat the TypeScript implementation as the reference and the CLI as a convenience front end.
ImmunogenicityRisk.riskLevel has four values, including 'very-high', while DrugCandidate.immunogenicityRisk, which feeds candidate validation in Chapter 6, accepts only 'low' | 'moderate' | 'high'. Code that pipes one into the other needs an explicit mapping, and the natural choice, collapsing very-high into high, loses information. Likewise, ADAData.clinicalImpact is a four-level qualitative scale that has no defined relationship to the numeric 1–10 severity used by the score. A future revision of the standard should either align these vocabularies or document the mapping; until then, implementers should record the mapping they chose alongside the data.
WIA-BIO-016 §6.3 names two sequence-level strategies. It also quotes percentage reductions for each; those figures are unsourced and are not repeated here, because the achievable effect depends heavily on the molecule and the patient population.
The historical path ran from murine antibodies to chimeric antibodies (murine variable domains on human constant regions), to humanized antibodies (murine complementarity-determining regions grafted onto human frameworks, often with a few framework back-mutations to restore affinity), to fully human antibodies from phage display or transgenic animals. Each step reduced foreign sequence. None of them made immunogenicity disappear, because the complementarity-determining regions of any antibody are new to the patient's immune system, and because the other risk factors in section 5.1 remain.
A durable, class-switched ADA response generally requires T-cell help, which in turn requires peptides from the drug to be presented on MHC class II molecules. Deimmunization identifies candidate T-cell epitopes, using in silico MHC-binding prediction followed by in vitro assays with human immune cells, and removes them with conservative substitutions. Each substitution must be checked against binding affinity (Chapter 4), stability and expression, because an epitope often sits in a structurally important region. The specification's shorthand, “remove T-cell epitopes, maintain binding affinity”, is the right summary of this trade-off.
The third lever is not in §6.3 but is at least as important: controlling the product-related factors measured in Chapter 2. Aggregates, particularly large or highly ordered ones, can present repeated epitopes that activate B cells more efficiently than a monomer does. Oxidised or deamidated variants, non-human glycan structures from certain expression systems, and host-cell protein impurities can all add risk. This is why the aggregate limits and SEC data from Chapter 2 are also immunogenicity controls, and why a stability programme that tracks aggregate growth over shelf life is, in part, an immunogenicity programme.
Stability work begins with deliberately damaging the molecule. Forced degradation studies expose the product to conditions harsher than it will meet in storage in order to answer two questions: which degradation pathways exist, and can the analytical methods see them? A method that can separate and quantify the degradation products is called stability-indicating, and demonstrating that property is the main deliverable of a forced degradation study.
WIA-BIO-016 §7.1 lists the classic stresses. The table below summarises them together with the main pathways and the methods that typically reveal them. Temperatures and concentrations shown are the specification's example conditions, not requirements.
| Stress | Spec example conditions | Main pathways | Methods that reveal them |
|---|---|---|---|
| Thermal | 40, 50, 60 °C; days to months | Unfolding, aggregation, fragmentation (hinge clipping), deamidation | SEC, CE-SDS, DSF/Tm, peptide mapping |
| pH | pH 3–9; 24–48 h | Deamidation (favoured at higher pH), isomerization of aspartate, acid-induced aggregation and fragmentation | IEX/cIEF, peptide mapping, SEC |
| Oxidation | Hydrogen peroxide, low concentration; hours | Methionine and tryptophan oxidation; in IgG, Fc methionines near the FcRn binding site are sensitive | Peptide mapping by LC-MS, FcRn binding assays |
| Light | Per ICH Q1B photostability design | Tryptophan and histidine photo-oxidation, cross-linking, colour change | Peptide mapping, SEC, appearance |
| Mechanical and freeze–thaw | Agitation, repeated freezing | Interface-induced aggregation, sub-visible particles | SEC, light obscuration, flow imaging |
The specification attaches acceptance criteria of “≥90% recovery” to thermal stress and “≥85% recovery” to pH stress. Those criteria are omitted here on purpose. Forced degradation is not a pass/fail test of the product: its aim is to produce measurable, interpretable degradation. A study in which nothing degrades has failed to learn anything, and a study that destroys the molecule completely has produced secondary products that are irrelevant to real storage. Stress levels are therefore tuned to generate informative amounts of change, and the result is a map of pathways, not a verdict.
Oxidation deserves one additional note because it links three chapters. Oxidation of the Fc methionines can weaken binding to FcRn, the receptor that recycles IgG and gives antibodies their long half-life (Chapter 4). A degradation pathway discovered here can therefore translate into a pharmacokinetic change, and oxidised material may also carry immunogenicity risk. When a forced-degradation study finds a sensitive residue, the finding should appear in three places in the dossier: the CQA list, the stability-indicating method description and the risk assessment.
Formal stability studies for biotechnological products follow ICH Q5C, supported by the general stability design in ICH Q1A(R2) and photostability in ICH Q1B. For a refrigerated product the long-term condition is 5 ± 3 °C, and ICH Q1A(R2) uses 25 °C with controlled humidity as the accelerated condition for such products. An important principle of ICH Q5C is that the shelf life of a biological product is based on real-time data at the intended storage condition; accelerated and stress data support the picture, reveal pathways and help evaluate excursions, but they do not by themselves justify the expiry date.
WIA-BIO-016 §7.2.1 summarises a typical design in three rows, which map directly to the StabilityConditions.type enum:
| Condition (spec §7.2.1) | Temperature | Duration in spec | Purpose | type value |
|---|---|---|---|---|
| Long-term | 2–8 °C | 24–36 months | Establish shelf life | 'long-term' |
| Accelerated | 25 °C | 6 months | Support, excursion assessment, trend prediction | 'accelerated' |
| Stress | 40 °C | 3 months | Degradation pathways | 'stress' |
| (not in table) | As labelled | Ongoing | Real-time confirmation on commercial lots | 'real-time' |
The enum's fourth value, 'real-time', overlaps with 'long-term': for a refrigerated product, long-term storage is real-time storage. A practical convention is to reserve 'long-term' for registration studies that support the initial shelf life and 'real-time' for ongoing post-approval monitoring of commercial lots. Whatever convention is chosen should be written down, because the type itself does not say.
A single test result at a single time point is represented by StabilityResult, which embeds its StabilityConditions. An illustrative 12-month result for a hypothetical IgG1 drug product looks like this (all values are examples):
{
"conditions": {
"temperature": 5,
"duration": 12,
"type": "long-term",
"lightExposed": false
},
"potency": 98.0,
"monomerContent": 98.9,
"aggregates": 0.9,
"pH": 6.0,
"appearance": "clear",
"passesSpec": true
}Four modelling details matter when a laboratory system produces these records:
duration is ambiguous. The comment says “duration in months”; in a result record the useful interpretation is the elapsed time point (12 months in the example), not the planned length of the study. Say so in your implementation notes.passesSpec is a derived value. Store the specification limits that were applied, or the boolean cannot be re-evaluated when the specification changes.Chemical rates increase with temperature, and the Arrhenius equation gives the classic description:
k = A × e^(−Ea / RT)
k degradation rate constant
A pre-exponential factor
Ea activation energy (J/mol)
R gas constant, 8.314 J/(mol·K)
T absolute temperature (K)
Taking the ratio of the equation at two temperatures removes A:
k2 = k1 × exp[ (Ea / R) × (1/T1 − 1/T2) ]If potency is lost by first-order kinetics, the fraction remaining after time t is e−kt, and the time to reach 90% of the initial value is t90 = −ln(0.90) / k = 0.1054 / k. That is the origin of the 0.105 in §7.2.2. The formula is exact only for first-order loss; aggregation is frequently higher-order or nucleation-limited, so the same shortcut can be misleading for aggregate growth.
BiopharmaSDK.predictShelfLife(k25C, activationEnergy, targetTemp = 5) implements the two-temperature form with R = 8.314 J/(mol·K) and a reference temperature of 25 °C, then converts t90 to months. Its unit assumptions are implicit and must be respected by the caller:
| Parameter / step | Unit assumed by the code | Consequence of getting it wrong |
|---|---|---|
k25C | per second (s−1) | The code divides t90 by 86,400 s/day. A rate in per day or per month gives a shelf life that is wrong by a factor of 86,400 or more. |
activationEnergy | J/mol | Passing 100 (kJ) instead of 100000 (J) makes the temperature correction almost vanish; the predicted shelf life at 5 °C becomes nearly the same as at 25 °C. |
targetTemp | °C | Converted internally with 273.15. |
| Month conversion | 30-day months | A small, systematic bias of about 1.4% against calendar-average months; acceptable if documented. |
| Gas constant | 8.314 J/(mol·K), hard-coded | BIOPHARMA_CONSTANTS.GAS_CONSTANT is 1.987, in cal/(mol·K). Mixing that constant with an activation energy in J/mol inflates the exponent by a factor of about 4.18. |
Suppose, for illustration, that potency loss at 25 °C has been fitted to a first-order rate constant of 2.0 × 10−9 s−1, and that the apparent activation energy is 100 kJ/mol. Both numbers are invented for this example.
import { BiopharmaSDK } from './src/index';
const sdk = new BiopharmaSDK();
const k25 = 2.0e-9; // s^-1, fitted at 25 °C (illustrative)
const Ea = 100_000; // J/mol (illustrative)
console.log(sdk.predictShelfLife(k25, Ea, 25).toFixed(1)); // 20.3
console.log(sdk.predictShelfLife(k25, Ea, 5).toFixed(1)); // 368.4
console.log(sdk.predictShelfLife(k25, 100, 5).toFixed(1)); // 20.3 <- Ea passed in kJ by mistakeBy hand: at 25 °C, t90 = 0.105 / 2.0 × 10−9 = 5.25 × 107 s, about 608 days, or 20.3 thirty-day months. For 5 °C, Ea/R = 12,028 K and 1/298.15 − 1/278.15 = −2.41 × 10−4 K−1, so the rate falls by a factor of e−2.90 ≈ 0.055, to 1.10 × 10−10 s−1. The predicted t90 becomes about 368 months, roughly thirty years.
That answer is arithmetically correct and practically useless, which is the most important lesson of this section. Protein degradation is rarely governed by one reaction with one activation energy across a 20 °C interval. Partial unfolding, a change in the rate-limiting pathway, the glass transition of a frozen or lyophilised matrix, and interfacial processes that do not scale with temperature at all can each dominate at different temperatures. An Arrhenius fit is valuable for ranking formulations, estimating the impact of a temperature excursion, or checking whether real-time trends are consistent with accelerated ones. It does not replace real-time data, and ICH Q5C does not allow it to. If an application displays the output of predictShelfLife(), it should label the value as a model estimate and show the real-time data beside it; the optional StabilityResult.predictedShelfLife field should be populated only with that understanding.
Formulation is the set of excipients, pH, concentration and container that keep the protein stable and deliverable. It is also, as the epoetin example in section 5.1 showed, part of the immunogenicity picture. WIA-BIO-016 §7.3 gives a short list of buffers and stabilisers; this section explains the reasoning behind them and corrects one rule.
A buffer works best within about one pH unit of its pKa. The three buffers named in §7.3.1 follow from that rule: histidine, whose imidazole group has a pKa near 6, is the workhorse for monoclonal antibodies formulated around pH 5.5–6.5; citrate, with several pKa values spanning the acidic range, covers lower pH; phosphate, with its second pKa near 7.2, covers neutral pH, although it shows large pH shifts on freezing because its salts crystallise at different rates.
The specification's rule “Optimal pH = pI ± 1” points in the wrong direction. At the isoelectric point a protein carries no net charge, so electrostatic repulsion between molecules is weakest and solubility is typically at its minimum; colloidal instability and precipitation are most likely there. Formulators therefore usually choose a pH at least one to two units away from the pI, within the range where the chemical degradation pathways from section 5.5 are slowest. For many IgG1 antibodies, whose pI is often in the basic range, this leads to a slightly acidic formulation.
| Class | Examples | Mechanism | Cautions |
|---|---|---|---|
| Sugars and polyols | Sucrose, trehalose, sorbitol | Preferential exclusion stabilises the folded state in solution; in freeze-dried products they replace water and form a glassy matrix | Contribute to osmolality; reducing sugars can glycate proteins, which is why non-reducing sugars are preferred |
| Amino acids | Arginine, glycine, histidine, proline | Suppress protein–protein interactions, reduce viscosity of concentrated solutions, act as bulking agents in lyophilisates | Effects are molecule-specific; arginine can reduce aggregation in one product and not another |
| Non-ionic surfactants | Polysorbate 80, polysorbate 20, poloxamer 188 | Occupy air–liquid and solid–liquid interfaces so the protein does not unfold there during shaking, pumping or filling | Polysorbates degrade by hydrolysis and oxidation; degradation products can form particles and peroxides can oxidise the protein |
| Salts and tonicity agents | Sodium chloride | Adjust tonicity and screen charges | Can reduce colloidal stability for some molecules |
§7.3.2 quotes concentration ranges for these excipients (sugars 5–10%, amino acids 50–200 mM, polysorbate 80 0.01–0.1%). Treat them as a sketch of the design space rather than targets: the right amount is found experimentally, usually with a design-of-experiments screen that reads out the stability-indicating methods from section 5.5.
A growing share of antibody products is given subcutaneously, which limits the injected volume and pushes protein concentration upward. At high concentration, viscosity and self-association become the dominant problems, and they are governed by the same weak intermolecular interactions that formulation pH and excipients control. Subcutaneous administration is also one of the treatment-related immunogenicity factors listed in section 5.1. A formulation choice made for patient convenience therefore needs to be re-examined against both the stability data and the immunogenicity plan, which is precisely the integration this chapter has been arguing for.
Immunogenicity, stability and formulation are three views of one question: what does the patient actually receive, and how does their body respond to it? The WIA-BIO-016 artefacts that carry the answer can be connected as follows:
Chapter 2 CQAs (aggregates, oxidation, glycans, HCP)
| |
v v
Forced degradation (§7.1) Formulation (§7.3)
-> stability-indicating methods -> pH away from pI, excipients
| |
+--------------+---------------+
v
Stability programme (ICH Q5C)
StabilityConditions / StabilityResult
predictShelfLife() -> model estimate only
|
v
Clinical lots -> tiered ADA assay -> ADAData
|
v
assessImmunogenicity() -> IRS (WIA triage heuristic)
|
v
DrugCandidate.immunogenicityRisk (Chapter 6 gating;
map 'very-high' explicitly)Three rules summarise how to use the standard responsibly in this area. First, report inputs alongside every derived number, whether that number is an IRS or a predicted shelf life. Second, validate the parameters the SDK leaves unchecked, especially the tolerance factor and the units of the Arrhenius inputs, and record the conventions chosen for the undefined scales. Third, never let a model output stand in for the measurement it approximates: real-time stability data set the shelf life, and a validated ADA assay, not a score, establishes whether patients are making antibodies.
Chapter 6 takes these outputs into the development protocol, where stability scores and immunogenicity classes become gates for candidate selection, and where the same analytical attributes are compared against a reference product to establish biosimilarity.