AI in SEC and IEX for biologic characterisation addresses a different problem from the process-optimisation applications elsewhere in downstream bioprocessing: not how to purify a molecule, but how to characterise it once purified. Size-exclusion chromatography and ion-exchange chromatography are the workhorse analytical techniques for assessing the size and charge heterogeneity of monoclonal antibodies and increasingly complex bispecifics, and machine learning is beginning to automate the interpretation of their data, particularly in high-throughput comparability and lot-release contexts where the volume of chromatograms strains manual review.
This article covers where AI genuinely helps and where it does not. It sits within Separation Science's guide to AI in process chromatography, and complements the process-side coverage of AI-assisted Protein A loading and elution, where the same molecules are purified before they reach the characterisation bench.
Key Takeaways
|
SEC for Biologics: What the Data Shows
Size-exclusion chromatography separates molecules by hydrodynamic size, and in biologic characterisation it is the preferred analytical technique for quantifying aggregates and fragments in a therapeutic protein product. Aggregation is a critical quality attribute because aggregates can affect both efficacy and immunogenicity, so accurate quantification of even low-abundance high-molecular-weight species matters directly for product safety and regulatory acceptance.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
What SEC data reveals, and where interpretation gets hard:
- High-molecular-weight species. Aggregates elute earlier than the monomer, and their accurate quantification, often at well below one percent, is a core release and stability measurement. Small, early-eluting peaks near the detection limit are where integration consistency matters most.
- Monomer purity. The main peak quantification is usually straightforward, but its accuracy depends on consistent baseline and boundary decisions relative to adjacent aggregate and fragment peaks.
- Fragments and low-molecular-weight species. Clipped or degraded species elute later and require the same integration care, particularly when partially co-eluting with the monomer.
- The interpretation variability problem. Where peaks are small or partially resolved, manual integration decisions vary between analysts, introducing a source of variability that consistent automated assignment can reduce.
This is the specific place AI adds value in SEC: not in the easy, well-resolved cases where manual integration is already reliable, but in the consistency of the difficult, low-abundance, partially resolved decisions where analyst-to-analyst variability is real and consequential for a release or stability result.
IEX Charge Variant Analysis: The Interpretation Challenge
Ion-exchange chromatography, and cation-exchange in particular, is the workhorse for charge variant analysis, which characterises the heterogeneity of a therapeutic protein's charge-related forms. As one detailed study of the technique notes, charge variant analysis depicts variant forms that may differ by only minor modifications of a single amino acid, and the profiles often display partially resolved peaks on the shoulders of larger peaks, which puts considerable pressure on method robustness and on the consistency of peak assignment.
Why charge variant interpretation is genuinely difficult:
- Subtle chemical differences. Acidic and basic variants arise from modifications such as deamidation, C-terminal lysine variation, and glycation, which shift charge slightly and produce closely spaced, sometimes overlapping, peaks.
- Partially resolved shoulders. Many charge variant peaks appear as shoulders on a dominant main peak rather than as baseline-resolved features, making integration boundaries a matter of judgement that varies between analysts and runs.
- Method sensitivity. Charge variant separations are sensitive to mobile phase pH and gradient, so the profile itself shifts with method conditions, complicating consistent assignment across a method's lifetime and between laboratories.
- Regulatory weight. Charge variant distribution is a monitored critical quality attribute, so the consistency and defensibility of its measurement carry direct regulatory significance.
Charge variants can differ by a single amino acid modification and often appear as shoulders on a larger peak. That combination, subtle chemistry and partial resolution, is exactly where consistent, automated interpretation earns its place.
AI-Assisted Peak Assignment in SEC and IEX
Given that both techniques produce partially resolved peaks whose integration varies between analysts, the clearest AI application is consistent, automated peak detection and assignment. This is the same peak-picking problem that appears across chromatography, applied to the specific patterns of biologic size and charge profiles.
What AI-assisted assignment contributes:
- Consistent integration. A trained model applies the same integration logic to every chromatogram, removing the analyst-to-analyst and day-to-day variability that manual integration introduces, particularly for shoulder peaks and low-abundance species.
- High-throughput processing. In comparability and stability studies generating hundreds of chromatograms, automated assignment processes the volume in a fraction of the time manual review requires, which is where the throughput case is strongest.
- Pattern recognition across a profile. Models can learn the characteristic pattern of a given molecule's charge variant or size profile, flagging deviations from the expected pattern that may indicate a real product change rather than an analytical artefact.
- Automated method optimisation. Beyond interpretation, algorithmic approaches to optimising charge variant separations systematically tune pH and gradient to improve resolution of the critical variants, reducing the manual method development burden the technique traditionally carries.
The consistent limitation is that automated assignment is a starting point, not a final answer, for a result that carries regulatory weight. Where an assignment is uncertain, particularly for a novel variant or an unexpected peak, orthogonal confirmation is needed, and the pairing of these separations with mass spectrometry, in workflows that couple SEC and ion-exchange directly to MS, is increasingly how that confirmation is obtained at characterisation scale.
Automated Aggregation Monitoring With AI
Aggregation monitoring deserves specific attention because it is one of the most consequential characterisation measurements and one where automation delivers clear, practical value across development, release, and stability.
Where automated aggregation monitoring helps:
- Stability study throughput. Stability programs generate large numbers of SEC chromatograms across timepoints and conditions, and consistent automated integration of the aggregate peak across that entire dataset removes a significant manual burden and a source of variability.
- Low-abundance quantification consistency. Because aggregate specifications are often at very low percentages, consistent integration of small early-eluting peaks directly affects whether a result meets specification, making assignment consistency a quality-relevant matter, not just an efficiency one.
- Trend detection. Models tracking aggregate levels across a stability study or a manufacturing campaign can flag an emerging upward trend earlier and more consistently than periodic manual review, supporting proactive investigation.
- Linking to the process. Because aggregation can be introduced during purification, particularly at low-pH elution steps, consistent aggregate characterisation connects the analytical result back to the Protein A elution conditions that may have influenced it, closing the loop between process and product.
As with charge variants, the value is concentrated in consistency and throughput rather than in replacing analytical judgement. A model that integrates the aggregate peak the same way every time is valuable precisely because it makes the measurement more reproducible, which is what a stability or release decision depends on.
AI for Biosimilar Comparability Analysis
Comparability and biosimilarity assessments are among the most data-intensive applications of SEC and IEX, and therefore among the most natural fits for AI-assisted interpretation. Demonstrating that a biosimilar matches its reference product, or that a process change has not altered a product, requires extensive analytical characterisation, and automated multi-attribute monitoring workflows that assess size, charge, and other attributes across many samples are increasingly used to generate the necessary comparative data efficiently.
How AI supports comparability work:
- Consistent cross-sample comparison. Comparability depends on comparing profiles between products or process versions, and consistent automated assignment across all samples removes the analyst variability that could otherwise obscure or falsely suggest a difference.
- High-volume characterisation. A biosimilarity dossier requires characterising many batches across multiple attributes, and automated interpretation makes that volume of SEC and IEX analysis practical within realistic timelines.
- Pattern-level similarity assessment. Beyond individual peak quantification, models can assess whether two profiles are similar at the pattern level, supporting the holistic comparison that comparability fundamentally requires.
The caution specific to comparability is that the stakes make validation non-negotiable. A comparability conclusion supports a regulatory filing, so any AI-assisted interpretation contributing to it must be validated for that use, with the consistency of its assignments demonstrated rather than assumed.
Regulatory Expectations for AI-Assisted Characterisation Data
Biologic characterisation data supports regulatory filings, release decisions, and comparability conclusions, which places AI-assisted interpretation squarely within the regulatory framework for analytical data. The specifications and acceptance criteria for biotechnological products under ICH Q6B define what must be characterised and controlled, and AI-assisted methods that generate that characterisation data are held to the same standards as any other analytical approach.
The regulatory considerations that apply:
- Method validation. An AI-assisted peak assignment or aggregation quantification method must be validated for its intended use, with its accuracy, precision, and consistency demonstrated against the attributes it measures.
- Interpretability and defensibility. A characterisation result may need to be explained to a regulatory reviewer, so an assignment approach that can be examined and justified is more defensible than an opaque one, the same interpretability preference that applies across regulated analytical AI.
- Data integrity. The record of what was measured, how it was integrated, and any human review must meet data integrity expectations to the ALCOA+ standard, including for automated integration decisions.
- Orthogonal confirmation. For consequential or uncertain assignments, confirmation by an orthogonal method, typically mass spectrometry, remains expected, and AI-assisted optical peak assignment does not remove that expectation.
These obligations connect biologic characterisation to the wider analytical QC and compliance domain, which is the subject of its own dedicated coverage in this series. The through-line is consistent: AI improves the consistency and throughput of characterisation, but the validation, interpretability, and data integrity requirements that govern regulated analytical data apply undiminished.
What This Means for Your LabFor biologic characterisation, the near-term value of AI is in consistency and throughput, not in replacing analytical judgement. Target the difficult, high-volume interpretation problems first: low-abundance aggregate integration in stability studies, partially resolved charge variant shoulders, and the large sample sets that comparability work generates, since these are where manual variability is highest and automation pays off most. Treat automated assignment as a consistent first pass that still requires review for consequential results, keep orthogonal MS confirmation in the workflow for uncertain or novel peaks, and validate any AI-assisted method that contributes to a release or comparability decision. Approached this way, AI makes characterisation more reproducible and scalable without compromising the rigour that regulated biologic data demands. For the process side of the same molecules, the AI in process chromatography guide and the AI in analytical science overview provide the surrounding context. |
This article was produced under Separation Science’s AI Editorial Guidelines
