# Protein aggregation: causes, prediction and prevention

> Protein aggregation is the self-association of protein molecules into amorphous clumps, soluble oligomers or ordered amyloid fibrils. Learn why proteins aggregate, how aggregates are measured, what TANGO, AGGRESCAN, CamSol, Aggrescan3D, Protein-Sol and NetSolP predict from sequence and structure, and how formulation, mutations and expression conditions prevent it, with a worked Aggrescan3D run on transthyretin.

Transthyretin is a tetramer, and the first step toward transthyretin amyloid is that tetramer coming apart. In transthyretin amyloid cardiomyopathy, wild-type or variant protein becomes unstable, misfolds and deposits as fibrils in the heart. In a phase 3 trial of 441 patients, tafamidis, which binds transthyretin and prevents the tetramer from dissociating, lowered all-cause mortality over 30 months (29.5% against 42.9% on placebo).

The same behavior costs time at the bench. A designed binder that clouds the tube after purification, an enzyme that ends up in inclusion bodies, an antibody that gets viscous and opalescent at high concentration: all of these are aggregation, and much of it can be checked before the gene is ordered. Run through [Aggrescan3D](/app/aggrescan3d), the transthyretin structure turns out to light up in the very segment that fibril studies had already singled out, which the worked example below shows.

## What counts as an aggregate

"Aggregate" is a loose word. The FDA's immunogenicity guidance defines it as any self-associated protein species and sorts aggregates by size, reversibility, conformation, chemical modification and morphology, from dimers up to visible particles hundreds of micrometers across. A dimer that dissociates on dilution and a fibril that does not are both aggregates, and they behave very differently.

At one end are amorphous aggregates, disordered clumps with no repeating internal structure. At the other are amyloid fibrils, which are ordered. X-ray fibre diffraction of six fibrils from patient tissue and two synthetic fibrils gave similar patterns, consistent with beta sheets running along the fibril axis with their strands perpendicular to it. This is the cross-beta structure, and it appeared even though the precursor proteins share neither sequence nor native fold.

Between the two sit soluble oligomers, small assemblies that stay in solution and are easy to miss. They matter because they can be harmful: species formed early in the aggregation of two proteins unrelated to any disease were found to be highly cytotoxic. Inclusion bodies are the version biotechnologists meet most often, the insoluble deposits that recombinant protein forms inside bacteria when folding loses the race against aggregation.

## Why proteins aggregate

Aggregation usually starts from chains that are partly or fully unfolded. Chiti and colleagues concluded that aggregation generally begins in molecules that are partly or completely unfolded. Anything that loosens the fold, such as heat, extreme pH, a destabilizing mutation or a denaturant, raises the odds. [Protein denaturation](/guides/protein-denaturation) covers how folds come apart; the point here is what happens next.

Once chains are loose, the culprits are aggregation-prone regions. TANGO models aggregation as intermolecular beta-sheet formation and assumes that the core of an aggregate is fully buried, so segments that can form beta sheets and bury their side chains are the ones it flags. In a folded protein those stretches are normally hidden in the core.

The Aggrescan3D authors explain why structure matters: in folded globular proteins, aggregation-prone sites are often scattered along the sequence or hidden inside the native structure, which sequence-only methods struggle to predict.

Mutations change the picture in predictable ways. Across many peptides and proteins, the effect of a mutation on aggregation rate correlated to a remarkable extent with changes in hydrophobicity, secondary-structure propensity and charge. A few residues can also dominate the kinetics. In human beta-2 microglobulin at pH 2.5, a single region of about 10 residues set the rate of fibril formation, although other regions had significant intrinsic propensity too.

Folded proteins can aggregate as well, and this is where conformational and colloidal stability part ways. Conformational stability is resistance to unfolding. Colloidal stability is the tendency of intact molecules to stay apart, set by their surface charge and by sticky patches. Sugars and polyols mainly improve the first, while arginine is described as acting mostly on the second. A protein can be thermally stable and still clump at high concentration.

The conditions around the protein push in both directions. Therapeutic antibodies tend to aggregate when stored at the high concentrations their use requires.

Solubility usually reaches its minimum near the isoelectric point, where molecules carry no net charge to repel each other; in ribonuclease Sa variants engineered to different pI values, the pH of minimum solubility moved with the pI ([isoelectric point guide](/guides/protein-isoelectric-point)).

Interfaces and agitation also generate aggregates: in an IgG1 antibody, shaking and stirring produced different aggregate species, vial headspace had a large effect, and polysorbate was protective at some concentrations but increased aggregation at others.

## Why aggregation is a problem

In disease, the best-studied cases are the amyloid proteins. Alzheimer disease is linked to deposition of amyloid beta peptide in plaques, and the amyloid hypothesis proposes that its accumulation is the primary driver of the disease process. A mutation in the alpha-synuclein gene was identified in families with Parkinson disease in 1997, and the same year alpha-synuclein was reported in Lewy bodies. Transthyretin, as above, deposits when the tetramer becomes unstable.

For biologics, the concern is safety and shelf life. Aggregates from dimers to visible particles have been recognized for decades for their potential to provoke immune responses to protein drugs, and the FDA lists several mechanisms, including cross-linking of B-cell receptors and enhanced antigen uptake.

There is evidence that higher-molecular-weight aggregates and particles are more potent than small ones, though the guidance also states that key information is still missing on which aggregates, and how much of them, trigger a response for a given product.

Animal data illustrate the effect. In mice, aggregation alone raised the IgM and IgG2a responses to a single-chain antibody fragment, and a contaminating bacterial chaperone, DnaK, boosted the IgG and IgG2a responses further but only when aggregated protein was present. A review of the field concludes that aggregation can enhance immunogenicity while the mechanisms remain poorly defined.

The FDA asks manufacturers to minimize aggregation, choose formulations and container closures that limit it during storage, and let expiration dating reflect any increase in aggregates.

It also asks sponsors to assess subvisible particles of 2 to 10 micrometers over the shelf life and to characterize smaller ones (0.1 to 2 micrometers) as methods allow.

The pharmacopeial limit sits at the larger end: USP chapter 787 allows at most 6,000 particles of 10 micrometers or larger and 600 of 25 micrometers or larger per container for small-volume protein injections, counted primarily by light obscuration.

## How aggregation is measured

No single assay sees every aggregate. A review of analytical methods notes that none covers the entire size range or type of aggregate, that each has limits of detection, and that sample preparation can create or destroy aggregates, so results from orthogonal methods should be compared. The FDA gives a concrete example: size-exclusion chromatography alone may miss high-molecular-weight aggregates that never get through the column prefilter, yet these may matter most for immunogenicity.

  <caption><strong>Table 1. Common ways to detect protein aggregates and what each one misses.</strong> Dynamic light scattering is described from its general principle; the other limitations come from the cited sources. Sources: FDA (2014), Biancalana and Koide (2010), USP 787 summary, Kiese et al. (2008).</caption>
  <thead>
    <tr>
      <th scope="col">Method</th>
      <th scope="col">What it reports</th>
      <th scope="col">What it can miss</th>
    </tr>
  </thead>
  <tbody>
    <tr><td>Size-exclusion chromatography (SEC)</td><td>Monomer, dimer and larger soluble species separated by size</td><td>Large aggregates that do not pass the column prefilter</td></tr>
    <tr><td>Dynamic light scattering (DLS)</td><td>Size distribution of particles in solution</td><td>Scattering is dominated by the largest particles, so small species of low abundance are hard to resolve</td></tr>
    <tr><td>Thioflavin T fluorescence</td><td>Fluorescence rises when the dye binds amyloid fibrils</td><td>Amorphous aggregates, which lack the cross-beta structure the dye recognizes</td></tr>
    <tr><td>Light obscuration</td><td>Counts and sizes of subvisible particles, with USP limits at 10 and 25 micrometers</td><td>Samples that are cloudy or viscous, for which microscopy is the fallback method</td></tr>
    <tr><td>Visual and turbidity checks</td><td>Visible particles and cloudiness</td><td>Anything smaller than the eye or the instrument can resolve</td></tr>
  </tbody>

Thioflavin T is the standard dye for amyloid, because its fluorescence increases strongly when it binds fibrils. It is therefore the wrong tool for amorphous clumps, and a negative ThT result does not mean a sample is free of aggregates. The antibody stress study combined visible inspection, turbidity, subvisible particle analysis, SEC and DLS, and concluded that an array of methods was needed to cover the aggregate morphologies it produced.

## Predicting aggregation from sequence and structure

Prediction methods answer three different questions that are easy to confuse. Some locate aggregation-prone regions within one protein, some rank variants of the same protein by solubility, and some estimate whether a whole sequence will be soluble when expressed. Table 2 sorts the most used methods by what each paper reports.

  <caption><strong>Table 2. Aggregation and solubility predictors and what their papers report.</strong> SCM is the spatial charge map and SAP the spatial aggregation propensity. Sources: Fernandez-Escamilla et al. (2004), Conchillo-Solé et al. (2007), Tartaglia et al. (2008), Sormanni et al. (2015), Chennamsetty et al. (2009), Zambrano et al. (2015), Kuriata et al. (2019), Hebditch et al. (2017), Thumuluri et al. (2022), Kalejaye et al. (2024).</caption>
  <thead>
    <tr>
      <th scope="col">Method</th>
      <th scope="col">Input</th>
      <th scope="col">What the paper reports</th>
    </tr>
  </thead>
  <tbody>
    <tr><td>TANGO</td><td>Sequence</td><td>Statistical-mechanics model of beta-sheet aggregation with a buried aggregate core; predicted 179 literature peptides and 71 peptides from disease proteins, and the effect of mutations in the Alzheimer beta peptide, human lysozyme and transthyretin</td></tr>
    <tr><td>AGGRESCAN</td><td>Sequence</td><td>Aggregation-propensity scale for the amino acids derived from in vivo experiments, assuming short specific stretches drive aggregation; finds hot spots and compares mutants</td></tr>
    <tr><td>Zyggregator</td><td>Sequence</td><td>Estimates folding and aggregation propensity together from sequence, extending predictions to molecules that are folded or partly structured</td></tr>
    <tr><td>CamSol</td><td>Sequence</td><td>Screens tens of thousands of mutations for the largest solubility gain while preserving the native state and activity; tested on a single-domain antibody against amyloid beta</td></tr>
    <tr><td>SAP</td><td>Structure from molecular dynamics</td><td>Finds the location and size of aggregation-prone regions in full antibody simulations; mutants of those regions were more stable than wild type</td></tr>
    <tr><td>Aggrescan3D</td><td>Structure</td><td>Applies the AGGRESCAN scale in three dimensions with solvent exposure; a dynamic mode adds structural fluctuations, and version 2.0 adds multimers and predicted solubility and stability changes on mutation</td></tr>
    <tr><td>Protein-Sol</td><td>Sequence</td><td>Weights 35 sequence-based properties from an E. coli cell-free expression dataset and reports which features deviate most from average</td></tr>
    <tr><td>NetSolP</td><td>Sequence</td><td>Transformer language models predict solubility and usability for purification in E. coli, trained on datasets curated with strict sequence-identity partitioning</td></tr>
    <tr><td>DeepSP</td><td>Antibody variable sequences</td><td>Neural network trained on 20,530 antibody sequences predicts 30 SAP and SCM descriptors, correlating 0.76 to 0.96 (mean 0.87) with values from molecular dynamics</td></tr>
  </tbody>

Two cautions follow from the table. First, the solubility predictors learned from specific data. Protein-Sol was built on E. coli proteins made in a cell-free system, and NetSolP predicts solubility on expression in E. coli. That is a different question from whether a purified antibody stays monomeric at 100 mg/mL, so a good expression score does not clear a candidate for storage. The NetSolP authors flag a related problem: existing solubility datasets are biased, which is why they curated their own.

Second, a score is not a probability. Aggrescan3D scores residues in their structural environment and is best used to compare structures run with the same settings; positive values mark aggregation-prone environments and negative values favor solubility. They are not percentages. Structure-based methods also inherit the quality of the structure: for a predicted model, check the per-residue confidence first ([pLDDT guide](/guides/plddt)), because misplaced side chains give misplaced scores.

### Worked example: transthyretin in Aggrescan3D

To see what a structure-based run flags, we loaded PDB entry 1F41, wild-type human transthyretin, into [Aggrescan3D](/app/aggrescan3d) with default settings (10 Å radius, native pH scale). The file holds two chains, 231 residues in total, numbered 10 to 125 in the mature protein. The average score is −0.67, so most residues sit in solubility-favoring environments. Only 26 of 231 residues score above zero (14 in chain A and 12 in chain B), and just four exceed 1.0.

  <caption><strong>Table 3. Highest-scoring residues in the Aggrescan3D run on transthyretin (1F41).</strong> Scores are Aggrescan3D residue scores for chains A and B with default settings. "In 105 to 115" marks residues inside the segment that forms amyloid fibrils as an isolated peptide. Sources: Aggrescan3D run on PDB 1F41, Jaroniec et al. (2002).</caption>
  <thead>
    <tr>
      <th scope="col">Residue</th>
      <th scope="col">Chain A</th>
      <th scope="col">Chain B</th>
      <th scope="col">In 105 to 115</th>
    </tr>
  </thead>
  <tbody>
    <tr><td>Val 121</td><td>1.68</td><td>1.25</td><td>No</td></tr>
    <tr><td>Leu 110</td><td>1.17</td><td>1.15</td><td>Yes</td></tr>
    <tr><td>Val 20</td><td>0.91</td><td>0.63</td><td>No</td></tr>
    <tr><td>Tyr 114</td><td>0.91</td><td>0.84</td><td>Yes</td></tr>
    <tr><td>Leu 17</td><td>0.82</td><td>0.80</td><td>No</td></tr>
    <tr><td>Ser 112</td><td>0.66</td><td>0.62</td><td>Yes</td></tr>
    <tr><td>Thr 119</td><td>0.61</td><td>0.56</td><td>No</td></tr>
    <tr><td>Ser 115</td><td>0.55</td><td>0.52</td><td>Yes</td></tr>
  </tbody>

The two chains agree closely, and the flagged residues cluster. In chain A, 11 of the 14 positive residues lie between positions 105 and 123, and seven of them fall within residues 105 to 115 (Thr 106, Ala 108, Leu 110, Ser 112, Pro 113, Tyr 114, Ser 115). That segment is known from experiment: the isolated peptide TTR(105-115) forms amyloid fibrils in vitro, adopting an extended beta-strand conformation. Aggrescan3D was never told this.

Three limits keep the result in proportion. The file contains the two chains of the crystal's asymmetric unit, while the biological assembly is a tetramer, so surface that is buried at the tetramer interface may be scored as exposed. The run is static, so it says nothing about the dissociation step that precedes amyloid formation in the disease. And the output shows where the sticky segments sit, not whether the protein will aggregate under a particular buffer, temperature and concentration.

## Preventing aggregation

### In the buffer

Start with pH. Because solubility usually bottoms out near the pI, formulation and storage buffers are normally set away from it; compute the value with the [isoelectric point calculator](/app/isoelectric-point-calculator) or the [pI Calculator](/app/pi-calculator) and read [the pI guide](/guides/protein-isoelectric-point) for the caveats, since cysteines, tags and blocked termini shift real pI values.

Excipients act on different failure modes. Sugars such as sucrose stabilize the folded state.

Arginine suppresses protein-protein interaction and later-stage aggregation, was the most effective of the tested additives at reducing opalescence in concentrated IgG solutions, and is widely used to lower viscosity, but results are not uniform: arginine hydrochloride decreased colloidal stability in two antibodies and increased aggregation of one with a strongly positive Fv charge at pH 5.5. The review's practical advice is to pair a sugar or polyol with arginine rather than rely on either alone.

Polysorbates 20 and 80 are added to prevent surface adsorption and to stabilize against aggregation. Concentration matters, as the stress study showed: polysorbate protected the antibody at some levels and promoted aggregation at others, and headspace and the type of mechanical stress changed the outcome.

### In the sequence

When the buffer cannot fix it, change the protein. CamSol was built for this: it screens thousands of point mutations and picks the ones that raise solubility while keeping the native state and activity. In a library of 17 variants of a humanized antibody designed to span a range of predicted solubility, predicted and measured solubility correlated strongly. Aggrescan3D 2.0 likewise predicts solubility and stability changes on mutation, so a candidate substitution can be checked for both effects.

Charge is the other lever. In the ribonuclease Sa work, replacing acidic surface residues with lysines raised the pI from 3.5 to 6.4 and 10.2 and shifted the pH of minimum solubility with it, without moving the pH of maximum activity or stability.

For designed proteins, sequence design itself can help. [SolubleMPNN](/app/soluble-mpnn) samples sequences for a fixed backbone with a checkpoint trained only on structures of soluble proteins. It is not a solubility predictor, and its output does not guarantee soluble expression, so score its designs with the tools above; see the [protein sequence design](/use-cases/protein-sequence-design) use case for the full workflow.

### In the expression host

For recombinant protein in E. coli, the simplest lever is temperature. Lowering the incubation temperature is the most common way to slow synthesis, with 15 to 25 °C the typical range for limiting inclusion bodies, at the cost of lower yield and less efficient chaperones. Co-expressing chaperone sets such as GroEL/GroES or DnaK/DnaJ/GrpE helps nascent chains fold, and high expression can overwhelm the host's own quality control.

Fusion partners are the third option. When maltose-binding protein, glutathione S-transferase and thioredoxin were compared on six proteins that normally accumulate insoluble, MBP was far more effective than the other two. Large tags can also distort solubility assessments, so the free protein still needs testing after cleavage.

## Using aggregation predictors to pick candidates

A practical order for a set of designed binders or variants starts with the cheapest screens. Run sequence predictors first: [NetSolP](/app/netsolp-1-0) for solubility and usability in E. coli and [Protein-Sol](/app/protein-sol) for a solubility estimate with the features that deviate from average. Compare pI and net charge at your buffer pH with the [protein charge plot](/app/charge-plot), and look for long hydrophobic stretches with the [hydrophobicity plot](/app/hydrophobicity-plot).

Then fold the survivors with [ESMFold](/app/esmfold), keep those with confident models, and score them with Aggrescan3D. Rank on the average score and on the exposed hot spots, not on a single threshold, and compare only runs with the same settings.

Whatever passes still needs SEC and a particle check, because most of these predictors were developed on natural proteins or antibodies, and designed sequences can fall outside that experience. The [protein binder design](/use-cases/protein-binder-design) use case places these checks in the wider selection workflow, and [protein language models](/guides/protein-language-models) explains what models like those behind NetSolP learn from sequence.

Antibodies have their own tools. [DeepSP](/app/deepsp) predicts the SAP and SCM descriptors from the variable-domain sequences, [DeepViscosity](/app/deepviscosity) classifies high-concentration viscosity, and [TAP2](/app/tap2) profiles developability with a structure model; the [antibody engineering guide](/guides/what-is-antibody-engineering) explains where these checks fit.

## Frequently asked questions

### Is protein aggregation reversible?

Sometimes. The FDA classifies aggregates partly by reversibility and dissociation, which means that some come apart again, for example on dilution, and others do not. Reversibility has to be measured for each sample.

### Is amyloid the same as aggregation?

No. Amyloid is one ordered kind of aggregate, defined by the cross-beta structure of its fibrils, while amorphous aggregates and soluble oligomers have no such order.

### Does a high solubility score mean my protein will not aggregate?

No. Tools like Protein-Sol and NetSolP predict solubility on expression in E. coli, which is a separate question from colloidal stability at high concentration in a formulation buffer.

### Which aggregation predictor should I use?

Use sequence tools for a first pass over many candidates and a structure-based tool such as Aggrescan3D to see where the problem residues sit. For antibodies, add the antibody-specific descriptors. Experiments remain the final test.