Characterization/Property Prediction

​

Tamarind Scoring Tool

Every score described below can be accessed as part of the Protein Scoring tool, which lets you score a (potentially large) set of protein sequences for any number of properties. You can also use these tools in Pipelines, creating stages with filters between them to successively score for more computationally intensive properties.

​

Antibody Developability Properties

​

General Developability

Therapeutic Antibody Profiler (TAP): TAP compares structural metrics of an antibody to known therapeutic antibodies. We’ve reproduced its methodology to calculate Patches of Surface Hydrophobicity (PSH), Patches of positive charge (PPC), Patches of negative charge (PNC), Structural Fv charge symmetry, and CDR length. Antibodies outside the typical range for therapeutic antibodies are flagged. TAP is a good starting point for general screening because it covers multiple risk factors at once.

​

Aggregation

Aggregates reduce efficacy, trigger immunogenic reactions, and complicate manufacturing. AggreScan3D takes a structure (modeled or experimental) and computes per-residue aggregation scores, identifying surface-exposed hotspots. Addressing predicted hotspots — substituting problematic residues or inserting glycosylation sites near aggregation-prone regions — can improve solubility and stability without harming binding. A practical workflow: run a fast sequence-based method (e.g. TANGO, Aggrescan) on many candidates to filter out the worst, then apply AggreScan3D on top leads.

​

Solubility

An antibody with poor solubility tends to precipitate or aggregate, especially at high concentration. NetSolP provides fast, sequence-based solubility prediction. That said, solubility should usually be optimized after primary metrics like binding affinity. If you have experimental solubility data, our ESM finetuning tool can learn what drives solubility in your specific context. Use solubility scores to remove candidates below a threshold or in the bottom percentile of your library. If an antibody scores low, consider conservative mutations outside the binding interface (e.g. substituting a surface-exposed leucine with threonine in a non-critical framework position).

​

Viscosity and Self-Interaction

High viscosity is a problem for therapeutic antibodies at high concentrations (100+ mg/mL) for subcutaneous injection. A major contributor is large electrostatic patches on the antibody surface that drive transient self-association. DeepSP predicts spatial properties like Spatial Charge Map (SCM) and Spatial Aggregation Propensity (SAP) directly from sequence, outputting ~30 spatial descriptors for variable regions. This enables high-throughput viscosity-risk screening without needing a 3D model. TAP’s surface charge asymmetry guideline also catches antibodies likely to have formulation issues.

​

Stability and Chemical Liabilities

ThermoMPNN predicts stabilizing point mutations given a 3D structure by estimating ΔΔG upon mutation. DeepStabP predicts melting point from sequence. For nanobodies, TEMPRO and NanoMelt predict VHH melting temperatures from sequence. Antibodies are also subject to chemical liabilities — sequence motifs that degrade over time:

• Asn deamidation: Asn-Gly (NG) motifs are well-known hotspots, especially in CDR loops.
• Asp isomerization: DG motifs risk isoAsp formation.
• Met/Trp oxidation: Exposed methionine residues, especially in CDRs, are flagged.
• Unpaired cysteines: Extra cysteines beyond expected disulfide pairs cause mispairing or aggregation.
• Unintended glycosylation: N-X-[S/T] motifs in variable domains create glycosylation sites leading to product heterogeneity.

The Tamarind Antibody Annotator screens for PTMs and chemical liabilities, marking prone residues and indicating whether they’re solvent-exposed or buried. Antibodies with multiple liabilities are usually filtered out; manageable ones can be engineered away (e.g. Asn→Gln to prevent deamidation, Met→Leu to avoid oxidation).

​

Immunogenicity

We offer DeepImmuno for Class I and TLimmuno for Class II immunogenicity. These tools scan every possible n-mer peptide and assess its response with every HLA allele, reporting the maximum and average score per peptide and the allele where it peaked. CDR loops can contain sequences not common in the human germline that could be immunogenic. Antibodies with multiple strong predicted epitopes in CDRs are candidates for deimmunization. Use a broad set of HLA alleles to ensure population coverage. For research or diagnostic antibodies used only in vitro, this filter can be skipped.

​

Isoelectric Point and Charge Profile

Most therapeutic mAbs have pIs between about 6.1 and 9.4. Extreme pI signals risk — high pI (above 9) leads to nonspecific binding and faster clearance; low pI (below 6) can cause purification issues. The Tamarind Protein Properties tool calculates pI from sequence. Beyond pI, charge distribution matters. TAP’s Structural Fv Charge Symmetry Parameter (SFvCSP) quantifies heavy vs. light chain charge imbalance, and DeepSP identifies charge patches from sequence. pI is typically used as a secondary filter — prefer candidates with moderate pI when binding properties are similar.

​

Manufacturability and Expression

Highly aggregation-prone or unstable antibodies often manifest as low expression yield. By filtering out antibodies with poor biophysical properties using the tools above, you increase the chances remaining candidates express well. Additional liabilities to watch: free cysteines (ER retention), heavy-light chain pairing mismatches, and atypical N-terminal sequences. For novel formats (bispecifics, fragments, scFvs), the same principles apply but format-specific evaluation may be needed.

​

General Protein Characterization

​

Solubility

NetSolP also works for non-antibody proteins, providing fast sequence-based solubility prediction.

​

Stability

ThermoMPNN finds stabilizing point mutations for any protein with a 3D structure. DeepStabP predicts melting point from sequence.

​

Aggregation

AggreScan3D works for any protein structure, not just antibodies.

​

Enzyme Activity

We offer DLKcat and CatPred to predict activity of a protein given a protein sequence and SMILES substrate. You can also try one of our protein/small molecule docking tools to use binding affinity as a proxy for activity.

​

Integrating Predictions into Your Pipeline

Multi-parameter screening: Use a panel of predictors covering all key properties rather than a single score. Each tool catches different issues. Stage-wise filtering: Apply fast predictions early, reserve intensive analysis for later:

1. Sequence-level filters on large pools: Run liability scans and basic rules. Eliminate clones with egregious liabilities (NG motifs, free cysteine, extreme hydrophobicity/charge).
2. Refined profiling on leads: Run deeper analysis on top 20–50 leads — structure modeling, AggreScan3D, DeepSP, TAP, immunogenicity prediction.
3. Experiment and iterate: Run experimental assays on remaining leads. Compare with predictions to validate and refine.

Benchmark against known therapeutics: Use databases like Thera-SAbDab. If your candidate falls well inside the range of known successful antibodies, that’s a good sign. Design for developability: Incorporate predictions during engineering, not just after selection. If a mutation increases affinity but introduces an aggregation hotspot, explore alternatives.

​

Key Developability Prediction Tools