Structural validation for protein binder design
Every designed binder candidate is validated by structure prediction before advancing to synthesis. Boltz-2, ESMFold, and ColabFold filter candidates that fail to fold correctly or engage the target as intended.
Three models, layered validation
ESMFold
Single-sequence structure prediction using a protein language model. No multiple sequence alignment required, making it fast enough to screen thousands of designed sequences as an initial fold quality check.
Seconds per sequence. Screens the full design pool rapidly.
First-pass filter. Removes sequences that are predicted to misfold before running more expensive complex prediction.
Boltz-2
Structure prediction model for validating designed binder-target complexes. Predicts the three-dimensional structure of the binder in complex with the target and scores the predicted interface quality.
ipTM, interface pLDDT, and PAE at the binding interface. All three must pass thresholds.
Primary complex validator. Confirms that the designed binder engages the target at the intended epitope with correct geometry.
ColabFold (AlphaFold2)
Accelerated AlphaFold2 for complex structure prediction. Provides an independent validation check using a different model architecture, reducing the risk of model-specific false positives from Boltz-2 alone.
Same confidence framework: pLDDT, PAE, ipTM. Cross-referenced against Boltz-2 predictions.
Orthogonal validator. Candidates that pass both Boltz-2 and ColabFold have higher confidence of correct fold and interface geometry.
How we interpret predictions
Per-residue confidence score (0-100). Regions above 70 are generally reliable. Below 50 indicates disorder or low-confidence prediction.
Predicted aligned error between residue pairs. Low PAE between binder and target indicates high confidence in their relative positioning at the interface.
Interface predicted TM-score. Measures the quality of the predicted complex interface. The primary metric for assessing whether a designed binder engages the target correctly.
When your target lacks a crystal structure
Many client targets do not have experimental crystal structures in the PDB. AlphaFold2 models serve as the structural input for design campaigns, provided the binding region has high confidence (pLDDT >70, low inter-domain PAE).
We assess model quality before accepting it as a design input. Low-confidence regions at the proposed binding site are flagged, and we recommend experimental structure determination if the model is unreliable at the epitope.
Have a target sequence but no crystal structure?
We accept AlphaFold models as design inputs with appropriate confidence assessment. Send us your target and we will evaluate feasibility.
Start a project →