Ask about our free Pilot2Partnership program
Learn more
Case study · AI protein design
Results at a glance
50/50
variants showed soluble misfolded protein before purification
17/50
contained only misfolded protein, giving a clean negative call
33/50
were mixtures; CLīM isolated the folded, active fraction in one step
95%
of the His6/Ni-NTA eluate on one target was misfolded, inactive protein
The challenge
Standard tags can’t tell folded from misfolded.
Standard affinity tags such as His6 and Twin-Strep are short peptides. They capture soluble protein whether it is folded and active or misfolded and inactive. In this screen, Ni-NTA resin purified soluble misfolded protein right alongside the folded protein, diluting it with an inactive population and skewing activity, binding and stability data. For a team training models on that data, every mixed sample is a mislabeled data point.
What we did
One purification step, 50 variants.
We expressed 50 point-mutation variants of an inhibitor-binding enzyme in E. coli and purified each in a single step with CLīM affinity resin and on-column SUMO protease cleavage.
The CL7 tag must be folded to bind the Im7 resin
Ultra-tight CL7/Im7 affinity (published KD ~10⁻¹⁴ M)
Stringent washes up to 3.0 M NaCl remove background
Folded protein elutes tag-free; misfolded protein flows through
Results
Definitive calls instead of ambiguous readings.
17 clean negatives
These variants made only misfolded protein and did not bind the CLīM resin: a definitive negative for functional expression rather than an ambiguous low-activity reading.
33 folded fractions isolated
In the mixed variants, folded protein bound and eluted tag-free while the misfolded species flowed through.
5% vs. recovered
On one target compared directly, standard His6/Ni-NTA purification yielded just 5% active protein and needed a second MBP step. Single-step CLīM recovered the active enzyme.
Why it matters
Better data in, better models out.
1
Less systematic bias
Unstable variants misfold most, so co-purified inactive protein skews exactly the designs your model learns from.
2
Less assay noise
Activity, affinity and stability are measured on the true folded population, not a mixture.
3
Compounding advantage
Each design cycle trains on cleaner values, widening the gap over teams using noisy data.
Results are from a 50-variant screen of a single enzyme performed for a client. The His6/Ni-NTA versus CLīM comparison was run on one representative target and is directional, not a formal head-to-head across all 50 variants. The KD value is the published figure for the CL7/Im7 interaction (Vassylyeva et al., PNAS 2017).