Custom service

AI-assisted Protein Engineering

Structure prediction and machine learning, applied where they actually help.

Design methods: AlphaFold2, inverse folding models, stability predictors, co-evolution analysisDesign scale: 10–50 variants per cycle, 2–3 cycles typicalAssays: Expression, thermal stability, activity, binding — per goal

AI in protein engineering is useful in exactly one way: it reduces the number of variants you have to make and test. Structure prediction turns a sequence into a hypothesis about which residues matter; inverse folding proposes sequences for a desired backbone; machine learning models rank mutations by predicted effect. None of it removes the experiment — it shrinks the search, which is where the value lies.

The workflow starts with structure. AlphaFold2 predictions, supported by experimental structures where they exist, identify active sites, interfaces, instability hotspots and glycosylation sequons. From that map, the engineering question is posed concretely: which mutations are plausible, at which positions, in which combinations. Design tools — inverse folding, model-predicted stability, co-evolution signals — generate a ranked candidate list instead of a combinatorial explosion.

The candidates are then made and measured, because prediction is a hypothesis, not a result. Small-scale expression and activity assays score the designs, the data feed back into the model, and the next design cycle starts from a better-trained position. Two or three cycles typically deliver what a library screening campaign delivers in one, at a fraction of the variants.

AI-assisted engineering fits well-defined single-objective problems: stabilising an enzyme by twenty degrees, improving a binding interface, solubilising an aggregation-prone domain.

It underperforms on poorly defined objectives and on targets with no structure and no homologues — a weak prediction produces weak designs. It is also not a substitute for directed evolution when the objective is one that no model can rank, such as a novel catalytic activity.

Each cycle's report pairs predictions with measurements, variant by variant, including the misses. Final engineered variants are delivered with full expression, stability and activity data, sequences and constructs.

Service Highlights

  • Structure-first hypothesis. AlphaFold2 and experimental structures used to locate the residues that matter before any variant is designed.
  • Ranked design lists. Inverse folding, stability models and co-evolution analysis produce a shortlist, typically 10–50 variants rather than thousands.
  • Model feedback loop. Experimental data from each cycle retrain the ranking, so designs improve rather than restart.
  • Multi-objective optimisation. Stability, activity and solubility weighted together, since maximising one at the expense of the others is a common failure.
  • Wet-lab validation included. Every design cycle ends in measured expression and activity, not in a report of predictions.
  • Transparent reporting. Predicted effects are reported alongside measured outcomes, including where the model was wrong.

How the work runs

  1. Problem definition

    The property to engineer — stability, affinity, activity, solubility — and its acceptance criteria are agreed before any prediction is run.

  2. Structure acquisition

    AlphaFold2 prediction, experimental structure, or homologous structures assembled into a working model.

  3. Computational design

    Positions and mutations proposed and ranked by the relevant models; the candidate list is sized to the wet-lab capacity.

  4. Construct and expression

    Designed variants synthesised, cloned and expressed in parallel at small scale.

  5. Measurement

    Activity, stability and solubility measured on every variant, with the assay chosen against the engineering goal.

  6. Iteration

    Data fed back into the model; a second and third cycle refine the design until the acceptance criteria are met.

Specifications & timeline

Design methods
AlphaFold2, inverse folding models, stability predictors, co-evolution analysis
Design scale
10–50 variants per cycle, 2–3 cycles typical
Assays
Expression, thermal stability, activity, binding — per goal
Timeline
2–6 weeks per cycle

Related Products

Next step

Tell us the target, we will tell you what is feasible

Send the sequence, assay format and sensitivity you need. You will get a technical assessment — recommended platform, timeline and the risks worth knowing before you commit budget.