AI-Guided TCR Engineering: How Etcembly Unlocked a Clinical-Stage TCR Candidate

The drug development industry has a lead optimization problem. Traditional T cell receptor (TCR) engineering campaigns require scientists to synthesise and test hundreds of variants in the lab, spending months and significant budget before discovering whether any candidates are worth pursuing. Hit rates are low, timelines are long, and the cost of each wet lab experiment adds up fast, often before a viable lead is even identified.

EMLy workflows are purpose-built to change that. Not by replacing scientific expertise, but by enhancing it and helping focus it more precisely. By using EMLy AI to do the heavy computational lifting before a single experiment is run, you can save time and money and avoid using wet-lab resources on candidates without promise.

In this blog post, we share the story of how this worked in practice, enabling the optimization of a TCR therapeutic in just weeks, rather than months.

The Challenge: Quickly turning weak-binding parental TCRs into therapeutically viable candidates

In a partnership project with Zelluna Immunotherapy, the goal was clear but technically demanding. Zelluna had a lead MAGEA4 (Melanoma-Associated Antigen A4) TCR with binding affinities in the micromolar range. Although this reflects a typical affinity for natural TCRs, it is far below the potency required for a viable cell therapy. This TCR needed to be  engineered to a level that would improve therapeutic viability and enable further clinical development.

The conventional approach would have involved broad exploratory mutagenesis: making and testing large numbers of variants in the lab, iterating slowly, spending heavily on protein production and testing at each round. It is a well-established process, but it is neither fast nor cheap.

We took a different approach entirely.

How EMLy works: structure-guided design from sequence alone

Early programmes like the Zelluna collaboration were instrumental in shaping how Etcembly approached this structural modelling challenge. EMLy's starting point is the TCR sequence itself. From that sequence alone, the platform generates 3D structural models of each TCR docked to its peptide-HLA complex, producing hundreds of candidate poses and identifying the conformations most likely to represent productive binding.

The insights gained from real-world engineering projects like this ultimately led to the development of DoRIAT, Etcembly's proprietary Bayesian framework for interpreting and annotating docking runs. General-purpose tools often generate structural predictions that appear plausible but fail to capture the specific geometry of TCR-pHLA engagement. DoRIAT is specifically designed to truly understand this interaction. 

By optimising for the crossing angle, CDR loop positioning, and interface dynamics that determine whether a TCR will actually bind its target at therapeutically relevant affinity, it can identify candidates that have real-world potential.The result is a ranked shortlist of the most promising mutation candidates, assessed entirely in silico before any lab work begins.

From this computational foundation, EMLy runs a recursive cycle of improvements: expert-guided mutagenesis, thousands of in silico variant assessments, structural perturbation analysis, and candidate selection, all before synthesis. The best models are selected based on optimal TCR contact geometry with the peptide and HLA complex, and only the most promising variants proceed to the lab.

The result: 40 variants, 20-fold affinity improvement, and a candidate in the clinic

For the Zelluna programme, EMLy predicted and shortlisted TCR variants within three weeks. In total, 40 variants were designed and sent to the lab for testing.

The results validated the approach:

• Over 20-fold increase in binding affinity compared to parental, with the best variants reaching KD values in the low micromolar range, which is the ideal target affinity for cell therapy

• 9 shortlisted hits were selected by Zelluna based on potency, specificity, expression levels, number and location of mutations, and consistency across experiments

• No detectable off-target cross-reactivity to a panel of mimetic peptides, achieving an essential safety requirement for any clinical candidate

  
  

Superior functional activity in T cell assays with Zelluna's IFNγ ELISpot data showing Etcembly's engineered clones outperforming parental clone.

Note: these TCRs are examples of optimised TCRs, but none is the actual clinical TCR

The lead molecule progressed through Zelluna's internal safety, potency, and specificity assessments and is now in the clinic for its first-in-human trial.

Why EMLy’s TCR-specific AI outperforms general-purpose AI tools

The broader AI drug discovery landscape is increasingly crowded, and it is worth being specific about what makes EMLy different from general-purpose protein modelling tools.

Models like AlphaFold 2, Boltz-1 (AlphaFold 3), and Chai-2 are remarkable achievements in structural biology. But they were not built specifically for TCR-pHLA co-complex modelling and that specificity matters enormously when the goal is to predict whether an engineered variant will bind its target at therapeutically relevant affinity.

When we compared EMLy's predicted TCR-pHLA crossing angles against experimentally determined crystal structures, EMLy consistently produced predictions closer to the real structure than AlphaFold 2 or Boltz-1:

The crossing angle, which reflects the geometry of how the TCR sits in the peptide-HLA complex, is a critical determinant of binding specificity and potency. Getting it wrong means engineering for a binding pose that does not exist in reality. EMLy gets it right, consistently.

When assessing structural accuracy benchmarks across multiple TCR-pHLA complexes, EMLy produced lower RMSD values and higher DockQ scores than both Boltz-2 and Chai-2, and did so with an 8-million-parameter model running inference 15 times faster than billion-parameter alternatives.

EMLy’s sequence-derived model (magenta) aligned with crystal structure (cyan) with RMSD of <2.5Å (Fig. A), and highly aligned CDR loop positions (Fig. B)

This is not a general-purpose model retrofitted for biologics. It is a purpose-built platform, validated against clinical outcomes.

Beyond TCRs: Applying EMLy to across antibody developability programmes

The Zelluna programme is not an isolated result. Etcembly has applied the same EMLy-driven approach across antibody engineering programmes with consistent outcomes.

In separate developability optimisation programmes run in partnerships with Vector Laboratories and Twist Bioscience, EMLy was tasked with improving the expression and yield of several hard-to-express antibodies, where poor expression and aggregation were blocking progression.

For Vector Laboratories, more than 80% of EMLy's predictions improved expression, yield, and aggregation issues over the parental antibody. As a result, they were able to achieve faster progression to developable antibody leads.

For Twist Biosciences, the programme achieved a 96% success rate for yield improvement, enabling faster progression to developable antibody leads while avoiding more than £50,000 in anticipated wet-lab costs.

The pattern is consistent: a small number of precisely selected variants, tested in the lab, with a high success rate. The economics of drug discovery look different when the hit rate is this high.

What this means for your biotech or pharma pipeline

TCR and antibody engineering have historically been defined by the scale of the experimental search: more variants equals more experiments, which means more time. EMLy inverts that logic. The search happens in silico at incredibly high computational speeds, before a single experimental run. The lab resources are used more cost-effectively to confirm predictions rather than to discover them.

For a pharma or biotech team with a challenging molecule - whether it's a weak binder, a hard-to-express antibody, or a bispecific that needs optimising for selectivity without sacrificing potency - this is a fundamentally different kind of collaboration.

We do not offer a black-box AI service. Every programme is built around Etcembly's validated structural and computational capabilities, delivered by a team with a proven track record of taking engineered molecules into the clinic.

If you have a TCR or antibody engineering challenge, we would like to hear about it.

Get in touch at hello@etcembly.io - or book a consultation to learn more about how EMLy can accelerate your pipeline.