AI-guided TCR affinity maturation for biologics engineering

Therapeutic candidates can only advance as far as their engineering allows. For every molecule that reaches the clinic, hundreds of others are synthesised, tested, and discarded. The problem is rarely that the target was wrong, but rather that the tools used to optimise antibody and TCR affinity, specificity, functional potency, and developability required for a clinical candidate are inefficient.

Conventional biologics engineering approaches - whether for antibodies or T cell receptors (TCRs) - demand a broad and expensive experimental search: make many variants, test them all in the lab, iterate slowly, and hope that something worth pursuing emerges before the budget runs out.

ETC-101, Etcembly's lead PRAME-targeting T cell engager (ETCer), reflects a new approach to TCR-based biologics engineering and affinity maturation.

Its development highlights what becomes possible when AI-guided affinity maturation is placed at the front of the discovery process, not as a supplement to traditional methods, but as the engine driving candidate selection and optimisation.

PRAME and the case for a high-affinity TCR engager

Preferentially Expressed Antigen in Melanoma (PRAME) is one of the most compelling oncology targets available. It is expressed across a diverse range of cancer types, including gynaecologic malignancies, melanoma, sarcoma, lung, and leukaemia, while remaining largely absent from healthy adult tissue.

The market for PRAME-targeting therapeutics across melanoma, gynaecological cancers, head and neck cancers, and AML has been valued at approximately $8 billion¹. Early clinical data from various drug developers has shown a 61% clinical benefit rate and up to 78% tumour shrinkage.

That expression profile and strong evidence for therapeutic benefit create a strong rationale for PRAME-targeting therapeutics. But it also raises the bar for engineering. In a competitive PRAME landscape, our goal was to engineer a molecule that would stand out among the strong competitor candidates.

For a TCR-based T cell engager to be therapeutically viable, the TCR component which is responsible for cancer cell recognition must bind its target peptide-HLA complex at picomolar affinity to support effective tumour targeting.

Getting to picomolar affinity from a naturally occurring TCR, which typically binds its target in the micromolar range, requires precision engineering across multiple CDR loops. Done experimentally, that is a long and expensive optimisation process. Done computationally with EMLy, it becomes a more focused TCR engineering challenge.

The TCR discovery challenge: finding the right parental PRAME with machine learning

Before any engineering could begin, Etcembly needed the right starting PRAME TCR to engineer from. The discovery phase drew on EMLy's repertoire analysis capabilities, which combines tumour-infiltrating lymphocyte (TIL) repertoire data, network analysis, machine learning-based sequence assessment, and structural modelling to identify promising parental TCR candidates.

This wasn't a random search. EMLy's large language models (LLMs) were trained on hundreds of millions of natural TCR sequences, including Etcembly’s paired TCR dataset, which were used to embed and cluster candidate TCRs. Candidates were assessed to identify structural and sequence features consistent with productive PRAME peptide-HLA binding.

From this process, EMLy produced a shortlist of candidates for experimental validation. One parental TCR in particular, ETC-101 showed a binding affinity in the micromolar range providing a natural-affinity starting point with the right geometry to build from.

Figure 1 Structural modelling of the parental TCR: HLA (aqua) and peptide (yellow) shown with all 6 CDR loops, coloured by Root Mean Square Fluctuation (RMSF) determined over a 150 ns MD simulation. Higher RMSF values (coloured red) indicate regions with higher flexibility, whereas lower RMSF values (coloured blue) indicate regions with more rigidity. Specific loop residues have been predicted to be ideal for affinity enhancement.

The affinity optimisation challenge: Taking ETC-101 from micromolar to picomolar with EMLy

With a parental TCR in hand, the engineering challenge of optimising its affinity began.

EMLy's starting point is always the TCR sequence itself. From that sequence, the platform generates three-dimensional structural models of the TCR bound to its PRAME peptide-HLA complex (specifically the HLA-A*02:01/SLL epitope). It then produces hundreds of candidate poses and uses Etcembly's proprietary DoRIAT Bayesian framework to identify the conformations most likely to represent real, productive binding.

The crossing angle, the geometry of how the TCR sits relative to the peptide-HLA surface, is a critical determinant of whether a binding prediction reflects reality. When benchmarked against experimentally determined crystal structures, EMLy's predictions consistently outperformed general-purpose protein modelling tools. For the ETC-101 programme, in silico models achieved an RMSD of less than 2.5Å against its crystal structure, with highly aligned CDR loop positions. That level of structural accuracy gave the team a strong foundation for making engineering decisions.

From this foundation, EMLy ran a recursive engineering cycle:

1. Expert-guided mutagenesis design

2. Thousands of in silico variant assessments

3. Molecular dynamics simulations to understand CDR loop flexibility

   
   

Candidate selection was made, all before a single variant was synthesised.

When the first round of mutations was introduced computationally, EMLy predicted an approximate 5 million-fold improvement in binding affinity relative to wild type. Laboratory measurements confirmed what EMLy had predicted, validating the platform's ability to guide TCR engineering before experimental resources are committed.

Figure 2: Reaching target TCR affinity through the synergistic combination of EMLy predicted CDR mutations.

The results: 5 million-fold affinity improvement and a best-in-class ETCer

The numbers from the ETC-101 programme are striking:

The parental TCR had started at 58 µM. With EMLy optimisation, the lead engineered ETC-101 TCR achieved a KD of 11 pM with an 8 hour half-life. This improvement put ETC-101 firmly in the picomolar range required for effective tumour targeting and retention.

Expression yield improvements were equally significant. Combining alpha and beta chain mutations resulted in a 5-fold increase in expression yield. For therapeutic candidates, this matters because higher expression translates directly into manufacturing viability and lower cost of goods.

Importantly, the TCR showed no response to PRAME-negative cell lines and no cross-reactivity against a panel of PRAME mimetic peptides, confirming the selectivity essential for a viable clinical candidate.

When formatted as a bispecific ETCer (combining the high-affinity TCR with an anti-CD3 scFv effector domain), ETC-101 specifically targeted HLA-A*02:01 PRAME-positive cancer cells with no detected binding to PRAME-negative cells. In head-to-head comparison with a competitor benchmark, ETC-101 demonstrated superior anti-tumour efficacy and sensitivity in a patient-derived breast cancer organoid model (p < 0.0001).

Etcembly’s second-generation trispecific ETCer, incorporating an additional T cell co-stimulatory effector arm in addition to the anti-CD3 domain, pushed performance further still. Trispecific ETC-101 outperformed the bispecific format in serial killing assays and remained efficacious at low effector-to-target cell ratios. Crucially, it also overcame T cell exhaustion in a tumour microenvironment model, which is one of the key challenges facing current solid tumour therapies.

The safety and specificity challenge: Selective PRAME recognition

For any TCR-based therapeutic, safety is as important as potency. A TCR that binds well but cross-reacts with healthy tissue is not a viable drug candidate.

ETC-101 was evaluated against a broad panel of non-cancerous primary human cell types, including renal, cardiac, skeletal muscle, bronchial, lung, and female tissue. Across this testing negligible off-target reactivity was observed. The only low-level signal detected was in two renal cell types where background PRAME expression has been previously reported in the literature. Therefore, these findings were consistent with known biology rather than unexpected cross-reactivity.

mimetic peptide screening further confirmed ETC-101's highly selective peptide binding motif, with no cross-reactivity detected against a panel of five PRAME mimetic peptides in either the parental or engineered TCR formats.

A New Model for TCR Engineering

ETC-101 is now Etcembly's lead portfolio asset. Its development represents a proof of concept not only for PRAME as a target, but for what is possible when AI-guided engineering replaces the broad experimental search that has historically defined TCR programmes.

The economics of TCR engineering look different when the hit rate is this high. EMLy's predictions are validated, not discovered, in the lab. That means fewer variants to synthesis, faster iteration cycles and more confident progression.

The same platform that produced ETC-101 has now been applied to ETC-201, a second TCR engineering programme. Each engineering programme builds on the same computational foundation: purpose-built TCR-specific AI, delivered by a team that has demonstrated it can engineer molecules that work.

For biotech and pharma teams working on difficult TCR or antibody engineering challenges, ETC-101 shows how EMLy can help move from sequence to candidate with greater speed, precision, and confidence.

If you have a TCR or antibody engineering challenge, we'd like to hear about it. Get in touch at hello@etcembly.io

References:

¹ Pipeline and clinical development data sourced from Patsnap Pharma Intelligence (accessed June 2026); expression prevalence data from published literature; market size estimates based on analyst synthesis