AI biotech Galux rolls out new T-cell immunogenicity model to cut protein drug failures

Korean AI drug discovery startup Galux has built an immunogenicity prediction model that it says can spot T-cell risks earlier in protein drug development, with supporting data now published in Science Advances.

Working with a team at Seoul National University, the company developed an AI model dubbed T-SCAPE (T-cell Immunogenicity Scoring via Cross-domain Aided Predictive Engine) to forecast T-cell immunogenicity for drug candidates, from peptide–MHC (pMHC) complexes to full therapeutic antibodies.

In benchmark tests on neoantigen and infectious disease vaccine datasets, Galux’s T-SCAPE model (black bars) delivered the highest precision and overall performance in T-cell immunogenicity prediction compared with existing tools. (Courtesy of Galux)

According to Galux and the study’s authors, immunogenicity remains a persistent hurdle in protein therapeutics, with unwanted immune responses that can blunt efficacy or trigger safety problems.

They argue that sparse datasets and complex immune biology have made it difficult to generate reliable, quantitative predictions of which candidates are most likely to provoke T-cell responses.

T-SCAPE is designed to work around that constraint by training on multiple immunology datasets rather than a single readout.

The model ingests human and non-human peptide sequences, MHC binding data, T-cell receptor interaction information and T-cell activation assay results, linking these signals to capture patterns that single-source models may miss, the company said.

Because direct immunogenicity data are limited, Galux first “pre-trains” the system on broader biological principles, then applies an adversarial domain adaptation approach to narrow the gap between heterogeneous datasets and search for common rules, said Noh Jin-sung, an AI researcher at Galux who co-led the work. The company said that strategy led to a sharp improvement in predictive performance.

In benchmark testing, T-SCAPE outperformed existing tools on pMHC immunogenicity prediction and was also able to estimate antidrug antibody (ADA) risk for therapeutic antibodies with high accuracy, according to Galux, which suggested the model could serve as a reference metric for broader immunogenicity screening.

CEO Seok Cha-ok said the platform adds an immune risk filter on top of Galux’s AI protein design stack, with the aim of reducing uncertainty and dead ends in the candidate selection step and improving overall efficiency in drug development.