How KSQ Therapeutics uses CRISPR screens and machine learning to engineer better T cells for solid tumour immunotherapy

Most cancer patients have T cells in their tumours that are trained to recognise and kill cancer cells. The problem is that these T cells are outpaced, suppressed, and exhausted by the time they matter most. KSQ Therapeutics is changing that with a combination of genome-wide CRISPR screening, gene editing, and machine learning.

Engineering T cells to survive the tumour microenvironment

KSQ extracts tumour-infiltrating lymphocytes (TILs) from a patient's solid tumour, engineers them ex vivo by knocking out one or two specific genes, and infuses them back at billions of cells. Unlike CAR-T therapies that redirect T cells toward a single antigen target, TILs are polyclonal: their native TCRs have already been selected by the patient's immune system to recognise neoantigens and tumour-associated antigens across the full heterogeneity of the tumour. That is a structural advantage in solid tumours, where cancer cells routinely downregulate surface markers to escape single-target therapies.

The gene knockout targets, SOCS1 and REGNASE-1, were identified through genome-wide CRISPR screens covering all 20,000 genes across multiple cancer cell lines and immune cell types. Knocking out SOCS1 removes a negative feedback brake on cytokine signalling, making T cells hyperresponsive. Knocking out REGNASE-1, an RNA-binding protein whose targets include genes governing effector cytokine function, prevents the T cells from differentiating into an exhausted phenotype and allows them to persist and proliferate in hostile tumour environments.

Two programmes are in clinical trials: KSQ-01EX (single SOCS1 knockout) and KSQ-04X (SOCS1 plus REGNASE-1 double knockout). Safety data from KSQ-01EX was presented at AACR, with early hints of efficacy.

Predicting tumour reactivity before manufacturing: TRACE

A fundamental challenge in TIL therapy is understanding what is in the drug product. After ex vivo expansion, thousands of T cell clones are present, but only a fraction are tumour-reactive. That fraction ranges from low single digits to high double digits depending on the patient and tumour type.

KSQ built TRACE, a machine learning classifier that uses paired single-cell TCR sequencing and RNA sequencing from the initial tumour harvest to predict, at individual clone level, whether a given T cell is likely to be tumour-reactive. Rather than relying on TCR sequence alone, TRACE reads the transcriptional state of each clone. Tumour-reactive T cells occupy a distinct transcriptional niche after sustained exposure to tumour-associated antigens, characterised by a signature detectable by single-cell RNA sequencing.

TRACE was trained on publicly available datasets, supplemented with KSQ's own annotated internal data. The model code and weights are fully open-source and available for any group to use, retrain, or adapt.

TCR sequencing across the manufacturing and clinical pipeline

Bulk TCR sequencing runs throughout KSQ's operations. During manufacturing it monitors polyclonality, ensuring that expansion does not accidentally collapse a diverse TIL repertoire into a narrow set of dominant clones. After infusion, blood-based TCR sequencing tracks drug product clonal dynamics in the patient's circulation, distinguishing infused clones from the patient's recovering endogenous repertoire. With TRACE it should be possible to identify putative tumour-reactive clones before infusion, and then track exactly those clones in post-treatment blood and tumour biopsies, connecting repertoire dynamics to clinical response.

About the guest

Dipen Sangurdekar is VP and Head of Data Science at KSQ Therapeutics, where he leads computational biology and bioinformatics across the R&D pipeline. He trained as a chemical engineer before completing a PhD and postdoc in computational biology and has spent fifteen years at the preclinical-to-clinical translation interface in oncology and neuroscience.

Listen to the full episode: YouTube | Spotify