Synthetic Lethality & AI: IP Strategy in Oncology

What is Synthetic Lethality?

Synthetic lethality (SL) describes an interaction between genes where the combined disruption of two genes results in cell death, but the disruption of either gene alone allows cell survival. Calvin Bridges discovered this phenomenon in fruit flies over 100 years ago. It has since been hailed as a promising treatment strategy in precision oncology, where targeting an SL partner gene in tumours with a specific mutation selectively kills cancer cells while sparing healthy cells.

Conventional targeted anti-cancer therapies tend to focus on direct mutation targeting and the inhibition of a specific oncogene; a mutant gene that can cause cancer. However, many tumours do not have targetable oncogenes, especially those with loss-of-function mutations, where a protein product no longer performs its normal function. Therefore, a significant proportion of cancer patients do not currently have access to appropriate targeted therapies. SL treatment strategies encourage mutation indirect targeting by considering gene interactions and thus provide an expanded spectrum of targets.

Mutations in BRCA1/2 genes are common in hereditary breast and ovarian cancers. PARP inhibitors, such as olaparib and rucaparib, have been used to treat BRCA1/2-mutated cancers and were the first clinically approved drugs designed to exploit SL. The PARP enzymes repair single-stranded DNA breaks, while BRCA1 and BRCA2 are tumour suppressor genes that repair double-stranded DNA breaks, through homologous recombination. BRCA1/2 mutations impair homologous recombination and cells then become more dependent on PARP-mediated repair. Inhibiting PARP causes DNA damage to accumulate, which BRCA1/2-deficient cells cannot repair, leading to the selective death of BRCA1/2-mutated cancer cells.

Despite the initial success of PARP inhibitors, there are still very few SL-based drugs clinically available. Reasons for this include a high proportion of patients treated with PARP inhibitors developing drug resistance, with tumour cells adapting through different pathways to counter the inhibition. Most importantly, there has also been a lack of effective techniques to accurately identify new clinically relevant SL gene pairs.

Emerging technologies and the use of Machine Learning

Following the initial success of PARP inhibitors, experimental methods such as yeast screens were used to identify SL targets. This led to the use of higher throughput human cell methods such as drug screens, RNAi screens, and CRISPR/CAS9 screens to identify gene pairs with SL relationships. However, with over 200 million gene pair combinations, genetic context considerations, and epigenetic modifications, screening all the potential SL pairs by wet-lab methods was impractical. Despite this, the results from these experiments remain a key resource for SL prediction and there are large databases that compile these gene-gene interactions.  

The Cancer Dependency Map project (DepMap) is currently the largest database for human cancer cell lines and uses data from multiple large-scale projects such as the Broad Institute’s Project Achilles [1]. Advancements in Machine Learning have led to the development of algorithms that use such large databases for SL prediction. Within Machine Learning, Deep Learning methods present a way of capturing non-linear relationships between inputs and outputs by leveraging multilayered architectures to build networks that simulate cellular systems, enabling the identification of more complex SL gene interactions.

Currently, most Deep Learning methods for SL predictions are based on supervised learning, where models are trained on high-quality labelled data to identify patterns. These data include both gene pairs known to have an SL relationship and those that do not exhibit an SL relationship, known as positive and negative-labelled data, respectively. Negative SL pairs are often difficult to prove by screening methods due to the complex interactions between genes. Moreover, the “black box” nature of Deep Learning models makes it difficult to explain the output SL predictions, meaning they may be overlooked by biologists, hindering the translational applications of these data.

Recent strategies designed to overcome these challenges include NSF4SL (Negative Sampling Free for SL) [2]. This is a prediction model based on contrastive learning that avoids the use of negative samples. It captures the characteristics of positive SL samples by using two parallel neural networks that interact with each other to learn the patterns of positive SL genes. Most importantly, the SL gene candidates are then ranked rather than assigned to a binary yes or no classification. This method can accelerate precision oncology as it allows researchers to streamline their work and focus on potential hits.

The interpretability of outputs can be improved by knowledge and data dual-driven Artificial Intelligence (AI) frameworks for SL prediction, such as KDDSL [3]. Unlike traditional “black box” models, dual-data driven methods integrate biological knowledge with gene expression data to identify SL candidates, enhancing both the predictive accuracy and interpretability of results. Such frameworks enable a bidirectional workflow, where findings from wet-lab experiments refine computational models and the model-generated predictions guide subsequent wet-lab investigations.

While recent advances show promise, several challenges remain. Deep Learning approaches are still highly dependent on the availability and quality of prior biological knowledge and large datasets, which can often be incomplete, non-standardised, or disease specific. Models with improved interpretability can still be highly complex and translating SL predictions into clinical applications is resource-intensive, often incurring lengthy timelines.

The future of Synthetic Lethality

Investment in synthetic lethality is expanding and is projected to grow over the next decade. There have been several major strategic deals in the SL sector of precision oncology, notably a deal reportedly valued at over €500 million between Boehringer Ingelheim and Tessellate BIO, announced in 2025. This partnership aims to develop targeted treatments for ALT-positive tumours, which represent 10-15% of all cancers. Tessellate Bio has developed inhibitors to target SL partner genes of ALT, including a method that reduces FANCM expression, for which a related European patent application (EP 20754058.4) has been filed.

There have been over 1,200 clinical trials involving SL-based interactions, encompassing both monotherapies and combination treatments. While the most established class of SL drugs remains PARP inhibitors, the field has extended into targeting new pathways such as metabolic vulnerabilities and novel DNA damage response targets. Advances are also likely to be made in identifying responder and non-responder patient cohorts more accurately, which can have a transformative effect on clinical trial design. With investments on the rise and a high number of ongoing clinical trials, IP will play a key role in protecting and developing pipelines for SL treatments and related advancements.

Recent growth in SL treatment filings include those that target BRCA1/2-mutant cancers, such as POLΘ inhibitors, as well as expansion beyond this to target new SL pathways, such as PRMT5 inhibitors that target MTAP-deleted cancers and WRN inhibitors that target microsatellite instability and/or defective DNA mismatch repair.

AI/ML-driven drug discovery platforms are increasingly used to identify SL interactions and therapeutic targets, raising questions around inventorship and ownership. In most jurisdictions, including the EPO and the USPTO, AI/ML systems cannot currently be named as inventors. As a result, human inventors must be identified even where AI plays a significant role in generating the invention. This is likely to create uncertainty in a scenario where models can autonomously generate candidates without human guidance through tightly defining guardrails.

Synthetic Lethality is transitioning from a promising biological concept to a viable treatment strategy in precision oncology, driven by expanding clinical pipelines, technological innovation, and growing investment in the sector. The progression of SL as a viable treatment strategy will be dependent on the integration of discovery approaches with translational applications and the navigation of a complex but promising IP landscape.

Author - Daniel Allen (Patent Assistant)

‍