A deep learning approach for the discovery of SOS1-KRAS interaction inhibitors for anti-cancer treatment

Abstract

Cancer is one of the global diseases that poses a serious threat to human health. At present, the global burden of cancer incidence and mortality is increasing year by year, especially in low- and middle-income countries, where the growth rate is even faster. According to the latest statistical report released by the International Centre for Research on Cancer (IACR) of the World Health Organisation (WHO) in 2024, there are estimated to be 20 million new cases of cancer and 9.7 million cancer deaths worldwide, accounting for 16.7% of the total global deaths, second only to cardiovascular disease. Among all malignant tumour cases, lung cancer is still the number one killer of cancer in the world, followed by pancreatic cancer and colorectal cancer, accounting for 12.4%, 9.6% and 3.3% of new cases in 2022, respectively.
Kirsten Rat Sarcoma Viral Oncogene Homolog (KRAS) gene represents one of the most frequently mutated oncogenes in human cancers and significantly influences the occurrence and development of malignant solid tumours with high mortality rates, including pancreatic, colorectal, and especially lung cancers. Activating RAS mutations are present in about 30% of human cancers, whereas KRAS is significantly more frequently mutated compared to HRAS and NRAS. KRAS normally cycles between its active state-bound guanosine triphosphate (GTP) and its inactive-state-bound guanosine diphosphate (GDP). However, compared to its normal state, KRAS continuously remains activated following mutations, especially at codon 12. Further disruption of normal KRAS function by these mutations and sustained activation promotes sustained proliferation, survival, and metastasis of tumour cells through the activation of downstream signalling pathways, such as the Mitogen-activated protein kinase (MAPK) and PI3K-AKT pathways.
Over the past few decades, KRAS has been considered an "undruggable" target and thus difficult to target with GTP-competitive inhibitors, with a smooth and almost spherical surface lacking well-defined "pocket" structures that small-molecule inhibitors can bind to. This structural feature can make it difficult for traditional small-molecule drugs bind to appropriate sites. Moreover, KRAS exhibits a very high binding affinity for GTP, in the picomolar range, whereas the intracellular concentration of GTP is approximately 500 μM. This low binding affinity makes it hard for drugs to compete with GTP, thereby limiting their ability to bind to and inhibit the function of the KRAS protein.
Recent advances in structure-based drug design, high-throughput screening, and covalent inhibitors have revitalised efforts to develop potent KRAS inhibitors, overcoming historic challenges. In May 2021, Amgen’s KRAS G12C inhibitor, Sotorasib (Lumakras, AMG510), received accelerated US Food and Drug Administration (FDA) approval, challenging the view of KRAS as an “undruggable” target and offering new hope for patients with KRAS mutations. Following this, on February 16, 2022, the FDA accepted Mirati’s New Drug Application (NDA) for another KRAS G12C covalent inhibitor, Adagrasib (KRAZATI, MRTX849).
Although the therapeutic strategy for KRAS mutations has made significant breakthroughs, research on KRAS inhibitors still faces major challenges due to the complex structure of the KRAS protein and complex signalling pathways. Currently, FDA-approved small-molecule KRAS inhibitors are only available for patients with KRAS G12C mutations. Clinical data show that the efficacy of these inhibitors sometimes fails to meet expectations, and there is a risk of toxicity. Moreover, the long-term challenge of treating KRAS mutants is further complicated by emerging clinical resistance.
Given the current research landscape, developing therapies for a wide range of KRAS mutants and overcoming KRAS inhibitor resistance are urgent problems. As a critical guanine nucleotide exchange factor (GEF) involved in KRAS activation, Son of Sevenless homolog 1 (SOS1) protein plays a crucial role in this process. As a direct upstream regulatory molecule of KRAS, SOS1 activates KRAS and maintains its key functions in cell proliferation and tumorigenesis by promoting the exchange between GDP and GTP of KRAS protein at specific catalytic sites. Therefore, targeting SOS1 to intervene in the activation of the KRAS signalling pathway provides a new research direction for the development of new anti-KRAS driven tumour treatment options. SOS1 serves as a potential therapeutic target for KRAS mutations and represents a significant breakthrough in overcoming KRAS inhibitor resistance.
In recent years, targeting the inhibition of the SOS1-KRAS interaction has emerged as a new strategy for the development of pan-KRAS drugs. Although the related treatments are still in clinical research, due to their innovative designs and preliminary clinical results, SOS1 inhibitors such as BI 1701963, MRTX0902, and BAY3498264 have shown potential for treating KRAS-driven cancers. These inhibitors are entering clinical trials as single agents or in combination, which is expected to broaden the treatment options for patients with KRAS mutations and become an essential part of pan-KRAS treatment strategies. However, currently, the known SOS1 inhibitors remain structurally limited. Compounds such as BAY293, BI-3406, and MRTX0902 exhibit strong homogeneity, predominantly featuring quinazoline-based scaffolds and binding to similar sites on the SOS1 protein. This high degree of structural homogeneity limits the comprehensive exploration of the diverse binding pocket of the SOS1 protein.
Coupled with the rapid development in both biotechnology and computer science, drug-screening technology has leaped forward from traditional laboratory methods to modern High-throughput Screening (HTS) and virtual screening. Since its invention, HTS has totally changed the concept of new drug discovery and development. The capability of this technology to rapidly screen millions of compounds has greatly enhanced the speed and accuracy of drug discovery, especially with the integration of machine learning, which has accelerated this process. The classically long and complicated drug discovery pathway includes time-consuming and costly steps such as target identification, compound screening, and optimisation. On the other hand, machine learning can virtually screen potential drug candidates from large compound databases with high accuracy using virtual screening and molecular dynamics simulations, thereby greatly speeding translation from laboratory to clinic. Thus, the integration of high-throughput screening technology with machine learning techniques hastens not only the revolution in drug screening but also greatly shortens the timeline needed with reduced associated costs to open unprecedented possibilities for rapidly developing a new drug.
Building on the aforementioned background, this study systematically conducted a multi-dimensional discovery and optimisation project to develop novel SOS1 inhibitors, with SOS1 as the central therapeutic target. First, KRAS G12C/SOS1 and KRAS G12D/SOS1 protein protein interaction (PPI) homogeneous time resolved fluorescence (HTRF)-based molecular screening assays were independently optimised and established, providing robust and reliable platforms for accurate evaluation of the molecular activities of candidate SOS1 compounds. On this basis, cell proliferation inhibition assays were further optimised to construct a three-dimensional (3D) cell-based screening platform highly sensitive to SOS1 inhibition. In parallel, functional screening systems were established to systematically assess the inhibitory effects of candidate compounds on KRAS downstream signalling pathways across cell lines harbouring multiple KRAS mutations. Meanwhile, tumour engraftment conditions in nude mice were systematically optimised, leading to the successful establishment of stable NCI-H358 lung cancer and MIA PaCa-2 pancreatic cancer xenograft models, which laid a solid experimental foundation for in vivo pharmacodynamic evaluation of candidate SOS1 inhibitors.
Leveraging these integrated molecular-, cellular-, and animal-level evaluation platforms, multiple drug discovery strategies were employed, including machine learning-based virtual screening, structure modification of known active scaffolds, and a combination of conventional screening with structure-based virtual screening. Through these multi-path exploration efforts, a series of structurally novel SOS1 inhibitor candidates was identified, among which several compounds exhibited favourable antitumour activities in both in vitro and in vivo.

Date of Award	15 Sept 2026
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Bencan Tang (Supervisor), Hua Xie (Supervisor), Sze Shin Low (Supervisor) & Wan Yong Ho (Supervisor)