1233 Predictive and generative AI to guide the clinical development of HFB200301, a first-in-class TNFR2 agonist: drug intelligence science (DIS®)

<h3>Background</h3> Predictive and generative artificial intelligence (AI) models can complement decision-making in Discovery and Early Development. For instance, generative AI can bridge data gaps by identifying patterns across data modalities such as single-cell, imaging, and medical data. Predictive models can then link these patterns to drug efficacy and safety outcomes in patients. HiFiBiO has developed a multi-modal generative AI and predictive modeling platform called Drug Intelligence Science (DIS®) to support decision making for cancer and autoimmune therapies. This case study presents results from the platform that have guided the development of HFB200301, our first-in-class TNFR2 agonist (NCT05238883). <h3>Methods</h3> <i>Pathological cell identification and target credentialing:</i> To identify the pathological cell types and their targets for immune therapies, we employed generative AI and deep learning with variational autoencoders to create a Disease Cell Atlas with &gt; 15 million human cells from public and internally generated single-cell data. For oncology targets we emphasize curation and modeling of data from patients refractory to anti-PD-(L)1 therapy. <i>Tumor type selection:</i> We have developed a series of multi-modal machine learning and predictive models that integrate preclinical data with accumulating peripheral and tumor data from our clinical trials to select and to refine the tumor types and subgroups most likely to benefit from immune therapies like HFB200301. <i>Proof of mechanism in the tumor:</i> To identify and to quantify the on-mechanism action of our compounds we employed convolutional neural networks on paired H&amp;E and multiplex immunofluorescence images (mIF) to delineate individual cells, their states, densities, and spatial locations. <h3>Results</h3> Analysis of our Disease Cell Atlas in the anti-PD-(L)1 refractory setting revealed that TNFR2 is primarily expressed on CD8+ T-cells lacking PD-1 expression, suggesting that agonism may mobilize a distinct set of CD8+T-cells. In our HFB200301 TNFR2 agonist clinical trial (NCT05238883), several tumor types selected by our predictive models, such as anti-PD-(L)1 refractory pleural mesothelioma and EBV+ gastric cancer, have shown either monotherapy or combination anti-tumor activity with the anti-PD-1 inhibitor tislelizumab, which suggests that our predictive models are surfacing the tumor types that will benefit from HFB200301. Moreover, in the same patients exhibiting anti-tumor activity, mIF analysis revealed marked increases in activated CD8+ T-cells following compound treatment, suggesting that the anti-tumor activity results from the on-mechanism action of HFB200301. <h3>Conclusions</h3> These case studies demonstrate how our multi-modal generative AI and predictive modeling platform, Drug Intelligence Science (DIS®), can support decision-making and increase the probability of success in the clinic. <h3>Trial Registration</h3> NCT05238883.