Turning science into medicine with computational & ML-empowered tools for biologic molecule design. Andrew will give an overview of the languages of biotherapeutics & machine learning, discuss the impact of quality curated data, and finally, demonstrate applications of machine learning in biomolecule design.

Accurate prediction of binding interaction between TCRs and host cells is fundamental to understanding the regulation of the adaptive immune system as well as to developing data-driven approaches for personalized immunotherapy. In this talk, I demonstrate the importance of devising a suitable embedding technique and present our context-aware amino acid embedding models (catELMo) designed explicitly for TCR analysis.

We developed a machine learning model to predict immune checkpoint blockade (ICB) response by integrating genomic, molecular, demographic and clinical data from a comprehensively curated cohort of patients treated with ICB across 16 different cancer types. In a retrospective analysis, the model achieved high sensitivity and specificity in predicting clinical response to immunotherapy and predicted both overall survival and progression-free survival in the test data across different cancer types. Our model significantly outperformed predictions based on tumor mutational burden, which was recently approved by the U.S. Food and Drug Administration for this purpose. Thus, this approach has the potential to substantially improve clinical decision-making in immunotherapy and inform future interventions.

• Mastering the Data Lifecycle: Best practices for end-to-end data product engineering while ensuring data quality and relevance. ​
• Data Governance Done Right: Implementing data management strategies to maintain data integrity and consistency throughout the pipeline.
  
• Streamlined Data Pipelines: Architecting efficient data processing workflows that enable traceability, data integrity, and reproducibility.
  
• Collaboration is Key: Fostering interdisciplinary teamwork between data scientists, engineers, and domain experts to design robust data products.
  

Optimising therapeutic antibodies across multiple properties is challenging. For T cell engagers (TCE) targeting solid tumours, cancer-vs.-normal cell selectivity is particularly difficult to achieve. LabGenius’ lead optimisation platform generates high-quality data from complex assays for machine learning to decipher design-fitness relationships and guide screening efforts to fruitful areas of the design space. We demonstrate our capability by discovering HER2 TCEs up to 400-fold more tumour-selective than a clinical benchmark.

In cancer immunotherapy, AI, ML, and DL are critical in developing precision-based targeted therapeutics and personalized molecular engineering. The complex cell signaling pathways, protein-to-protein interactions (PPI), cancer hallmark expressions, and the integral relationship of the differential multi-omic data expression necessitate a high confidence ranking, mapping, modeling, and selection architecture in developing small molecules, peptides, and peptide oligonucleotide conjugates. Cancer immunoediting is a molecular control process requiring a robust multi-omic design pipeline. The key features to consider in creating a design are immunoediting in cancer (elimination, equilibrium, escape) and mechanisms of action (immunosurveillance, immune induction, regulation, augmentation, signal editing, and adaptation.)

Accurate identification of T cell targets is critical for the development of specific and potent vaccines against cancer or infectious diseases. Gritstone’s EDGE is a state-of-the-art AI-driven platform that can identify T cell targets with high accuracy. The platform is powered by the best advancements in AI and rich datasets. We have demonstrated that the targets identified by EDGE elicit immune response in patients.

During the development of our clonal neoantigen T cell therapy (cNeT) for treatment of solid tumours we have identified and screened circa 10,000 clonal neoantigens for T cell reactivity using cells grown from tumour infiltrating lymphocytes (TIL).  Using this unique dataset of bona fide TIL-derived memory T cell reactivities for both CD4+ and CD8+ T cells, we have developed and validated an AI method for predicting neoantigen immunogenicity. When combined with our proprietary capability to identify clonal neoantigens this has broad applicability for optimising target selection across all types of personalised neoantigen therapies.

Multiplex immunofluorescence tissue imaging is increasingly used to identify new spatial signatures that predict responses to immunotherapy. However, identifying new signatures typically requires a prior selection of cell types, marker expression levels, and their spatial interactions, which can be challenging as the number of potential signatures quickly increases with the number of markers. Here, we show how annotation-free AI can identify spatial signatures in an unsupervised fashion, potentially revealing previously unappreciated facets of the tumor microenvironment which can be validated using independent validation cohorts. We also explain how AI-driven patient predictions may be applied in potential clinical scenarios.

The aim of this analysis was to evaluate the role of multimodal machine learning in predicting pre-treatment response in advanced urothelial carcinomas receiving immunotherapy. Relevant data was acquired from gene expression omnibus datasets project PRJNA735749. Responders were categorized as patients with either stable disease, complete, or partial response while non-responders were those who had progressive disease. Four distinct algorithms were developed using different data and analytical profiles and their performances were evaluated. The MLP/deep-learning classifier shows better overall results compared to other algorithms. However, further external validation is required for these results.

We utilized AlphaSeq protein-protein interaction (PPI) measurements and machine learning to generate tens of thousands of diverse high-affinity pembrolizumab variants. We generated over one hundred thousand PPI measurements from three design-test cycles, incorporating transfer learning of unrelated antibody:antigen PPI data and state-of-the-art protein language modeling techniques (ESM2). Finally, we validated 20 predicted improved variants from the model and demonstrated significant improvements across multiple biodevelopability metrics.