“Surface ID” is a geometric deep learning system for high-throughput surface comparison based on geometric and chemical features. Surface ID offers a novel grouping and alignment algorithm useful for clustering proteins by function, visualization, and in silico screening of potential binding partners to a target molecule.

Significant advances in computational methods and hardware-accelerated scientific computing have enabled the dawn of a new era of medicine in which lifesaving therapeutic molecules can be designed and optimized with greater speed and precision than ever before. At Johnson & Johnson, we are leveraging data from across the pharmaceutical value chain, manifested in a knowledge graph to inform novel computational, deep learning models to drive innovation and disrupt the therapeutic research and development lifecycle; building and leveraging our collective institutional knowledge across therapeutic programs and indications in order to inform novel AI/ML models to accelerate the development of lifesaving therapies for patients across the globe.

This presentation will provide an overview of the key challenges faced when using AI/ML to guide the design and optimization of biologics, including multi-specifics. We will then discuss some of the techniques used within Amgen to address these issues. Finally, we will argue that certain challenges are best solved through collaborative mechanisms, such as federated learning.

The deep-learning model is no longer a black box. This talk will cover how we design a high-throughput assay to generate a protein-stability dataset for training the language model. In addition to predicting novel stable sequences using the trained model, such a model can educate us on what patterns or motifs make a particular sequence stable. Therefore, such protein stability knowledge extracted directly from the trained model becomes valuable for scientists redesigning a more stable protein library.

Machine learning has become an integral part of antibody discovery and development. I will describe how we designed the J.HAL antibody discovery library with a generative machine learning method and share experimental results from recent discovery campaigns. I will also outline our computational approach to the problem of antigen-specific antibody design and our plans for experimental validation.

I will discuss a new protein structure prediction model based on a novel coarse-grained protein structure that achieves atomic accuracy on antibody structure prediction and is orders of magnitude faster than other state-of-the-art models. In addition, I'll describe its application in various antibody property prediction and design tasks.

The disposition of biologics (including clearance and immunogenicity) are key properties that influence efficacy (no exposure, no effect); tolerability/safety (neutralizing or clearing anti-drug antibodies); and commercial success (route of administration, patient convenience). Compared to small molecules, there is a dearth of predictive preclinical models of clinical pharmacokinetics. I will discuss recent progress and challenges in predicting human PK.

The Larman laboratory creates technologies for unbiased characterization of serum antibodies at cohort scale. This seminar will provide an overview of our current antibody profiling capabilities, recent findings, and ongoing developmental efforts that seek to overcome existing limitations of high-throughput antibody analyses.

• ML doesn’t understand protein. Digital representation (numerical features) is needed as input
• Meaningful representation (features) is a key for ML models
• Protein can be represented as 1D sequence (one hot or embedding), 3D structure (point cloud of cartesian coordinates, or graphs with nodes and edges), or surface patches
• Surface ID is deep learning derived representation, encoding geometric and chemical properties, that can be used for surface patch comparison
• Applications of Surface ID include paratope clustering, PPI classification, database mining etc.

• Explore common challenges for end-to-end integration and enterprise deployment of AI/ML models across the R&D product lifecycle   
• How are organizations leveraging the growing suite of predictive models to inform and accelerate generative design and optimization of protein therapeutics?
• How can we foster collaboration between different departments, including research, development, and CMC, to establish AI as a core organizational discipline? 
• What are the opportunities & best practices for incorporating AI/ML models and integrated lab automation platforms from discovery to development?
• How are advancements in computational hardware and infrastructure driving innovation in our digital platforms and business processes?

AlphaFold has revolutionized the structure prediction of proteins alone or in the complex. The need for co-evolutionary sequence constraints for structure prediction limits its use against antibody- antigen complexes. We predicted the structure of antibody- antigen complexes using traditional physics-based protein-protein docking tools. We evaluated the ability of AlphaFold in the quality assessment of models. Our results highlight that AlphaFold can rescue poorly-ranked models and better discriminate good-quality models from decoys.

Photooxidation of methionine (Met) and tryptophan (Trp) residues is common and includes major degradation pathways that often pose a serious threat to the success of therapeutic proteins. We applied the random forest machine learning algorithm to in-house liquid chromatography-tandem mass spectrometry (LC-MS/MS) datasets (Met, n = 421; Trp, n = 342) of tryptic therapeutic protein peptides to create computational models for Met and Trp photooxidation. We show that our machine learning models predict Met and Trp photooxidation likelihood with accuracy and further identify important physical, chemical, and formulation parameters that influence photooxidation.

Developability, the set of physicochemical properties of an antibody relevant for manufacturing and success in clinical trials, is one of the key determinants for success during clinical testing and any developability parameters can be computed from the antibody sequence and structure. Although the distribution of developability parameters of natural antibody repertoires may provide guidance on the potential suitability of therapeutic antibody candidates, the sequence and structural distributional landscape of the natural antibody repertoire has not yet been described. We quantify the redundancy, sensitivity, and predictability of developability parameters in natural and clinical-stage antibodies. Exploiting the vast amount of available antibody high-throughput data will facilitate the derivation of the rules underlying developability profiles to guide antibody therapeutic discovery.

Multi-specific biologics are of interest due to the advantage of engaging distinct targets. One important component is the scFv, but their relatively poor thermostability often hampers development. As experimental methods are laborious and expensive, computational methods are an attractive alternative. Here, we show two machine learning approaches – one with pre-trained language models (PTLM), and second, a supervised convolutional neural network (CNN) trained with Rosetta energetics – to better classify thermostable scFv variants from sequence. On out-of-distribution sequences, we show that a simple CNN model outperforms a general PTLM trained on diverse protein sequences (Spearman ?=0.4 vs 0.15).

Over the years, there has been an increased focus on decreasing non-specific binding during early-stage drug development. It has been recognized as a root cause for failure in many drug programs due to unexpected pharmacokinetics and elevated toxicity. From a computational design perspective, prediction has remained a challenge. Proposed work describes the development of sequence-based machine learning models for prediction of this property with accuracy of up to 74%, enabling flagging and deselection of non-specificity at an early-stage.