BigHat Biosciences is a novel protein therapeutics company and developer of an AI-guided antibody design platform
Development of therapeutic antibodies usually starts with antibody libraries that are synthetically constructed or obtained from immunized animals. From these libraries, candidate molecules with affinity to the desired target can be selected. However, the initial antibodies found by these display technologies or immunization are rarely clinic-ready, especially with next generation antibodies showing affinities barely in the therapeutically relevant range. Therefore these initial candidates require significant engineering to improve not just their affinity but also other biophysical properties that make a high quality therapeutic such as solubility, thermostability, and aggregation propensity of the molecule. BigHat’s novel AI-guided experimental platform offers an innovative approach using ML techniques ideally suited for multi-parameter optimizations to chart an efficient experimental path to find higher quality variants of the initial antibody. BigHat’s platform is based on using machine learning coupled with iterative design-build-test cycles of sequences in an automated wet lab. By repeating such rapid design-build-test cycles multiple times and using the test data from all previous cycles to inform the design for the next cycle, the platform can guide a highly efficient multi-parameter sequence optimization. This AI-guided smart selection of high-value test sequences provides a highly efficient and cost-effective approach to discovering better antibody sequences with the required biophysical properties, allowing BigHat to optimize existing lead therapeutic candidates for more effective treatment and create antibodies addressing unmet needs inaccessible to traditional design approaches. An introduction to BigHat: https://www.youtube.com/watch?v=K4raSxBjryk
Currently, antibody engineering is performed in sequential expert-guided steps to improve primarily one or, at most, a few of the relevant properties. This stepwise approach is not only costly and time-consuming, but also prone to failure because the relevant properties are not independent, making sequential optimization difficult. Changes to the antibody molecule that improve some properties in one step may negatively impact others in the next step. In addition, expert-driven engineering approaches typically only explore a limited number of the antibody molecule during each optimization step, based on data from other engineering campaigns and the literature; thereby likely missing novel and superior antibody variants. BigHat is able to solve this more efficiently through their platform which is well suited for multi-parameter optimization, enabling the discovery of novel variants that evade current methods and also find pathways to optimize existing lead therapeutic candidates to increase treatment efficacy. Unlike other platforms that have simply tacked on bioinformatics to their programs, BigHat’s platform was designed and built from the ground up with a ‘fit-for-purpose’ philosophy. Using an automated high speed wet lab combined with BigHat’s proprietary LIMS, all experimental data is seamlessly fed into the platform allowing machine learning techniques to continuously iterate through design cycles, significantly reducing the latency between experiments and more efficiently search an intractably large sequence space. This unique platform design enables a highly efficient and faster path to drug discovery than other existing platforms as well as enabling novel designs beyond today’s capabilities
There are many ways that BigHat works with partners to create high-quality clinic-ready antibodies. This can be as narrow as designing antibodies to overcome inadequate biophysical properties in lead candidate antibodies, where BigHat works with these partners to onboard these antibodies to iteratively optimize away the liabilities while preserving or sometimes enhancing the other biophysical properties. BigHat can also work with partners in a broad end-to-end partnership starting with de novo discovery through design of high quality antibodies across multiple properties to target unmet indications. This is the most flexible way BigHat works with partners starting at the target of the therapeutic hypothesis and working jointly to create an end-to-end discovery and development plan that leverages its platform to produce ultra high quality molecules. Currently BigHat is engaged in an undisclosed partnership with a large pharma and are refining collaborative proposals with multiple large pharma and biotech companies. BigHat recently published its first peer-reviewed journal article at the prestigious International Conference for Machine Learning (ICML) Workshop on Computational Biology. BigHat researchers collaborated with advisor and Bayesian deep learning expert Andrew Gordon Wilson of NYU, to develop a framework for protein design that requires only a small amount of labeled data. The real-world utility of these methods were demonstrated by optimizing the Stokes shift of multiple fluorescent proteins.