Unmet Need
Antibodies are one of the most important tools in medicine. Although they have long been essential components of research experiments and diagnostic assays, their use as therapeutics is revolutionizing the treatment of many autoimmune and cancerous conditions. The antibody drug market is estimated at $126.7 billion (BCC Research) and the antibody testing market is valued at $7.4 billion (Persistence Market Research). Antibodies’ enormous specificity for their target is responsible for a large portion of their value, but makes discovering and refining them for a specific task expensive. The custom antibody market is valued at $393 million (Markets and Markets). Antibodies suitable for a desired task are currently identified through experimental antibody screens that involve antigen identification, purification methods, usage of a living organism for development and collection of antibodies, and refinement to improve certain properties. The cost of this process depends on the specific antibody but is typically in the thousands of USD range at the least. The main limitation of this process is the expense and time involved in this labor-intensive process. Solutions that address the above limitations are likely computational in nature. Therefore, there is a strong need for a rapid, inexpensive, likely computational solution to be developed to simplify the difficulties in identifying promising antibodies.
Technology Overview
Researchers at Johns Hopkins have developed a machine learning method to identify promising antibody candidates from datasets of antibody sequences, and provides suggestions for which parts of the antibody may be involved in binding to the target. The researchers developed a model based on BERT, a popular language model, and trained it on 558 million antibody sequences. They used a multiple instance learning framework to predict whether a set of antibody sequences is likely to contain a subsequence of amino acids involved in binding to the target.
Further benefits of this method, in addition to its enormous time and expense-saving capabilities, include its easy potential for improvement and personalization. An improved model architecture or training dataset could be easily integrated. For personalized antibody treatments, the model could be further trained on an individual patient’s antibody sequences.
Stage of Development
The invention is in the proof of concept stage and requires experimental validation.
Publication
Ruffolo, J. A., Gray, J. J., & Sulam, J. (2021). Deciphering antibody affinity maturation with language models and weakly supervised learning. arXiv preprint arXiv:2112.07782.
