Benzinga | Jun 25, 2021 07:31AM EDT

Evaxion Biotech Reports Acceptance Of New Scientific Paper By Int'l. Conference On Machine Learning

Evaxion Biotech A/S (NASDAQ:EVAX), a clinical-stage biotechnology company specializing in the development of AI-driven immunotherapies to improve the lives of patients with cancer and infectious diseases, announced today the acceptance of a new scientific paper by the International Conference on Machine Learning (ICML 2021). A draft of the article is available on the open-access scientific server bioRxiv.org.

The paper is entitled "Efficient Generative Modelling of Protein Structure Fragments using a Deep Markov Model", and was written and developed by Evaxion personnel in collaboration with Assoc. Prof. Thomas Hamelryck's probabilistic programming group at the University of Copenhagen. The paper describes BIFROST, a novel predictive system based on deep probabilistic programming that enables the rapid conversion of sequence data into structural information on protein fragments, which we believe may be useful for drug or vaccine design. Deep probabilistic programming is a new development in machine learning that combines the principled treatment of uncertainty provided by Bayesian statistics with the capabilities of deep learning. Compared to existing protein structure prediction approaches, BIFROST appears to be computationally more efficient, only requires sequence information and, importantly, incorporates an assessment of the reliability of its own predictions.

Lars Wegner, CEO of Evaxion, said: "This work is an exciting development by the collaborative team that we believe has the potential to make vaccine development more efficient. We intend to apply our expertise to further the development of Bayesian machine learning and to integrate these methods fully into Evaxion's AI platforms, including both our EDEN and RAVEN platforms for vaccine development."

Protein structure prediction methods such as BIFROST have the potential to facilitate AI-driven pharmaceutical design by indicating the likely conformation that components of immunotherapies or vaccines and their target might adopt. Existing methods for predicting the conformation of protein fragments do not explicitly evaluate the probability of conformations given the sequence which can make it difficult to dissect the reliability of subsequent calculations. By including estimates of uncertainty in predictions, BIFROST's Bayesian approach may be particularly useful in drug development datasets that, typically, are incomplete and relatively small.