COMPARATIVE ANALYSIS OF ALPHAFOLD2 & ESMFOLD ALGORITHMS TO SUPPORT BIOSIMILARITY & DRUG DISCOVERY

Abstract

The elucidation of three-dimensional protein structure plays a pivotal role in comprehending biological phenomena. It directly governs protein function and hence aids in drug discovery. Development of protein prediction algorithms, AlphaFold2 and ESMFold, have the potential to shift the paradigm of protein-based therapeutic discovery. Turning an amino acid chain into 3D domains and docking them can aid in unlocking a protein’s full potential. Besides this, the effects of mutations on the domain structure can be studied meticulously. Prediction scores from extensive studies were examined in the hope of searching for newer modalities of transforming protein therapeutics. Most of these studies failed to find any utility of these algorithms, and a few suggested, despite their dismal findings, that their utility can be found. The inventors of the algorithms cautioned that the predicted structures and scores have no utility except regurgitate known structures from the known structure databases. A few possible applications, as considered in this study, are to predict pre-translation variations, mutations, and structural changes. A potential correlation of repeatedly manufactured batches of therapeutic protein is correlated with the structure prediction score as a measure of thermodynamic instability. 204 unmodified FDA-approved therapeutic proteins were correlated with their prediction scores and available physicochemical and functional properties. Slight residual differences among the commercial therapeutic proteins and structures reported in the PDB were found. The potential impact of mutations on the prediction scores is also studied. No observed correlation was found between the prediction score and any tested attribute. The algorithms exhibited lower confidence in predicting structures for sequences with low identity scores when tested against the UniProt and PDB databases. Other deployed algorithms (i.e., trRoseeta) were concluded to be more relevant to domain manipulation as well. Reliable structure prediction from these algorithms highly depends on the model’s architecture and training data. Ultimately, it was concluded that none of these algorithms have any 1 Abstract value except they show how good they can be at reproducing a known or partiallyknown structure. The comparison of AF2 and ESMF resulted in R 2 of 0.69, vouching for their orthogonality. However, the R 2 value of physiochemical attributes was as low as 0.07. Lack of significant correlation of predictability scores with physicochemical and functional properties cannot vouch for in-vivo stability and molecular functionality of a protein. Furthermore, when novel randomized and mutated sequences are provided to these algorithms, they fail to predict structures with acceptable accuracy. This is majorly due to the unavailability of similar folds in the training dataset (i.e., UniProt and PDB) of these algorithms. Although it might seem that these algorithms go beyond regurgitating available data, it might not be the case. In this context, these algorithms are considered no different than GPT4, which also regurgitates available data. These algorithms do not play well in proving the Levinthal paradox as solved, yet it remains unsolved.