TLmutation: predicting the effects of mutations using transfer learning

Abstract

A reoccurring challenge in bioinformatics is predicting the phenotypic consequence of amino acid variation in proteins. Due to recent advances in sequencing techniques, sufficient genomic data is becoming available to train models that predict the evolutionary statistical energies for each sequence, but there is still inadequate experimental data to directly predict functional effects. One approach to overcome this data scarcity is to apply transfer learning and train more models with available datasets. In this study, we propose a set of transfer learning algorithms, we call TLmutation, which implements a supervised transfer learning algorithm that transfers knowledge from survival data to a protein function of interest in the same protein followed by an unsupervised transfer learning algorithm that extends the knowledge to a homologous protein. We explore the application of our algorithms in three cases. First, we test the supervised transfer on dozens of previously published mutagenesis datasets to complete and refine 1 missing datapoints. We further investigate these datasets to identify which variants build better predictors of variant functions. In the second case, we apply the algorithm to predict higher-order mutations solely from single point mutagenesis data. Finally, we perform the unsupervised transfer learning algorithm to predict mutational effects of homologous proteins from experimental datasets. These algorithms are generalized to transfer knowledge between Markov random field models. We show the benefit of our transfer learning algorithms to utilize informative deep mutational data and provide new insights into protein variant functions. As these algorithms are generalized to transfer knowledge between Markov random field models, we expect these algorithms to be applicable to other disciplines.

Combining evolutionary and assay-labelled data for protein fitness prediction

“…We assess all methods on 19 labelled mutagenesis datasets, each comprising hundreds to tens of thousands of mutant sequences. Most related works 21,23,24,27 evaluate on a subset of or all of the mutation effect datasets introduced by EVmutation 20 . We included all EVmutation 20 protein data sets with at least 100 entries in order to have sufficient data to glean insights from.…”

“…Each of these labelled data sets is paired with an evolutionary data set found by searching through the UniRef100 database 39 with the wild-type protein sequence as query. A jackhmmer 43 search is commonly used for retrieving evolutionary data 20,21,24,26 ; however, these search results can differ depending on the jackhammer parameters used. Thus, for simplicity and ease of comparison with other papers 20,21,24 , we use the MSAs provided by EVmutation 20 when available, and in other cases conduct our own jackhmmer with parameters as used in the EVmutation paper (see Methods).…”

“…Regardless of the training data size, we always use 20% of the full data (not used for training or validation) as test set. In addition to these fixed sample sizes, for more direct comparison with TLmutation 24 and ESM 23 , we also test performance using train-test splits, training on 80% of all single mutants in each data set, and testing on the remaining 20% of the data set. For each sample size and the 80-20 split, we evaluate performance over 20 random seeds.…”

“…There have been recent attempts to combine the two aforementioned approaches, that is, combining the weakly-positive evolutionary data with assay-labelled data 24,26,27 . We refer to such a hybrid learning setting as weak-positive semi-supervised learning.…”

“…Shamsi et al . 24 instead “transfer” information learned from a first step of modelling evolutionary data to the supervised problem by learning a binary masking of an already-trained evolutionary Potts model using a supervised objective function. Biswas et al .…”