Function prediction for in silico protein mutagenesis using transduction and active learning

As wet lab mutagenesis is expensive and time consuming one would use instead in silico mutagenesis. Towards that end, this paper proposes a novel method, T-RF-AL, for in silico mutagenesis. The method combines the merits of transduction, random forests, and active learning, the latter driven by a criterion of maximum curiosity. The feasibility of the T-RF-AL is shown on predicting mutant activity for HIV-1 Protease (HIV-1) and Bacteriophage T4 Lysozyme (T4) datasets. The new method, incremental in nature, compares favorably against random forests using the same active learning criteria, that of maximum curiosity. The observed advantages include better prediction accuracy that is achieved faster and using less training data.