Link prediction is a fundamental problem in the field of graph mining. The aim of link prediction is to infer/discover unobserved links in graphs. Link prediction in biological graphs is highly challenging. There exist many similarity-based methods in the literature for link prediction. These methods compete for victory in graphs from various domains. Unfortunately, they are efficient only in some specific graphs, and no one wins in all graphs. In this paper, we study some well-known similarity-based methods and consider them as independent features to define a feature set. The feature set is then used to train traditional supervised learning methods for link prediction in biological graphs. We evaluate the methods on ten biological graphs from different organisms. Experimental results show that the similarity-based methods collaboratively improve prediction performance, and are even comparable to high-performing embedding-based methods in some biological graphs. We compute the importance score of similarity-based features in order to explain the leading features in a graph.