TL;DR
This paper introduces a novel supervised gene rank aggregation method that automatically weights individual gene ranking lists based on their agreement with prior biological knowledge and other lists. They leverage an Artificial Fish Swarm Algorithm (AFSA) to efficiently solve the complex optimization problem, demonstrating its superior performance in identifying informative genes compared to existing methods.
What Problem Does This Paper Address?
The explosion of high-throughput genomic and proteomic data has led to numerous experiments aiming to identify genes crucial for various diseases. To gain a more robust understanding, researchers need to aggregate findings from these multiple, often disparate, experiments. However, existing gene rank aggregation methods typically either treat all lists equally (unweighted) or assign fixed weights based on domain expertise. Both approaches have significant drawbacks: unweighted methods ignore varying list reliability, while fixed weights are difficult to determine and scale poorly to hundreds or thousands of experiments. This paper addresses the need for an objective, automated method to weight gene ranking lists, thereby improving the accuracy and biological relevance of aggregated results.
Methods
The core of this paper lies in two key innovations:
- Supervised Weighted Rank Aggregation: They propose a new method to automatically assign weights to each gene ranking list. This weighting considers two factors:
- Agreement with other ranked lists: Lists that are more consistent with the consensus of other lists are deemed more reliable. This is quantified using distance metrics like Kendall's Tau or Spearman's Footrule.
- Agreement with prior biological knowledge: This incorporates known informative (and implicitly, un-informative) genes. Lists that correctly rank known informative genes higher and un-informative genes lower are given higher weights. The weighting function is designed to penalize disagreements with both other lists and prior knowledge.
- Artificial Fish Swarm Algorithm (AFSA): The problem of minimizing the disagreement metric (the objective function) is NP-hard. To efficiently find an optimal solution, the authors employ AFSA, a swarm intelligence algorithm inspired by fish schooling behavior. AFSA is known for its ability to perform global searches, avoid local optima, and be insensitive to initial conditions, making it well-suited for this complex optimization task. The AFSA is adapted to search for the optimal ranked list of genes.
Datasets
The paper evaluates their proposed method on three types of datasets:
- Contrived Dataset: A dataset from a previous study (Lin et al., 2010) consisting of three top-40 ranked lists was used to evaluate the AFSA's optimization-searching effectiveness without relying on prior biological knowledge.
- Prostate Cancer Dataset: Five gene ranked lists from prostate cancer studies were used. This dataset incorporated prior biological knowledge by using nine genes previously identified as informative for prostate cancer.
- HIV Dataset: This dataset involved aggregating results from three gene selection methods (SAM, Student T Test, Wilcoxon signed-rank test) applied to 24 progressors, with a focus on identifying informative genes related to HIV progression.
Key Insights and Ideas
- The Power of Weighted Aggregation: Simply averaging or equally weighting gene lists misses crucial information about their individual reliability. Incorporating prior biological knowledge and inter-list agreement significantly enhances the quality of the aggregated results.
- AFSA as an Efficient Optimizer: For NP-hard problems like gene rank aggregation, traditional optimization methods can be computationally prohibitive. AFSA provides a robust and effective alternative for finding near-optimal solutions in a complex search space.
- Biological Meaning Matters: The authors emphasize that their method not only finds a statistically optimal ranking but also produces results with stronger biological significance, meaning the identified genes are more likely to be truly informative.
Key Results
- Contrived Dataset: The AFSA-based method (ASFAK for Kendall's Tau and ASFAS for Spearman's) achieved the minimal objective function values for both Kendall's Tau and Spearman's measures, outperforming other established rank aggregation algorithms like Borda's method, Markov Chain, and Cross Entropy Monte Carlo.
- Prostate Cancer Dataset: The supervised AFSA method resulted in an aggregated ranked list that had minimal disagreement with other ranked lists and included the maximal quantity of identified informative genes within the top-25 ranking. This demonstrated both statistical and biological superiority.
- HIV Dataset: The proposed method (ASFAK) showed the most commonality with the other three ranked lists and included all identified informative genes within the top-30, with a relatively low average ranking, indicating improved performance.
Limitations and Future Work
- Limitations:
- The paper notes that for the prostate cancer dataset, they did not discuss un-informative genes due to a lack of information.
- When comparing methods, they mention that it's hard to compare results when weighted methods differ, and they did not compare with other algorithms for the prostate cancer real dataset because no other automatically weighted methods were available.
- Future Work: Not explicitly detailed, but the success of AFSA in this domain suggests its potential application to other complex optimization problems in bioinformatics and related fields. The authors' focus on supervised weighting also opens avenues for exploring more sophisticated ways to incorporate diverse forms of prior biological knowledge.
Summary updated