Bibliography from Literature Review:

(Note: references include sources cited in this document as well as other journal articles read prior to the start of this project)

Ali, A. A., Almukhtar, S. E., Abd, K. H., Saleem, Z. S. M., Sharif, D. A., & Hughson, M. D. (2021). The causes and frequency of kidney allograft failure in a low-resource setting: Observational data from Iraqi Kurdistan. BMC Nephrology, 22(1), 272. https://doi.org/10.1186/s12882-021-02486-9

Bicalho, P. R., Requião-Moura, L. R., Arruda, É. F., Chinen, R., Mello, L., Bertocchi, A. P. F., Lamkowski Naka, E., Tonato, E. J., & Pacheco-Silva, A. (2019). Long-term outcomes among kidney transplant recipients and after graft failure: A single-center cohort study in brazil. BioMed Research International, 2019, 1–10. https://doi.org/10.1155/2019/7105084

Continuous distribution. (n.d.). Retrieved December 20, 2021, from https://optn.transplant.hrsa.gov/policies-bylaws/a-closer-look/continuous-distribution/

Davis, S., & Mohan, S. (2021). Managing patients with failing kidney allograft: Many questions remain. Clinical Journal of the American Society of Nephrology, CJN.14620920. https://doi.org/10.2215/CJN.14620920

Fiorentino, M., Gallo, P., Giliberti, M., Colucci, V., Schena, A., Stallone, G., Gesualdo, L., & Castellano, G. (2021). Management of patients with a failed kidney transplant: What should we do? Clinical Kidney Journal, 14(1), 98–106. https://doi.org/10.1093/ckj/sfaa094

Gómez, E. J., Jungmann, S., & Lima, A. S. (2018). Resource allocations and disparities in the Brazilian health care system: Insights from organ transplantation services. BMC Health Services Research, 18(1), 90. https://doi.org/10.1186/s12913-018-2851-1

Kamoun, M., McCullough, K. P., Maiers, M., Fernandez Vina, M. A., Li, H., Teal, V., Leichtman, A. B., & Merion, R. M. (2017). HLA amino acid polymorphisms and kidney allograft survival. Transplantation, 101(5), e170–e177. https://doi.org/10.1097/TP.0000000000001670

Lee, S., Abecasis, G. R., Boehnke, M., & Lin, X. (2014). Rare-variant association analysis: Study designs and statistical tests. The American Journal of Human Genetics, 95(1), 5–23. https://doi.org/10.1016/j.ajhg.2014.06.009

Lee, S., Emond, M. J., Bamshad, M. J., Barnes, K. C., Rieder, M. J., Nickerson, D. A., Christiani, D. C., Wurfel, M. M., & Lin, X. (2012). Optimal unified approach for rare-variant association testing with application to small-sample case-control whole-exome sequencing studies. The American Journal of Human Genetics, 91(2), 224–237. https://doi.org/10.1016/j.ajhg.2012.06.007

Li, B., & Leal, S. M. (2008). Methods for detecting associations with rare variants for common diseases: Application to analysis of sequence data. The American Journal of Human Genetics, 83(3), 311–321. https://doi.org/10.1016/j.ajhg.2008.06.024

Liu, Y., Huang, J., Urbanowicz, R. J., Chen, K., Manduchi, E., Greene, C. S., Moore, J. H., Scheet, P., & Chen, Y. (2020). Embracing study heterogeneity for finding genetic interactions in large-scale research consortia. Genetic Epidemiology, 44(1), 52–66. https://doi.org/10.1002/gepi.22262

Madsen, B. E., & Browning, S. R. (2009). A groupwise association test for rare mutations using a weighted sum statistic. PLoS Genetics, 5(2), e1000384. https://doi.org/10.1371/journal.pgen.1000384

Moore, C. B., Wallace, J. R., Frase, A. T., Pendergrass, S. A., & Ritchie, M. D. (2013). BioBin: A bioinformatics tool for automating the binning of rare variants using publicly available biological knowledge. BMC Medical Genomics, 6(2), S6. https://doi.org/10.1186/1755-8794-6-S2-S6

Morgenthaler, S., & Thilly, W. G. (2007). A strategy to discover genes that carry multi-allelic or mono-allelic risk for common diseases: A cohort allelic sums test (Cast). Mutation Research/Fundamental and Molecular Mechanisms of Mutagenesis, 615(1–2), 28–56. https://doi.org/10.1016/j.mrfmmm.2006.09.003

Pritchard, J. K. (2001). Are rare variants responsible for susceptibility to complex diseases? The American Journal of Human Genetics, 69(1), 124–137. https://doi.org/10.1086/321272

Reich, D. E., & Lander, E. S. (2001). On the allelic spectrum of human disease. Trends in Genetics, 17(9), 502–510. https://doi.org/10.1016/S0168-9525(01)02410-6

Van Loon, E., Bernards, J., Van Craenenbroeck, A. H., & Naesens, M. (2020). The causes of kidney allograft failure: More than alloimmunity. A viewpoint article. Transplantation, 104(2), e46–e56. https://doi.org/10.1097/TP.0000000000003012

Zhang, X., Basile, A. O., Pendergrass, S. A., & Ritchie, M. D. (2019). Real world scenarios in rare variant association analysis: The impact of imbalance and sample size on the power in silico. BMC Bioinformatics, 20(1), 46. https://doi.org/10.1186/s12859-018-2591-6

Research Plan/Project Summary

Initial plan completed on 4/1/21; Addendum completed 12/20/21 (see end of document)

RARE: Machine Learning Approach for Binning Rare Variant Features to Detect Association with Disease

Student Researcher: Satvik Dasariraju

Adult Sponsor: Dr. Ryan J. Urbanowicz

Rationale:

In numerous biomedical fields such as genetics, features with rare variant states pose challenges to analysis. While genome-wide association studies have demonstrated the role of common genetic variants in disease etiology, the effects of rare genetic variants (defined as having a minor allele frequency below 0.05) remain largely unexplained (Pritchard, 2001; Reich & Lander, 2001; Lee et al., 2014). Rare genetic variants are hypothesized to hold the key to explaining missing heritability in complex diseases, yet traditional univariate analysis methods are underpowered to detect rare variant association due to the low frequency of rare genetic variants (Li & Leal, 2008). Numerous alternatives have been presented in the literature, including the cohort allelic sum test (Morgenthaler & Thilly, 2007), nonparametric weighted sum test (Madsen & Browning, 2009), sequence kernel association test (Lee et al., 2012), and BioBin (Moore et al., 2013); these methods aim to overcome the challenges of rare variant association analysis by binning (i.e., grouping) rare variant features. Yet, all previous rare variant binning methods depend on some form of expert knowledge for initialization. Previous methods also have not considered interactions between features and/or interactions between features and bins despite epistasis and other forms of feature interactions being recognized as contributors to complex disease heritability (Liu et al., 2020). Finally, previous methods focused exclusively on genetics without developing algorithms that are versatile for a wide range of applications on biomedical data. 

Kidney transplant failure is a serious condition that poses a high burden of psychologic and medical morbidity, affecting about one in five kidney transplant recipients (Davis & Mohan, 2021). Transplantation failure poses an especially high burden on third world countries like Brazil and Iraq (Ali et al., 2016; Gómez et al., 2018). There is an urgent need to better understand the genetic and molecular mechanisms of transplantation failure. In kidney transplantation, features with rare variant states are particularly critical given the importance of donor-recipient HLA amino acid mismatches in kidney failure and the extremely low frequency of donor-recipient mismatches at many amino acid positions. Accordingly, an algorithm capable of binning donor-recipient HLA amino acid mismatches to detect association to kidney graft failure could elucidate relationships between HLA amino acids, overcome difficulties posed by the low frequency and multicollinearity of HLA amino acids, aid in prediction of kidney graft failure, and guide clinical and organizational efforts such as development of a new calculated panel reactive antibody test for HLA sensitization and the Organ Procurement and Transplantation Network’s new continuous distribution initiative (Continuous Distribution). 

This research plan proposes an evolutionary algorithm (a type of machine learning strategy) that constructs bins (i.e. groups) of features with rare variants and iteratively modifies bins to reach the highest possible association with outcome. The algorithm will be tested with simulation studies and applied to bin HLA amino acid mismatches to detect association to kidney graft failure.

Research Questions for Objective 1:

• What is the capability and efficiency of the proposed evolutionary algorithm to construct bins with univariate association to outcome?
• What is the capability and efficiency of the proposed evolutionary algorithm to construct bins with epistatic interactions predictive of outcome?
• What are the best parameters for the proposed evolutionary algorithm for binning rare variant features?
• How does expert knowledge input make the algorithm more reliable and/or efficient?

Research Questions for Objective 2:

• What is the capability and efficiency of a machine learning (specifically a genetic/evolutionary algorithm) to optimize bins of HLA amino acids to detect association of amino acid mismatches to kidney graft failure?
• At which locus of HLA amino acids (A, B, C, DR, or DQ) are mismatches most predictive of kidney graft failure?
• How do the proposed algorithm’s bins compare to expert knowledge sequence feature variant type bins of amino acid positions?

Hypothesis for Objective 1:

If an evolutionary algorithm for constructing bins of rare variant features is developed, then it will be able to reliably construct bins with univariate and epistatic effects (the latter of which will be achieved by scoring bins with methods sensitive to interactions), thus overcoming the limitations of previous rare variant association analysis methods by (1) being able to detect interactions, (2) not requiring but allowing expert knowledge input for initialization of bins, and (3) considering groups of features additively to reach the statistical power required to analyze rare variant features. 

Hypothesis for Objective 2:

If the proposed algorithm is applied to HLA amino acid mismatches and kidney graft failure, then it will be able to successfully construct bins of HLA amino acid mismatches with higher association to kidney graft failure compared to individual amino acid mismatches or expert knowledge sequence feature variant type bins because the algorithm adopts a data-driven approach and eliminates assumptions about relationships between amino acid positions across HLA loci. 

Engineering Goals:

• Objective 1: Development and validation of an evolutionary algorithm capable of constructing bins of rare variant features with optimal association to outcome
  • Sub-objective 1-1: The algorithm should be able to construct bins with univariate association to outcome
  • Sub-objective 1-2: The algorithm should be able to construct bins with epistatic interactions that are predictive of outcome
  • Sub-objective 1-3: Development of a synthetic data simulator to validate sub-objectives 1-1 and 1-2
  • Sub-objective 1-4: The algorithm should be able to either start/initialize without expert knowledge or use expert knowledge bins as input for initialization
• Objective 2: Application of the algorithm developed in objective 1 to construct bins of HLA amino acid positions with association to kidney graft failure
  • Sub-objective 2-1: Compare bins constructed by algorithm to expert knowledge sequence feature variant type bins
  • Sub-objective 2-2: Use algorithm to bin amino acids within each HLA locus separately and compare bins from each locus (especially DR vs. DQ locus)

Expected Outcomes:

Objective 1: An evolutionary algorithm (a type of machine learning algorithm) that can construct bins with both univariate and epistatic effects with above 90% accuracy when tested against simulated datasets with known/predetermined optimal bins

Objective 2: When applied to HLA amino acid positions and kidney graft failure, the algorithm should construct bins with higher association to graft failure compared to individual amino acid positions and sequence feature variant type bins

Materials:

1. Personal computer (PC)
2. Connection to computing cluster at Penn Medicine
3. Donor-recipient amino acid position mismatch and graft failure data from Scientific Registry of Transplant Recipients, obtained through collaborator Dr. Malek Kamoun

(Note: all the data is anonymous and de-identified)

4. Jupyter notebooks for documentation and coding in the Python programming language
5. Various open source Python packages including sklearn, numpy, and pandas.

Procedures:

Steps of Proposed Algorithm:

1. Initialize bins with either random grouping of features or expert knowledge input
2. Repeated evolutionary cycles (mimics natural selection process)
   1. Bin fitness evaluation with chi-square or MultiSURF
   2. Preserve elite bins for next generation of bins, remove poor bins
   3. Genetic operations to choose pairs of parent bins and then create offspring bins through crossover of parent bin pair and mutation
   4. Add offspring to next generation of bins
3. Final bin evaluation and summary

Objective 1:

1. Develop and code the proposed evolutionary algorithm for rare variant feature binning
2. Develop a rare variant data simulator to test algorithm’s ability to construct univariate bins
3. Develop a rare variant data simulator to test algorithm’s ability to construct epistatic bins
4. Run 30 trials of each of the following experiments (to validate the algorithm); for each experiment, one of the rare variant data simulators coded in the above steps will be used, so the features that belong in the optimal bin will be known, and the accuracy of the bin constructed by the algorithm can be measured against the optimal bin
   1. Test algorithm with MultiSURF scoring on univariate data simulator with 0 noise
   2. Test algorithm with MultiSURF scoring on univariate data simulator with 0.05 noise
   3. Test algorithm with MultiSURF scoring on univariate data simulator with 0.1 noise
   4. Test algorithm with MultiSURF scoring on univariate data simulator with 0.5 noise (negative control)
   5. Test algorithm with Chi-square scoring on univariate data simulator with 0 noise
   6. Test algorithm with MultiSURF scoring on univariate data simulator with 0 noise with bins forced to maintain fixed bin size
   7. Test algorithm with MultiSURF scoring on univariate data simulator with 0 noise, expert knowledge input for initialization
   8. Test algorithm with Chi-square scoring on epistatic data simulator with 0 noise
   9. Test algorithm with MultiSURF scoring on epistatic data simulator with 0 noise

Objective 2:

1. Apply algorithm to construct bins of HLA amino acid mismatches with association to kidney graft failure, where all amino acids are eligible for inclusion in bins
2. Repeat step 1, but run the algorithm five separate times (1 for each locus of HLA: A, B, DR, DQ)
3. Evaluate bins in comparison to expert knowledge sequence feature variant type bins

Data Analysis:

Due to the multicomponent nature of the study, there will be numerous sets of variables and data analyses.

Objective 1 Part A: Testing Algorithm on Univariate Simulated Data with Different Levels of Noise

Independent Variable: Level of noise in data (0, 0.05, 0.1, 0.5)

Dependent Variable: Performance of  algorithm, measured by percentage of correct features averaged across 30 replicate trials

Control Variables: 

• Parameters of algorithm 
• Characteristics of dataset (number of correct features, number of instances, etc.)
• The same computer and computing cluster will be used for all runs in this project

Objective 1 Part B: Testing Algorithm on Univariate Simulated Data with Different Configurations

Independent Variable: Configurations of algorithm

Dependent Variable: Performance of algorithm, measured by percentage of correct features averaged across 30 replicate trials

Control Variables: 

• Characteristics of dataset (number of correct features, number of instances, etc.)
• The same computer and computing cluster will be used for all runs in this project

Objective 1 Part C: Testing Algorithm on Univariate Simulated Data with Different Configurations

Independent Variable: Initialization (random vs. expert knowledge)

Dependent Variable: Performance of algorithm, measured by percentage of correct features averaged across 30 replicate trials

Control Variables: 

• Characteristics of dataset (number of correct features, number of instances, etc.)
• The same computer and computing cluster will be used for all runs in this project
• Parameters of algorithm aside from initialization

Objective 1 Part D: Testing Algorithm on Simulated Data with Epistasis

Independent Variable: Metric used for scoring bins in cycles of algorithm (MultiSURF vs. chi-square)

Dependent Variable: Performance of algorithm, measured by percentage of correct features averaged across 30 replicate trials

Control Variables: 

• Characteristics of dataset (number of correct features, number of instances, etc.)
• The same computer and computing cluster will be used for all runs in this project

Objective 2 Part A: Comparing Algorithm’s Bins vs. Expert Knowledge Bins

Independent Variable: Whether bins were generated from algorithm or sequence feature variant type categories

Dependent Variable: Association between bin and kidney graft failure, measured by chi-square

Control Variables: 

• The same computer and computing cluster will be used for all runs in this project

Objective 2 Part B: Comparing Bins from Different HLA Loci

Independent Variable: HLA locus (A, B, C, DR, DQ)

Dependent Variable: Association between bin and kidney graft failure, measured by chi-square

Control Variables: 

• The same computer and computing cluster will be used for all runs in this project

Risk and Safety Evaluation:

There are no major hazards associated with the conducting of the project because all of the set up and experimentation will be done on a computer. Potential risks associated with using a computer are the common hazards of the internet. I will protect myself from these hazards by only downloading data and journal articles from secure sources, not communicating with strangers, and not revealing any personal information on the internet. 

Addendum:

No major changes were made to the procedure or data analysis. The sole addition was the calculation of statistics to measure association between individual amino acid positions and kidney graft failure (which served as an additional point of comparison to demonstrate the utility of bins constructed by the algorithm). The name of the proposed algorithm was finalized as RARE (Relevant Association Rare-variant-bin Evolver). 

It is worth noting that I am the first author of a peer-reviewed research paper based on the algorithm described in this project (i.e., objective 1 of this research plan). The manuscript was accepted to and presented at the Genetic and Evolutionary Computation Conference as a workshop paper has been published in GECCO '21: Proceedings of the Genetic and Evolutionary Computation Conference (link: https://doi.org/10.1145/3449726.3463174). The application of the algorithm, also described in this project, resulted in an abstract of which I am first author. The abstract was accepted to and presented at the American Society for Histocompatibility and Immunogenetics 47th Annual Meeting and also published in Human Immunology. The code has been made open source at https://github.com/UrbsLab/RARE.

Abstract for Project:

Features with rare states, such as rare genetic variants and low frequency amino acid mismatches, pose a challenge for statistical and machine learning analyses due to limited detection power and uncertainty surrounding their role in predicting outcomes such as disease phenotype. Bins of rare variant features hold the potential to increase association detection power. However, previous binning approaches were wholly dependent on prior-knowledge assumptions, instead of data-driven techniques, and ignored multivariate interactions. Previous binning methods were also confined to genetics despite the pertinence of rare variant features across biomedicine and other scientific fields. This study develops the Relevant Association Rare-variant-bin Evolver (RARE), the first evolutionary algorithm for automatically constructing and evaluating rare variant bins with either univariate or epistatic associations, offering flexibility to initialize bins both with or without expert knowledge. RARE’s ability to correctly bin simulated rare-variant associations is evaluated over a variety of algorithmic and dataset scenarios. Specifically, this study examines (1) ability to detect rare variant bins of univariate effect (with varying levels of noise), (2) using fixed vs. adaptable bins sizes, (3) employing expert knowledge to initialize bins, and (4) ability to detect rare variant bins interacting with a separate common variant. Results demonstrating the feasibility and efficacy of RARE are presented alongside an application of the algorithm in constructing bins of donor-recipient HLA amino acid positions associated with kidney graft failure in order to elucidate the role of HLA mismatches in transplantation.