Excited to announce that I’ve been working with some fantastic computational biologists (that I met at ISMB2018) on a “10 simple rules” style paper for creating and promoting effective bioinformatics collaborations with wet-lab biologists. We will leverage our many years of combined bioinformatics core experience to create these “10 simple rules.”
We will touch on:
–experiment planning and design.
–data management plans, data QC, and record keeping.
–avoiding batch effects and contamination.
–managing expectations and developing clear communications.
–handling low-quality data when things do go wrong.
Hope to have this out by the end of 2018…watch this space!
Welcome to the second part of this post series on building artificial neural network models for copy number classification. In the first part, I described the problem with interpreting copy-ratio plots to find clinically-relevant CNV events. The data from targeted capture deep sequencing are noisy and biased, and finding clinically-relevant genotypes in genes that have CNVs requires the analyst to visualize the CNV event and assign a classification on the basis of experience and expert knowledge.
The LASSO model
Once my training data were in place (see part 1), I used a multiple linear regression LASSO model as a machine-learning benchmark. I did this to determine whether a more powerful neural network model would be warranted. The LASSO model uses an “L1” prior to perform feature selection, setting some coefficients to zero as warranted by the data. There is ample precedent for applying this type of model in bioinformatics settings where the goal is maximize predictive power without overfitting.
I fit the LASSO to the data, with 33% held out for validation. The best fit was obtained with the alpha parameter set to 0.001. k-fold cross validation (where k=10 and alpha=0.001) yielded an accuracy of 76%. These results are surprisingly good, given the complexity of the CNV signals in the noisy data. Unfortunately, 76% accuracy is simply not good enough for an automated method that will be used to predict genotypes in clinical data.
The ANN model
Next, I decided to construct an artificial neural network model. My goal was to keep the model as simple as possible, while reaching a very high classification accuracy needed for clinical work. To that end I constructed a one hidden-layer model with 19 input nodes corresponding to the 19 copy-ratio probes in the CNV data. The output layer contained five nodes, corresponding to the five classes of defined CNV event or other event (for example, a very distinct sequencing artifact that kept appearing in the data):
In between the input and output layers I constructed a 10-node hidden layer. A one hidden-layer neural network is the simplest form of the ANN model, and I tried to keep the number of hidden-layer nodes to a minimum as well. Specific details about the model, hyper-parameter tuning, and the code will be available in the near future when I put a pre-print of this work on biorxiv.
Model training and cross-validation
I trained the model on the 175 sample dataset and on a 350 sample “synthetic” dataset created by adding gaussian “noise” to the real data. The results are shown below, across 250 training epochs.
When the ANN model was tested with 10-fold cross validation, the accuracy reached a level of 96.5% (+/- 5.4%). This is obviously a big improvement on the LASSO model, and reaches a level of accuracy that is good enough for clinical pipelines (with the caveat that low confidence predictions will still be checked “by hand.”)
Below, I’m showing a sample of the model output (left) and ground truth (right) from the test data. The numbers (and colors) of the boxes correspond to the model’s probability in that classification. You can see that most CNV events are called with high probability, but several (yellow boxes) are called correctly but with lower probability. One event (red box) is called incorrectly with high probability.
Conclusions and caveats
Going into this project, I had no idea if the ANN model would be able to make predictions on the basis of so few examples in the training set. The classic examples you see about ANN/CNN models rely on handwriting training sets with 10,000 or more images. So I was surprised when the model did very well with extremely limited training data. Since this method was developed for a clinical pipeline, it can be improved as the pipeline generates new training data with each new patient sample. We would need many thousands of samples through our “legacy” pipeline to see enough examples of the rare star allele events in CYP2D6 that we could then classify them. That is why I limited my CNV calling to three star alleles.
The low confidence, true positive predictions concern me less than the high confidence false negative. Missing a real CNV that has impact on CYP2D6 function and therefore clinical relevance is very dangerous. This can lead to incorrect prescribing and adverse drug reactions for the patient. I really want to understand why the method makes predictions like this, and how to fix it. Unfortunately, I believe it will require a lot more training data to solve this problem and that is something I lack.
My goals for this project now are 1) to publish a preprint on biorxiv describing this work and 2) to obtain some additional training/test datasets. Because our pharmacogenomics test is not generating the kind of volume we expected, I may have to look around for another gene with clinically-relevant CNV events to test this method further. For example, we do have an NGS-based test of hearing and deafness genes with thousands of validated patient samples. One gene, STRC, has relevant CNVs that are complex and require analyst visualization to detect. This may be a good system for follow up refinement of this type of model.
tl;dr: It’s only 2X faster than kallisto; a marginal, rather than “ultra” improvement. If you’re already using kallisto in your pipelines there’s no reason to switch. Now onto the blog post:
Another RNA-seq read aligner
In July, a paper appeared in BMC Bioinformatics: “Matataki: an ultrafast mRNA quantification method for large-scale reanalysis of RNA-Seq data.” The title seems to imply that Matataki is much faster than anything else out there. After all, what does ‘ultrafast’ mean unless it’s in comparison to the next fastest method? The authors claim it also solves a unique problem of large-scale RNA-seq reanalysis that can’t be solved any other way. Is this true? Let’s dive into the details and find out.
The authors point out in the introduction that the amount of publicly available RNA-seq data is increasing rapidly and that many would like to do large-scale reanalysis of these datasets. Reanalysis (including realignment) is necessary because of the large batch effects of using different methods for alignment and quantification.
Although these alignment-free methods such as Sailfish, Kallisto, and RNA-Skim are much faster than the alignment-based methods, the recent accumulation of large-scale sequence data requires development of an even faster method for data management and reanalysis.
Counting unique k-mers
The method itself eschews traditional alignment for a k-mer based read counting approach. This approach takes two steps: build an index and quantify the reads against it. Building the index requires the algorithm to search for all k-mers unique to a gene and also present in all isoforms of that gene (to avoid effects of differential expression of isoforms). Quantification proceeds by matching the k-mers from a read with the unique k-mers in the index. If there is only one indexed gene match from several k-mers (with a minimum cutoff set by the user) that read is assigned to the indexed gene. If there is more than one match to several genes, the read is discarded.
If this sounds a lot like how kallisto operates under the hood, I would agree with you as would the authors:
Similar to Kallisto, our method uses k-mers that appear only once in a gene and quantifies expression from the number of unique k-mers.
Of course, kallisto creates a transcriptome de bruijn graph (tDBG) and uses compatibility classes to decide whether to skip k-mers while hashing a read. Matataki does not use a tDBG and simply skips along a read using a fixed interval set by the user. The code also skips the Expectation-Maximization (EM) step found in kallisto and focuses only on gene-level quantification. These two novel approaches are what set Matataki apart from other pseudo-alignment algorithms, according to the authors.
But does it work?
Well, it appears to. The authors start by reporting the performance of the method on simulated data using optimized parameters (I’ll discuss this below). In Figure 2, they describe the parameter tuning that led to those choices. For parameter tuning the performance of Matataki on real data, the TPM values from an eXpress run on sample ERR188125 were taken as “ground truth.” The choice of ‘eXpress’ as their benchmark for real-world dataset quantification is odd, given that eXpress is defunct and has been for some time. It also takes orders of magnitude longer to complete than a method like sailfish or kallisto.
Fig 2. Comparison of TPM when k was varied. The x-axis shows the TPM values of eXpress, the y-axis shows the TPM values of our method, and the color indicates the indexed k-mer coverage of each gene when changing k from 16 to 56 with a step of 8
Going back to Figure 1 (below) and the performance on simulated data, we can observe accuracy similar to that of kallisto, RSEM, sailfish and other methods (note that by using a scale from 0.93 to 0.99, the differences between the methods seem exaggerated). The one simulated sample was created by RSEM from the real sample plus three other related samples. So perhaps it’s not surprising that RSEM does the best in this comparison overall.
I think it’s important to remember as well that Matataki is being run on this simulated data with optimized parameters for this sample while the other methods are run with “default” parameters. We don’t know how Figure 1a and 1b would change for other samples, other organisms, and other sequencing types (PE vs. SE, long vs. short, etc…) It would be enlightening to find out.
Figure 1. Summary of the results using simulation data. a Spearman correlation coefficient with the expected expression and estimated expression values using each method. “Matataki” indicates the results of the proposed method, and “MatatakiSubset” indicates the results of the proposed method without uncovered genes. To compare the gene-level expression profile and transcript-level expression profile, the sum of TPM by each gene was calculated. b Means of absolute difference from the expected expression levels.
So how fast is ‘ultrafast’ anyway?
We finally get to the promised speed improvements near the end of the manuscript, only to find that the speedup relative to kallisto, a method that has been available publicly since at least 2015, is only 2-fold. Yep, that’s it. It’s only twice as fast. This is an incremental improvement, but not a paradigm shift. In fact, it shows how the leap from alignment to pseudo-alignment was in fact a paradigm shift when kallisto first came out. It’s been three years and nothing has substantially improved on it.
Fig 3. CPU time comparison.
Limitations?
Finally, the authors end the paper with a section on ‘expected use-cases and limitations’ where they say that the method is only suitable for large-scale reanalysis and not for normal, turn-the-crank RNA-seq.
Nevertheless, Matataki is not suitable for common RNA-Seq purposes because other methods are sufficiently fast and provide better accuracy. For example, a single nucleotide substitution has larger effects in Matataki than in other methods, because even a single point substitution changes the k-mer for 2 k − 1 bases, which ultimately affects the number of k-mers in a transcript and calculation of the TPM value.
For reasons that I don’t understand, the authors promote this tool (which is only 2X faster than kallisto) as being for large-scale reanalysis, yet they performed no large-scale reanalysis to benchmark the tool. Rather, they used it to perform normal RNA-seq as a benchmark, which they go on to say not to do in the wild.
To sum up, while a 2X speedup is useful, I see no reason to abandon kallisto to pseudoalign and quantify data when doing large scale reanalysis. Unlike Matataki, kallisto is robust to read errors and it delivers transcript-level estimates that can be later summed to gene level, if you so desire. It also provides bootstrapped estimates of technical variation, which helps to understand the biological variation as well. And finally, kallisto integrates nicely with sleuth for DE testing and visualization.
In early July, I attended the ISMB 2018 meeting, a computational biology-focused meeting held by the International Society for Computational Biology (ISCB). The meeting was held in the beautiful Hyatt Regency hotel in downtown Chicago, just across the street from the river and blocks from Navy Pier and Lakeshore Drive.
ISMB 2018 was a huge meeting, with at least 1500 attendees and up to ten parallel meeting tracks (called “COSIs” in ISCB parlance) at any one time. The meeting was so big I always felt as if I was missing something good, no matter which talk I went to (except for the keynotes, where nothing else was happening).
Here is Steven Salzberg’s excellent keynote on a historical overview of finding genes in the human genome:
Obviously, at a meeting this large, one cannot explore more than a tiny fraction of the talks and posters (but you can watch many of the ISMB2018 talks on youtube now if you’re interested).
I want to briefly summarize three talks that I particularly enjoyed and found very interesting:
1) Michael Seiler, H3 biomedicine. “Selective small molecule modulation of splicing in cancer”
Up to 60% of hematologic neoplasms (CLL/AML/MDS) contain heterozygous hotspot mutations in key splicing factor genes involved with 3′ splice site recognition. One such gene is SF3B1, an RNA splicing factor, which is recurrently mutated around the HEAT repeats. The cryo-EM structure of SF3B1 revealed the mutations cluster in the premRNA-interacting region. All of the observed mutations lead to alternate 3′ splice sites being chosen during splicing; so-called “cryptic 3′ splice sites (AG)”. In particular, the K700E mutant slips the spliceosome upstream to an internal AG almost 18 nucleotides from the proper splice site.
Seiler asked the question whether one could exploit this weakness to attack cancer cells? The idea is to modulate activity of SF3B1 using small molecules to disturb the WT function. At H3 they took a natural product compound library and optimized for chemistry and binding to find candidates. One molecule, H3B-8800, showed promise. The compound was optimized toward selective induction of apoptosis in SF3B1 mutant cells in vitro. When resistance mutations occur, they are all at points of contact with the drug. At only 13 nM dose, around 1% of splicing events were affected which was enough to cause lethality to the cancer cell. RNA-seq demonstrated that it was mainly the spliceosome genes themselves that were affected by the alternate splicing in the presence of the inhibitor. This demonstrated a delicate feedback loop between correct splicing and spliceosome gene expression.
2) Curtis Huttenhower, Harvard University. “Methods for multi-omics in microbial community population studies”
Huttenhower’s group at Harvard is well known for their contributions to methods for microbiome analysis, collectively known as the “biobakery.” In this talk, Huttenhower addressed the fact that the microbiome is increasingly of interest when looking at population heath, and that chronic immune disease incidence is rising around the world over the last few decades.
Huttenhower also spent a good amount of time describing the IBD-MDB, the inflammatory bowel disease multi ‘omics database. The database contains ~200 samples that are complete with six orthogonal datatypes, including RNA-seq, metagenome, and metabolome. He talked about how they can associate bacteria enriched in IBD with covariation in the prevalence of metabolites. For example, he showed how sphingolipids, carboximidic acids, cholesteryl esters and more are enriched while lactones, beta-diketones, and others are depleted in the guts of crohns disease suffers.
3) Olga Troyanskaya, Princeton University. “ML approaches for the data-driven study of human disease”
Troyanskaya’s group at Princeton aims to develop accurate models of the complex processes underlying cellular function. In this talk, she described efforts in her lab to use machine learning methods to understand how single nucleotide changes (SNPs) in non-coding regions can affect gene regulation and expression. She is also interested in how pathways and networks change in different tissues and whether tissue-specific maps can be used to identify disease genes.
In particular, Dr. Troyanskaya described the “DeepSEA” software, which uses a convolution neural network (CNN) to attempt to predict the chromatin remodeling consequences of a SNP in a genomic context. The model is a three layer CNN that takes 1000 base pairs of sequence as input. The 1000 bp region is centered around known TF binding locations (200 bp bins). The training data consists of a length 919 vector that contains binary values for the presence of a TF binding event across the genome (1 = binds, 0 = does not bind). The output of the model is the probability that the specific sequence variant will affect TF binding at each of the 919 bins.
Schematic of the DeepSEA model, showing input data, training data, and output.
The DeepSEA model can be used to predict important sequence variants and even eQTLs for further study or to prioritize a list of known non-coding sequence variants.
As a bioinformatics applications scientist, I work in a middle ground between those who develop code for analyzing next-gen sequencing data and those who consume that analysis. The developers are often people trained in computer science, mathematics, and statistics. The consumers are often people trained in biology and medicine. There is some overlap, of course, but if you’ll allow a broad generalization, I think the two groups (developers and consumers) are separated by large cultural differences in science.
Ensemble focus vs. single-gene focus
I think one of the biggest differences from my experience is the approach to conceptualizing the results of a next-gen seq experiment (e.g., RNA-seq). People on the methods side tend to think in terms of ensembles and distributions. They are interested in how the variation observed across all 30,000 genes can be modeled and used to estimate differential expression. They are interested in concepts like shrinkage estimators, bayesian priors, and hypothesis weighting. The table of differentially expressed features is thought to have meaning mainly in the statistical analysis of pathway enrichment (another ensemble).
Conversely, biologists have a drastically different view. Often they care about a single gene or a handful of genes and how those genes vary across conditions of interest in the system that they are studying. This is a rational response to the complexity of biological systems; no one can keep the workings of hundreds of genes in mind. In order to make progress, you must focus. However, this narrow focus leads investigators to sometimes cherry pick results pertaining to their ‘pet’ genes. This can invest those results with more meaning than is warranted from a single experiment.
The gene-focus of biologists also leads to clashes with the ensemble-focus of bioinformatics software. For example, major DE analysis packages like DESeq2 do not have convenience functions to make volcano plots, even though I’ve found that those kinds of plots are the most useful and easiest to understand for biologists. Sleuth does have a volcano plotting function, but doesn’t allow for labeling genes. In my experience, however, biologists want a high-res figure with common name gene labels (not ensemble transcript IDs) that they can circle and consider for wet lab validation.
New software to address the divide?
I am hopeful about the recent release of the new “bcbioRNASeq” package from Harvard Chan school bioinformatics (developers of ‘bcbio’ RNA-seq pipeline software). It appears to be a step towards making it easier for people like me to walk on both sides of this cultural divide. The package takes all of the outputs of the ‘bcbio’ pipeline and transforms them into an accessible S4 object that can be operated on quickly and simply within R.
Most importantly, the ‘bcbioRNASeq’ module allows for improved graphics and plotting that make communication with biologists easier. For example, the new MA and volcano plots appear to be based on ‘ggplot2’ graphics and are quite pretty (and have text label options(!), not shown here):
I still need to become familiar with the ‘bcbioRNASeq’ package, but it looks quite promising. ‘pcaExplorer‘ is another R/Bioconductor package that I feel does a great job of making accessible RNA-seq reports quickly and easy. I suspect we will see continued improvements to make plots prettier and more informative, with an increasing emphasis on interactive, online plots and notebooks rather than static images.
