Scientist Stories: Euan Ashley, Assessing Clinical Utility

It is a pleasure to be here. Did you notice they just found flowing water on Mars? Did anyone notice that? They just announced apparently they found flowing water on Mars. So, on Mars. Sounds funny with my Scottish accent. Mars, that's the planet, the red planet there. It's not supposed to have any water on it. Anyway, it's great to be here off a red-eye. I might be a little spicy. But I— when we planned the meeting, we thought that one opportunity might be for me to focus a little bit on some of the technological aspects of opportunities potentially for CSER 2.0. Here are my disclosures. So this is not news to anyone in the room, but I think it bears repeating. Genomic medicine is here. It's actually here today. We're practicing it already. You're familiar with many of these stories that we've seen. There are stories like this happening all the time. It's actually here in mainstream medicine today. Both for rare disease and for non-invasive prenatal testing. So we've arrived. So that's good. The exploration, at least to the limit, can stop, and the next phase of actually moving this to mainstream can go. But a few years ago, 5 years ago, we thought, "When do we get to the point when we get a genome in every medical record?" And we're certainly not there yet. So I think some exploration into that world is what is appropriate for CSER 2.0. This man said this phrase, Satya Nadella, actually, the current CEO of Microsoft, quoted Bill Gates, said, "Most people overestimate what they can do in 1 year, but underestimate what they can do in 10." And I think this mirrors what the comment that Eric Green was making this morning about the massive progress from the start of the Human Genome Project. I think we couldn't have imagined being where we are here today. The question is, how do we get from here to there? Probably more likely from here to there. And in order to do that, we have to expand clinical utility. So just mirroring a little of what Robert just said, We have to choose use cases for going forward. We have to design, implement, thinking about what was said this morning. We have to learn from what we do and have a virtuous cycle. We have to build the evidence base for effectiveness first, build cost-effectiveness second, because you can't assess cost-effectiveness without effectiveness, and include pairs early. You may have seen some of these ICD-10 codes. I learned last week there are 9 for being attacked by a turkey, including being pecked by a turkey, being struck by a turkey, and other interaction with a turkey. Actually true. So what are the challenges in front of us if we want to get from here to there? Well, they're these. The genome is actually pretty complex. You can't call it in relation to variants if you can't see it. And the technical performance of our algorithms is actually kind of upside down. It was optimized for cohort variant discovery, in the days of GWAS, and not for the n 1 that we face on a daily basis in our clinics. This is the sort of slide that keeps me awake at night, or rather, waking in cold sweat. Repeats make up, say, 56% of the genome. If you are sequencing with short reads, you do not know where those short reads came from, and particularly in relation to some diseases that are caused by repeats, but also just in general, knowing where short reads come from. There's also a lot of paralogous sequence. Segmental duplications, gene families, pseudogenes— there's probably 8,000 of those, and all associated with varying constraints, so they will vary differently. This new data from Evan Eichler with the Hydatidiform mole, looking at— with the PacBio long-read sequencing, gives us an insight into long reads and structural variation that we just haven't been able to see yet. And moving forward, we have to start to incorporate that in when we think about the technical accuracy of the genome sequencing we're doing. For clinical medicine. So coverage is one challenge here. This is from the ExAC server. It's been mentioned many times today. This is KCNH2, which is one of the long QT genes, which long QT came up just earlier in Bob's talk. You'd think, then, the coverage should be pretty even. Of course, it's not. This is not, again, news to anyone. But look at the exon towards the end. You think that might be an important exon to cover properly, especially since if you look in ClinVar, there are 19 pathogenic or likely pathogenic variants in that exon, but you can't call those if you can't see them. We think whole genome is better, but in fact, it's probably not a lot better at 35x. This is the percentage of the gene not covered for the 56 ACMG genes. You can see— and remember that in the days of Sanger sequencing, you would not release a clinical report if one base pair of the coding sequence was not called. You just wouldn't release it. And although we have made enormous strides and many discoveries that wouldn't be possible otherwise, we have reached a point where we accept 10x coverage of 90% of the gene and call it a day. I don't think we should, and we shouldn't for CESAR 2.0. We're good at calling certain classes of variation, single nucleotide variants, the concordance between two platforms, Complete Genomics and Illumina, shown here on the left in blue, 99% concordance is pretty good. But the concordance for small indels, arguably a more important variation of— because it's more disruptive to the open reading frame, is only 55%. We should be able to do better than that. In fact, I think you'd probably say we are better at calling single nucleotide variants that are overall less likely to cause disease than those that are much more disruptive, such as structural variants and indels. And I think that's only because we haven't paid attention to that. I think there is nothing specific about calling insertions and deletions that is particularly any harder than single nucleotide variants, although we could certainly have a long discussion about that. So what is the answer? Well, and one of the questions I was asked to address is, does it start with an E or a G? You know, is the future an exome world or a genome world? And so I think— snapshot, spoiler alert— probably somewhere in between. Get coverage. This is a picture from my native city of Glasgow, where it rains 300 days out of 365 a year. We were so concerned about the coverage issue that actually we started a company, and, you know, I'm showing you a graph from this company, but actually many other sequencing sites have now started to fill in the holes in exome sequencing, and this is good, because you can see if you know where the hole is, you can go cover it more. Particularly the first exon is a little harder because you need to do smarter things with the chemistry to really cover first exons properly, but if we know where the problem is, we can shine a light on it, be transparent about it, we can fix it. If you can do it for one gene, you can do it for all the genes as shown here. And then here's for the 56 ACMGs. So this is a graph a bit like the one I showed in two or three slides ago, and this is a few different exome providers and genome at the bottom. So this is to the question of, should we order an exome or a genome? And so you can see now we're doing base pairs along the x-axis, so this is the number of base pairs that are not callable at 20x. Q30 bases. And you can see the exomes actually do pretty well for these 56 genomes, the new augmented exomes that you can either buy off the shelf or order from a provider. But the 35x whole genome doesn't do quite as well, even though we're moving theoretically to a genome world, and that's for 10-fold the amount of more data, not necessarily 10 times the amount of pricing, depending on where you go. You'd think that the genomes would do better for inclusion of UTRs. That's one of the extra things you get from the whole genome. A little better, but even the newer chemistry shown here, the HiSeqX chemistry on the left, is actually worse than the HiSeq chemistry from prior. And so I think having these kind of metrics, and in the Undiagnosed Diseases Network, which is where I mentioned, of course, this morning, and I'll shout out again at the end, we've actually been developing these metrics further so that when new sequencing chemistry happens and the sequencing providers provide new approaches, then we can pull the same metric out and apply it. So that we can see what changes. Currently in our report at the Stanford Clinical Genome Service, we actually show the number of base pairs that are not callable per gene. So knowing the enemy is important, and bringing the groups together around the communities. National Institute of Standards and Technology has a Genome in a Bottle consortium, as you know. They actually have an outpost thing at Stanford, and so we see them on West Coast time not uncommonly. You've heard of some of the dollars from the Precision Medicine Initiative going towards the FDA for PrecisionFDA and a new arrangement where there's a collaboration with DNA Nexus and, most recently, Stanford, with the first meeting just last week. So one of the things that the NISC Consortium has done is talk about areas of the genome that can be called with high confidence and those that can be called with lesser confidence. So it might be important to know that when we're calling these, and so shown here is the number of bases from the ACMG genes that are in the high-confidence regions. Again, you would hope that they would all be up at number 1, 100%. In reality, it's much lower than that, and that's just to point out that there are areas of concern that we can approach. If we can see them, shine a light on them, we can potentially find a solution. Most recently, just 2 weeks ago, from the NIST Consortium, a new trio from the PGP has been put out with long read sequencing of the trio from multiple different platforms, and we can start to do an analysis, this one done just in the last 24 hours, from that. Okay, one— a couple more points to finish. The strength of family analysis, again, in the Undiagnosed Diseases Network, we are doing 2.2 extra family members in addition to the proband, and the reason is shown here. This is the number of genes or the number of variants for a compound heterozygosity model for a family-based analysis of a neurodegenerative developmental disease, you can see that if you only have the singleton, you're talking about 400 genes and 1,400 variants. That's a lot of time for manual curation. But by the time you can include inheritance state analysis of the trio or the quad, you can get that down to something that's actually manageable. So, in terms of designing better algorithms, I mentioned that insertions and deletions were particularly problematic. Small ones were not too bad. There's a sensitivity on the Y indel length. On the X for haplotype caller unified genotyper from GATK. As you can see, the sensitivity drops off markedly after 2 to 3 base pairs. So there are groups now, such as this group in Oxford, starting to think about how to improve that. I think, again, if we just shine a light on it, we can start to develop better tools. So this is in closing. I just want to say that I think it's possible that we can get to where we were before with Sanger sequencing, where we have 100% coverage of the genes that we're interested in. But then what is the best test? Is it an exome where we know that those coding bases are well-covered, or a genome which we can use better when we can do structural variant analysis and others? And really, going forward, the best test might be one that includes some whole genome coverage, but is actually augmented around the genes so that we can actually know that we can call every coding base pair. To be cost-effective, it probably will have to include some augmentation of the genes because Currently, it looks like we might have to get to a 50x or 75x whole genome in order to get the coverage that we would really want. Also, not really mentioned here, but this would allow higher sensitivity for calling mosaics. So, finally, just this slide about what I think about the current and future landscape of genomic medicine. We talked about what's already here, which is the box on the left, circulating cell-free DNA, of course, on the right. Cancer, I think, is coming. Pharmacogenomics has been out there, and I think many of us are disappointed that it's not more mainstream than it is now, despite great work from many people, including this room. But it's coming, and I think there'll be a tipping point for that, and it'll suddenly just turn. Infectious disease and complex disease are also coming. Final slide, a shout out for the Undiagnosed Diseases Network. I've done it already, mentioning it twice, but the gateway opened last week, and so we're happy to be, to be out and ready to to do business. All right, thank you for having me.