Earth BioGenome Project: the beginning of the end or the end of the beginning?

Key points • Reference genomes are insufficient to capture the genetic diversity of a species; we need pangenomes sequenced across cohorts of individuals from each species. • Integrating phenotypic data with artificial intelligence will be crucial to ensuring that the Earth BioGenome Project generates products or results useful to society. • The long-term impact of this project will depend on a new generation of scientists trained to perform activities beyond just the generation of reference genomes. Introduction "Now this is not the end. It is not even the beginning of the end. But it is, perhaps, the end of the beginning." Winston Churchill (November 10, 1942) The Earth BioGenome Project (EBP) aims to generate high-quality reference genome sequences of all named eukaryotic species, totaling an estimated 1.67 million species of plants, animals, fungi, etc. In their lead article for Frontiers in Science (1), Blaxter et al. celebrate the successes of the past and look ahead to the future. Under Phase I, over the last 3 years, EBP affiliated projects sequenced 1,667 high-quality genomes (from a wide diversity of taxa) at an average cost of US$28,000 per species. Under Phase II, and optimistically over the next 4 years, 150,000 genomes will be sequenced at an even lower average cost. Some species will be more difficult, but they are confident that the remaining technical challenges (extremely well-detailed in their manuscript) are surmountable. Among the non-technical geopolitical challenges is their commendable goal to engage the Global South, where many critical species must be sourced. Regardless of how everything shakes out, this ambitious endeavor is destined to become a milestone in science history, much like the Human Genome Project (HGP) that preceded it by a quarter of a century. HGP-EBP parallels The HGP provided a first version of the human genome (and the genomes of a few model organisms), whereas the EBP will do likewise for all other eukaryotic species. Both projects are noted for their sense of completeness, sequencing "every" base pair in a genome versus "every" named eukaryotic species, respectively. Both are expensive, costing billions of dollars, and justified by their potential impacts not only for science but also for society at large. For someone like me whose career spans both of these eras, the parallels are especially resonant. In the heydays of the HGP from the late 1990s to early 2000s, I was a top lieutenant for Maynard Olson's Genome Center at the University of Washington, which was a part of the Department of Molecular Biotechnology (now Genome Sciences) founded by Leroy Hood. Olson and Hood were members of the special committee of the United States National Academy of Sciences that defined the original goals for the HGP (2). Hence, I was privy to many of the internal discussions behind key policy decisions, and that experience now informs my views of the EBP. As Mark Twain once said, "history does not repeat itself, but it often rhymes." Two questions often asked of all "big science" projects are: i) When is the project finished? and ii) How do we make this useful? My comments on the former are based on how the HGP made its decisions, and what happened in the years after the project was declared finished. As for the latter, there are two interpretations. For the scientists involved, it suffices that their data is foundational. But for the public that funds such projects, they understandably want products or results that are useful even for non-scientists, like new medicines or new crops or improved conservation efforts. This requires much more than genomics, but it is in the interests of the EBP (as it was for the HGP) to facilitate such activities. When is the project finished? Originally, the HGP promised to sequence every base pair to an accuracy of 10−4. As human genetic variation occurs at a rate of 10−3, this ensures that any observed differences between two sequenced genomes will be more likely due to polymorphisms than to experimental errors. An exception was made for highly repetitive heterochromatic regions like centromeres and telomeres, as they could not be cloned, amplified, sequenced, or assembled with the technologies of the time. No one objected to our ignoring such regions, as there are few genes embedded within heterochromatin. When a competition arose between the public consortium led by Francis Collins at the National Human Genome Research Institute (NHGRI) and the private effort led by Craig Venter at Celera Genomics, there was a heated debate within the public consortium about abandoning the original quality standards. According to Lee Rowen at the Institute for Systems Biology (ISB), which side people supported depended on their outlook in life. Some were astonished that we raised $3 billion from the government. Those people took a "cathedral building" approach, and were determined to do it right, because there would never be a second chance. Others were prodigious fund raisers, and they argued that the data only had to be good enough to justify the money spent. Being optimists by nature, they also argued that regions too expensive or impossible to sequence with existing technologies could be deferred to future projects, when presumed technological advances would make such regions tractable. History has largely vindicated the techno-optimists perspective. The absolutist goal where every chromosome is assembled into one contiguous sequence was never met, by the public or the private HGP. The first telomere-to-telomere (T2T) assembly was not reported until 2022, two decades later (3). What made this possible was the invention of single molecule sequencing by Pacific BioSciences of California and Oxford Nanopore Technologies, with much longer read lengths (albeit moderate accuracies) that could sequence through heterochromatin. Another question often asked of the HGP was: whose genome did we sequence? No one genome can ever capture the panoply of sequences found across the human population, and indeed, a Human Pangenome Reference Consortium is now generating haplotype-phased diploid assemblies from a cohort of genetically-diverse individuals (4). This too is a consequence of technological advances. Without the million-fold reduction in costs of sequencing, largely driven by Illumina, this would have been unthinkable. Circling back to the EBP, the ultimate goal should clearly be a pangenome for all named eukaryotic species; but the costs are prohibitive with current technologies. EBP quality standards (5) (1 Mb contig N50, chromosome-scale scaffolds, and 10−4 base pair accuracy) are basically an acknowledgment of reality. Moreover, some species will be orders of magnitude more expensive, just because of genome size variation, and there might be a lot of reluctance to sequence these to the same "high" standards. People will be tempted to compromise because history has proven that even a less-than-perfect genome is valuable. And yet, there is another dimension to the EBP not faced by the HGP: the impossibility of sampling some species (either physically or ethically). This should be acceptable, much as it How do we make this useful? Multi-genome datasets are typically published with an evolutionary narrative; in part, because for evolutionarily focused papers, computationally derived conclusions often satisfy the reviewers. By contrast, developing a useful product requires phenotypic experiments, which are major projects in their own right and inconceivable on the EBP scale. What this means is that people will not fully appreciate the enormous value of the sequences being generated until phenotyping becomes more scalable. The hope lies in artificial intelligence (AI), which has been astoundingly successful at predicting 3-dimensional protein structures, as recognized by the 2024 Nobel Prize in Chemistry. We do not yet know how (or if) AI can be extended from relatively simple molecular phenotypes to complex organismal phenotypes. But it is a sign of our times that Demis Hassabis, co-developer of the AlphaFold algorithms for protein folding, has declared that: "one day … we can cure all disease with the help of AI … maybe within the next decade or so" (6). The success of AI depends on the examples used to train it. For protein folding, the training set was a quarter million experimentally determined protein structures from the Protein Data Bank (PDB). For phenotyping, the concept of a training set is inherently open-ended as we cannot even define all the phenotypes. To show how varied the concept can be, I present two examples. Both experiments were funded to solve an important problem unrelated to AI— one in neuroscience (optogenetics) and the other in agriculture. 1) Optogenetics uses light to activate (or deactivate) genetically defined classes of neurons in awake behaving animals, including even humans. Algae derived channelrhodopsin proteins are heterologously expressed in these animals. The original experiments used a protein from Chlamydomonas reinhardtii; but many years later, experimental characterization of channelrhodopsin proteins from 127 phylogenetically diverse algal species uncovered a series of new proteins with improved performance, e.g., red shifted, faster responses, higher currents (7). Multiple companies are using these proteins to restore vision in patients with retinitis pigmentosa. In the latest news (8), a woman who had been legally blind can now sit down on the beach and read a book again. 2) Modern agriculture suffers from a lack of genetic diversity. Although the germplasm banks abound with useful alleles, crossing them with elite cultivars often transfers unwanted traits along with the desired traits. To unlock this bounty, we must first identify the responsible haplotypes. A recent study demonstrated how, measuring 137 phenotypes in the A.E. Watkins collection of global bread wheat landraces and in recombinant inbred lines generated by crossing selected landraces with an elite cultivar (9). Useful haplotypes identified included ones for yellow rust resistance, for increasing yield and height, and an alternative dwarfing gene for the Green Revolution. These findings are already being used in commercial breeding programs. Biomedical research has the funding to generate incredible resources like the UK Biobank (UKB)1, with genomics proteomics and metabolomics data for half a million human subjects, complete with lifetime healthcare records (10). It is unrealistic to demand a comparable resource on the EBP scale. But is it necessary? In other words, how many phenotypes (and which phenotypes) and how many species (and which species) are necessary for AI to generalize? Can we bootstrap our way to a better future by producing a lot of phenotypic data now, on the promise that over time AI will gradually eliminate the need to do such experiments? Epilogue Scientific progress requires that young scientists continue/improve on what their predecessors accomplished. In the case of the HGP, it was a new generation of scientists who created the first T2T assembly, filling in the gaps and correcting the errors in the original HGP sequence, many years after their predecessors sailed off into the sunset with well-deserved accolades. Similarly, it was a new generation of scientists who constructed the UK Biobank, generating phenotypic data that are empowering the next cycle of discoveries. Long-term success for the EBP will not come from the sequence alone, regardless of quality. It will come from the follow up studies by the next generation of scientists. I urge the EBP to expand their vision to include the training of young researchers to conduct activities beyond the generation of reference genomes. If so, the eventual completion of the EBP will be seen as the fitting end to a glorious beginning.