JASON Evaluation Report of the Human Genome Project – Technology (cont’d)
Previous Section * Title and Table of Contents * Next Section
JSR-97-315
October 7, 1997
The MITRE Corporation
2. TECHNOLOGY
The technology to sequence the human genome is now in hand. Indeed, this was true when the Project was formulated and initiated in 1990, and there have been significant improvements in the intervening seven years. Nevertheless, as we have noted in Sections 1.2.2-4, there are ample reasons to improve the present technology, particularly if the Project’s cost, schedule, and quality goals are to be achieved. Further, improvements in sequencing technology will accelerate genomics research and applications beyond human biology and medicine.
The Project faces the classic dilemma inherent in any large technological project: when to freeze the technology available, to declare “good enough” at the risk of not pursuing the “better.” We believe that the likely inadequacy and ease of improvement of the present technology and the future importance and relative inexpense of developing radically different technology all argue for pursuing both tracks simultaneously. Our rationale is presented in the following sections.
2.1 Improvements of present genomics technology
In the course of our study, we identified two aspects of the present sequencing technology where improvements that could have a significant impact seemed possible. These are
- Electrophoresis
- Algorithms for base calling, assembly, and finishing
We consider these in turn.
2.1.1 Electrophoresis improvements and an ABI users group
The Applied Biosystems Inc. (ABI) automated DNA sequencers are the de facto standard for sequencing and will almost certainly carry the brunt of the sequencing load for the Project. These are “closed-box” instruments that utilize proprietary technology owned exclusively by ABI. The company has both the responsibility and the financial incentive to ensure reliable, standardized operation of its instruments, even if this results in sequencing that is less than optimal. On the other hand, the desire of many end users, especially those at major genome sequencing centers, is to push the performance of these instruments to the limit.
This tension raises issues both of technology per se and of how new technology can be inserted into ABI machines to the satisfaction of all. We first discuss possible technology improvements, then propose a users group.
It is clear that modifications could be made to the hardware, and especially the software, of the ABI sequencers without sacrificing accuracy of base calling or reliability of operation; one of our briefers spoke convincingly to this issue [C. Tibbets, briefing to JASON, July 1, 1997]. These instruments use the Sanger sequencing method to sample automatically molecules labeled with any of four (ABI-proprietary) fluorescent dyes. The samples undergo gel EP in 36 lanes. The lanes are scanned with an argon laser and bases are “called” by a combination of hardware and software.
Errors can (and do) arise from a number of sources, including lane tracking; differential migration of the four dyes; overlapping emission spectra of the dyes; and variable oligomer separations, due, for example, to secondary sources. There are a number of efforts underway to improve the software packages used for interpreting the (trace) data stream produced by the sequencing instrument. It is important to note that specific improvements might have a dramatic impact on the Project, but be of marginal significance for broad classes of commercial applications. One example is attaining longer read lengths.
Specific areas with clear potential for significant improvement include:
- increasing the lateral scan resolution, thus allowing for more lanes;
- indexing the lateral scan in space (instead of time) for greater trace precision and reproducibility;
- adding a fifth dye for enhanced lane tracking;
- allowing access to the raw (preprocessed) trace data, thus enabling improved base calling algorithms.
ABI has no obligation to respond to users’ requests for modifications such as those suggested above, nor are they required to make available detailed specifications that would allow users to make such modifications themselves. As a result, advanced users are taking matters into their own hands through reverse engineering, even if this risks invalidating the manufacturer’s warranty or service agreement. For both legal and sociological reasons these aftermarket modifications tend to be made at the level of individual genome centers. This may result in fragmentation of the standards of practice for acquisition of sequence data, complicating the establishment of quality-control measures across the entire genomics community.
It would be desirable to unify the genomics community’s efforts to enhance the performance of ABI instruments, without infringing on ABI’s right to control its products and to guard its proprietary technology. We recommend that DOE take an active role in setting up an ABI “users group” that would serve as a sounding board for issues pertaining to the operation of existing instruments, the modification of existing instruments for enhanced performance, and the development of next-generation instruments. The group would include members from each of the major genome centers, various private genomics companies that choose to participate, and a sampling of small-scale users who receive federal support for DNA sequencing activities. The group should also include a representative from DOE, NIH, and (if it wishes to participate) ABI itself.
The activities of the users’ group should be self-determined, but might include in-person or electronic meetings, generation of reports or recommendations concerning the operation and potential improvement of the ABI instruments, and distribution of information to the scientific community via journal articles or the World Wide Web. DOE should provide principal funding for these activities, although industry members and ABI should pay expenses related to their own participation. It must be understood by all participants that ABI is under no obligation to consider or follow the recommendations of the users’ group. We would expect, however, that by finding common ground and speaking with one voice, the users will have substantial impact on the improvement of automated DNA sequencing technology, while maintaining common standards of practice across the genomics field and respecting the proprietary rights to sequencing technology.
2.1.2 Algorithms
Algorithms (and the software packages in which they are embodied) for lane tracking, base calling, assembly, and finishing appear to be in a formative stage. Research into new algorithms, and development and dissemination of software packages containing them, can return significant dividends in terms of both productivity and accuracy.
2.1.2.1 Base calling
The base calling problem involves converting a four-channel record of dye fluorescence intensity to a sequence of bases, along with a confidence value for each base. Several factors make this a challenging problem. Spreading of the intensity function along the lane leads to inter-symbol interference. Overlap in the spectral response of the four dyes leads to cross-talk. The spacing between bases may be non-uniform, certain sequences of bases distort the record, and the signal levels are very low toward the end of a read.
All of the problems present in base calling are also present in the demodulation of signals in communication and magnetic recording systems. As a result, there is a rich literature of methods for dealing with these problem. For example, inter-symbol interference can be reduced by employing linear equalization or decision-feedback equalization. Clock-recovery methods can be applied to keep the base calls properly centered. Sequences can be decoded as multi-base symbols to compensate for sequence-dependent distortion. A trellis decoder or a hidden Markov model can be employed to exploit knowledge about expected sequences to compute the most likely sequence to be generated by a particular intensity record. It would be worthwhile to consider implementing new (or improving present) base calling algorithms on the basis of these techniques.
2.1.2.2 Assembly
Assembly algorithms stitch together a set of sequences (of perhaps 500 bases each) that are subsequences of a clone (of perhaps 30 kb in length) to generate the (hopefully) complete sequence of the clone. The process is similar to assembling a linear puzzle where the pieces are allowed to overlap arbitrarily. We saw considerable variability in the methods used for assembly. The PHRAP program uses a greedy algorithm where the segments with the closest matches are assembled first and the program builds out from this initial start. The group at Whitehead, on the other hand, uses an algorithm based on tags to find overlapping segments. All of these algorithms are heuristic and approximate, as a complete search for the optimum map is perceived to require excessive computation.
There are many directions for research on assembly algorithms. To start, better methods for comparing two sequences to determine if they match can be employed. The PHRAP program achieves more accurate assembly by using base-call confidence values in grading matches. This corresponds exactly to the use of soft-decision decoding in a communication system. One can further improve the accuracy of matching by taking into account the sequence-dependent probability of erasures and insertions, computing, for example, the probability of a compression based on the surrounding GC-rich sequence. Similar techniques can be used to handle assembly in the presence of repeats.
Better methods for searching the space of possible assemblies can also be developed. For example, the greedy algorithm employed by PHRAP can get stuck if it makes a wrong choice early in its processing. One should benchmark such algorithms against a complete branch-and-bound search on representative difficult sequences to determine how often such failures occur. If there is a significant advantage to a full search, one can construct special-purpose assembly computers to perform this computation in a reasonable amount of time. For example, one could use an ASIC (Application Specific Integrated Circuit) or a few FPGAs (Field Programmable Gate Arrays) to build an accelerator that plugs into a standard workstation that will compute (in less than a microsecond) matching scores for all shifts of two segments through an algorithm that employs confidence values and sequence-dependent insertions and deletions. Even with a complete search, the use of heuristics is important to guide the search to explore the most likely assemblies first, so that large parts of the search space can be pruned.
2.1.2.3 Finishing
The finishing process involves taking an assembled sequence and filling in the gaps through a combination of manual editing and directed sequencing. At some sequencing centers we saw that finishing accounted for roughly half of the entire sequencing effort. Yet the software available to assist finishing consisted of no more than simple sequence editors. Research into finishing software has the potential to automate much of this labor-intensive process.
The first step toward automated finishing is to improve assembly software. Generating a correct assembly without manual intervention would eliminate much of the need for manual editing, leaving only the genuine gaps or compressions to be dealt with using directed sequencing.
The directed sequencing process involves ordering new reads of the clone using primers designed to extend the ends of sections that have already been sequenced. Much of this process can be automated using a rule-based expert system. Such a system is built by having a knowledge engineer observe an expert finisher at work and capture the finisher’s thought process in a set of rules. For example,
when a contig of a particular length is terminated in a particular way at each end, order a set of primers that match part of the sequence and order new reads taken using these primers and dye-terminator sequencing.
By combining the approaches taken by several finishers from different centers, the system could, in some cases, outperform a single human finisher. At the very least, a set of a few hundred of these rules would be likely to cover most of the common finishing cases. This would allow the human experts to focus their effort only on the most difficult cases.
2.2 DOE’s mission for advanced sequencing technology
We heard briefings from nine experts describing various technologies that might bring radical improvements to the art of sequencing DNA. These are discussed in some detail below. They are all different, but they have several features in common: they are non-EP, small-scale, and currently absorb a small fraction of the DOE genome project budget (some $1.7 M of the $13 M DOE technology development budget); unfortunately, they are scheduled to receive even less in the future. These projects are long-range, aimed at developing technologies whose greatest use will be come in the sequel of applications following the initial sequencing of the human genome. They are, to some extent, high-risk, exploring ways to overcome obstacles that could prove to be insuperable. But they also are high-promise, offering a real possibility of new sequencing methods that would be significantly faster and cheaper than gel EP.
How much money should DOE spend on high-risk, high-promise ventures? This is one of the important questions addressed by our study. We recommend a gradual increase of funding for technology development by about 50% (to $20 M per year) with a substantial fraction of this money going to projects other than improvements in current gel EP techniques. One should be prepared to increase this level rapidly in case one or more of the new technologies becomes ripe for large-scale operation.
In making this recommendation for increased support for advanced technologies, we are well aware of the need for the DOE to play a significant role in the current stage of the Project. We also know of, and approve of, the technology goals of vastly improving current EP techniques by such means as high-voltage capillaries, ultrathin gels, and use of resonance ionization spectroscopy. It is likely that such improvements in gel EP are essential to completing the Project on time, and we have commented in Section 2.1 on improving gel EP throughput in the near term. However, we believe that in the long run DOE’s greatest impact will be in support of the development of advanced technology for various sequencing tasks that go beyond the current goals of faster gel EP.
There are two main reasons for DOE to support these high-risk technologies. First, this is the part of the Project that plays to DOE’s strengths: the history and traditions of DOE make it appropriate (indeed, natural) for the Department to explore new sequencing technologies based on the physical sciences. Second, existing gel EP technology is barely adequate for sequencing a single human genome, and new technologies will be required to satisfy the future needs of medicine, biological research, and environmental monitoring. The new ventures supported by DOE are the seed-corn of sequencing efforts, for a crop to be reaped far beyond the Project itself.
2.2.1 Institutional barriers to advanced technology development
Most of the current attention in the Project is currently focused on rapid, low-cost sequencing of a representative human genome, to be finished by FY05. As a result, there has been a tendency to freeze technology at a fairly early level of development, sometimes not much past the proof-of-principle level, in order to cut down lead times. This tendency is exacerbated by the subsequent commercialization of the technology, making it difficult, for the usual property-rights reasons, to incorporate improvements found by those outside the commercial sector. Even this would not be so bad if it were not that the majority of genome researchers are not oriented toward technology development per se, but to the biological research that the technology enables. There is a vicious circle in which lack of technology support leads to an insufficient technology knowledge base among the supported researchers, while this lack of knowledge among peer reviewers leads to a reluctance to support technology development.
2.2.1.1 A parallel in ultrasound technology development
Three years ago, a JASON study sponsored by DARPA [H. Abarbanel et al., Biomedical Imaging (JASON Report JSR-94-120, August 1995)] looked at the maturity and sophistication of technology both for ultrasound and for MRI. In both cases the study found concrete examples of the institutional barriers discussed in the previous section. Ultrasound was further behind in advanced technology than MRI, and we will comment only on ultrasound here. The problems of ultrasound are well-known to all who work in it: The transmission medium (flesh and bones) is so irregular that images have very poor quality, interpretable only by those devoting their lifetime to it. In-principle improvements were known, especially the construction of two-dimensional ultrasound arrays to replace the universally-used one-dimensional arrays (which severely degrade the resolution in the direction transverse to the array). But this was a difficult technological challenge, requiring sophisticated engineering beyond the reach of much of the ultrasound community, and not representing an obvious profit potential for the commercial suppliers.
The JASON study found that members of the ultrasound research community were largely limited by the pace of commercial technology development, which was conservative and market-oriented, not research-oriented. In some cases there were ultrasound researchers quite capable of making advances in the technology, but frustrated by the lack of NIH funding. The study recommended that DARPA occupy, at least temporarily, the niche of technology development for ultrasound, which existed because agencies like the NIH were not filling it.
In response to this study, DARPA put a considerable amount of money into advancing ultrasound technology, with emphasis on using (two-dimensional) focal-plane array techniques developed by defense contractors for infrared and other electrooptical arrays. While it is too early to foresee the ultimate impact, it appears that this funding will significantly improve ultrasound technology.
2.2.2 Purposes of advanced sequencing technology
The goal of sequencing 3 billion base pairs of a representative human genome requires a limited amount of redundancy (perhaps a factor of 10) to insure complete coverage and improve accuracy. However, further developments in genomics will have to address questions of diversity, rarity, and genomic function, which may make this sequencing effort seem small.
One can imagine the need for continuing (if not increased) sequencing capacity as diversity becomes the issue. Diversity arises from individual variation (RFLPs, VNTRs, and other manifestations of introns, mutations in genes, etc.) and from the desire to compare human genomes with those of other species, or to compare (parts of) one individual’s genome with another’s. If it is ever to become possible for MDs and laboratory technicians outside biotechnology laboratories to do sequencing routinely, the sequencing process itself will have to become much simpler, and not subject, for example, to fluctuations in the artistry of the experts who nowadays prepare gels. (Not everyone subscribes to such a goal, the alternative being large sequencing centers to which samples are submitted.). The databases that keep track of this diversity will grow correspondingly, as will the search engines needed to mine the databases. It is not out of the question to anticipate computing needs increasing even faster (a pairwise correlation search of a ten times larger database may require up to one hundred times more searching, for example).
The hunt for rare alleles or rarely expressed genes (associated with rare phenotypes or obscure functions) may call for advanced technology for constructing and searching cDNA libraries, perhaps massively-parallel machinery built on a considerably smaller unit scale than is now common.
Functional genomics (to oversimplify, the understanding of the roles and interactions of the proteins coded for by DNA) presents difficulties so specific to each individual case study that it is nearly impossible to summarize here, and we will not attempt to do so. But it is clear that many functional genomics activities will require a total sequencing rate substantially beyond that of the Project.
Advanced technologies also have a role to play in quality assurance and quality control. The chemical and physical bases of current sequencing technology result in intrinsic limitations and susceptibility to errors. Alternative sequencing methodologies at least as accurate and efficient as the present one would allow independent verification of sequence accuracy. An example is given in Section 3.2.2 below.
Some advanced technology development will be (indeed, is being) done by commercial companies, to be sure, and that is to be welcomed, but if ultrasound or even the current state of the Project is a guide for the future, there is a most important role for DOE advocacy and support of advanced technology beyond the goals of initial sequencing of the human genome.
2.3 Specific advanced technologies
One cannot, of course, confidently predict the future of advanced technologies in any area. Instead, we comment in the following subsections on three directions that seem particularly promising:
- Single-molecule sequencing (by STM, AFM, flow cytometry, etc.)
- Mass-spectrometric sequencing
- Massively-parallel sequencing (hybridization arrays)
2.3.1 Single-molecule sequencing
For at least thirty years, some molecular biologists have been dreaming that it might be possible to sequence DNA molecules one at a time. To do this, three steps would need to be taken:
- Step 1: stretch out a molecule of DNA in a known orientation, with one end attached to a solid surface and the other end free.
- Step 2: detach nucleotides one at a time from the free end.
- Step 3: count and identify the nucleotides in order as they are released.
Before any of these three steps were mastered, the technique of sequencing DNA by gel EP was invented and the three steps became unnecessary – gel EP became the standard method of sequencing. A significant disadvantage of this method was the requirement for a macroscopic quantity of identical molecules as input. This requirement initially limited its application to viral genomes and other small pieces of DNA that could be obtained in pure form. The invention of PCR made the preparation of pure macroscopic quantities of identical molecules routine and gel EP could then be applied to all kinds of DNA. Thus, the technology was ready for large-scale development when the Project began (indeed, its availability was one of the factors in initiating the Project) and the technology of single-molecule sequencing was left far behind. [Single-molecule spectroscopy and related fields are nevertheless very active areas of research; see, for example, the Symposium on Single Molecule Spectroscopy: New Systems and Methods, held last year in Ascona, Switzerland.]
The Human Genome Project has supported some single-molecule sequencing efforts. We heard about two serious programs to develop single-molecule sequencing. One, at LANL, was described to us in a briefing by Richard Keller. The other, a proprietary program at seQ Ltd. in Princeton, was mentioned but not described in detail. Neither program is now supported by the Project. Details of the LANL program have been published [P. M. Goodwin, W. P. Ambrose, and R. A. Keller, “Single-molecule Detection in Liquids by Laser-Induced Fluorescence”, Accounts of Chemical Research, 29, 607-613 (1996); R. A. Keller et al., “Single-Molecule Fluorescence Analysis in Solution”, Applied Spectroscopy, 50, 12A-32A (1996)]
Why should anybody be interested in single-molecule sequencing? There are two main reasons. First, each of the three steps required for single-molecule sequencing has recently been demonstrated to be feasible. Second, single-molecule sequencing, if all goes well, might turn out to be enormously faster and cheaper than EP. The following paragraphs explain the factual basis for these two statements.
The first step in single-molecule sequencing is the attachment of one end of a molecule to a solid surface and the stretching out of the rest of the molecule in a controlled manner. This has been done by the LANL team, using flow cytometry, a standard technique of microbiology. A single molecule of single-stranded DNA is attached by the covalent bonds of the biotin-avidin protein system to a plastic microsphere. The microsphere is held in an optical trap in a cylindrical fluid flow, which pulls the molecule straight along the cylinder’s axis. The second step is the detachment of nucleotides in sequence from the end of the molecule. This has also been demonstrated by the LANL team, using standard microbiological techniques. Exonucleases are dissolved in the flowing fluid. A single exonuclease molecule attaches itself to the free end of the DNA and detaches nucleotides, one at a time, at a rapid rate (many per second).
The third step, the identification of bases in the detached nucleotides, is the most difficult. It might be done in at least three different ways. The LANL team identifies the bases by passing the flowing fluid through a laser-beam. As each base passes though the beam, the molecule fluoresces at a wavelength that is different for each of the four bases. Because the passage through the beam is rapid, the fluorescence must be intense if it is to be detected reliably. To intensify the fluorescence, the DNA molecule is initially prepared for sequencing by attaching a fluorescent dye residue to each base, with four species of dye marking the four species of base. The four types of base can then be identified unambiguously during roughly one millisecond that each nucleotide spends in the laser beam. Unfortunately, the LANL team has not succeeded in eliminating spurious detections arising from unwanted dye molecules in the fluid. They expect to be able to reduce the background of spurious events to a level low enough to allow accurate sequencing, but this remains to be demonstrated; it will require faster-acting exonucleasus than those now used.
The seQ Ltd. team accomplishes the first two steps in the same way as the LANL team, but addresses the third step differently. The bases are not modified by addition of dye residues. Instead, the unmodified nucleotides are detected by fluorescence in an ultraviolet laser-beam. Since the fluorescence of the unmodified bases is relatively weak, they must be exposed to the laser for a longer time. This is achieved by depositing each nucleotide, immediately after it is detached from the DNA, onto a moving solid surface. The surface is then scanned by ultraviolet lasers at a more leisurely pace, so that each nucleotide is exposed to the lasers long enough to be identified unambiguously. The details of this technique are proprietary, and we were not told how well it is actually working.
A third possible way to do the third step in single-molecule sequencing is to use mass spectrometry. The state of the art of mass spectrometry is discussed in Section 2.3.2. Mass-spectrometric identification of the detached nucleotides would require their transfer from the liquid phase into a vacuum. This might be done by ejecting the flowing liquid into a spray of small droplets, letting the droplets evaporate on a solid surface, and then moving the solid surface into a vacuum. Molecules sticking to the surface could then be detached and ionized by MALDI. Once ionized, they could be detected and identified in a mass-spectrograph, since the four species of nucleotide have different masses. (As noted in the next subsection, it is considerably more difficult to differentiate the four base pairs by mass than to distinguish their presence or absence, as in Sanger sequencing.) However, none of the mass-spectrograph projects that we heard about has addressed the problems of single-molecule sequencing.
To summarize the present situation, although each of the steps of single-molecule sequencing has been shown to be feasible, no group has yet succeeded in putting all three together into a working system. Dr. Keller informs us that he is exploring the possibility of collaboration with a larger German-Swedish consortium headed by Manfred Eigen and Rudolf Rigler. The latter have published a plan for single-molecule sequencing essentially identical to the LANL program [M. Eigen and R. Rigler, Proc. Nat. Acad. Sci. (USA) 91, 5740 (1994)], although LANL is ahead of the consortium in the implementation of their plan. If the collaboration goes ahead, the skills of LANL will be leveraged by the larger resources of the consortium.
We turn now from the present situation to the future promise of single-molecule sequencing. The promise is that it might become radically faster and cheaper than gel electrophoresis. The claim that single-molecule sequencing might be extremely cheap stands or falls with the claim that it might be extremely fast. Sequencing by any method is likely to be a labor-intensive operation, with costs roughly proportional to the number of person-years devoted to it. The costs of machines and materials are likely to be comparable with the costs of wages and salaries. When we are concerned with large-scale operations, the number of bases sequenced per dollar will be roughly proportional to the number of bases sequenced per hour. The main reason why gel electrophoresis is expensive is that it is slow. If single-molecule sequencing can be a hundred times faster than gel electrophoresis, then it is also likely to be a hundred times cheaper.
The claim that single-molecule sequencing might be a hundred times faster than gel electrophoresis rests on a comparison of the inherent speeds of the two processes. The process of gel electrophoresis requires about eight hours to separate molecules with resolution sufficient to sequence 500 bases per lane. The inherent speed of gel electrophoresis is thus less than one base per minute per lane. In contrast, the elementary steps in single-base sequencing might have rates of the order of a hundred bases per second. The digestion of nucleotides in sequence from the end of a DNA molecule by exonuclease enzymes has been observed to occur at rates exceeding one hundred bases per second. And the discrimination of bases in ionized molecules detected by a mass-spectrometer can certainly be done at rates of hundreds of molecules per second. These facts are the basis for hoping that the whole process of single-molecule sequencing might be done at a rate of a hundred bases per second. That would imply that an entire human genome could in principle be sequenced by a single machine operating for a year.
Needless to say, this possibility is very far from being demonstrated. The three steps of single-molecule sequencing have not yet been integrated into a working process. And the rate of sequencing in a large-scale operation is limited by many factors beyond the rates of the elementary process involved. With either single-molecule or gel electrophoresis separation, the production of sequence will be slowed by the complicated manipulations required to prepare the molecules for sequencing and to assemble the sequences afterwards. Until single-molecule sequencing is developed into a complete system, no realistic estimate of its speed and cost can be made. The most that can be claimed is that single-molecule sequencing offers a possibility of radically increasing the speed and radically reducing the cost.
Two other potential advantages of single-base sequencing are longer reading-lengths and superior accuracy. The reading-length in gel EP is limited to about a thousand bases (roughly half of this in conventional practice). The LANL group has demonstrated attachment and suspension of single DNA molecules with many thousand bases. It is likely that DNA molecules with tens of thousands of bases could be handled, so that a single-molecule sequence could have a read length of tens of thousands of bases. As the short read length of gel EP makes final assembly and finishing an elaborate and costly process, these longer reads could greatly simply the process of assembly.
One of the major obstacles to accurate sequencing is the prevalence in the genome of repeated sequences of many kinds. Repeated sequences are a frequent cause of ambiguities and errors in the assembly process. Since the single-molecule system will have longer read lengths, it will be less vulnerable to effects of repetition. Repeated sequences will usually be displayed, without ambiguity, within the compass of a single consecutive read. As a result, it is possible that single-base sequencing may be not only faster, but also more accurate than gel EP.
There are some efforts directed towards single-molecule sequencing by non-destructive methods using microscopes. The idea of these efforts is to discriminate bases by scanning a DNA molecule with an Atomic Force Microscope or a Scanning Tunneling Microscope. These efforts are far from practicality; we have not examined them in detail. Since the art of microscopy is advancing rapidly, it is possible that some new invention will make it possible to visualize individual bases in DNA with enough resolution to tell them apart. However, without a new invention, it appears that the existing microscope technology cannot do the job.
In conclusion, this study’s recommendation is that DOE give modest support to single-molecule sequencing effortsWhile we have only reviewed two small efforts, it appears to us that ,with modest support, there is a finite probability that single-molecule sequencing will be developed into a practical system. There is a smaller, but still finite, probability that it will prove to be superior to gel EP by a wide margin. Of course, funding decisions for individual programs, including those we have reviewed, must be made through the usual mechanisms, including rigorous peer review of prior accomplishments, rate of progress, and future potential.
One can look at the support of single-molecule sequencing from two points of view. On the one hand, it is a gamble that DOE can afford to take, offering an opportunity to win a large pay-off by betting a small fraction of the genome budget. On the other hand, it is a premium that DOE can afford to pay for insurance against the possibility that the electrophoresis-based sequencing program might fail to reach its schedule, budget, and accuracy goals. From both points of view, modest support of single-molecule sequencing appears to be a prudent investment.
2.3.2 Mass-spectrometric sequencing
In the simplest terms, mass spectrometry (MS) in DNA sequencing replaces the gel EP step in Sanger sequencing. Instead of measuring the lengths of various dideoxy-terminated fragments by observing their rate of diffusion in a gel, one measures their mass with one of several possible MS techniques, including time-of-flight (TOF) and Fourier-transform ion cyclotron resonance (FTICR) spectroscopy. Presently, MS techniques are usable on fragments of about the same length as those used in gel EP (that is, several hundred bases), although this is not a fundamental limitation. The real advantage of MS sequencing is speed, since reading the output of the MS instrument is virtually instantaneous, compared to eight hours or so needed for the gel lanes to evolve to readable length. Many other techniques can be used, in principle, for sequencing with MS, and we will not go into all of them here. Some of these require a mass resolution capable of distinguishing all of the four base pairs by mass; this is a difficult job, since A and T differ by only 9 Da. (Sanger sequencing needs only to resolve one whole base pair, or about 300 Da.)
In early investigations into MS DNA sequencing, the methods for preparing and ionizing DNA (or protein) fragments were fast-atom bombardment or plasma ionization. (There are recent review articles on DNA MS, including references to the work described below [K. K. Murray, J. Mass Spect. 31, 1203 (1996); P. A. Limbach, Mass Spectrometry Reviews 15, 297 (1996)]; the discussion here is based on these articles and on remarks from several experts.) But spectroscopy was limited to oligonucleotides of ten or fewer bases.
One significant step forward is the use of MALDI (Matrix-Assisted Laser Desorption/Ionization) to prepare ionic fragments of DNA for MS. The general idea is to embed the DNA in a matrix, which can be as simple as water ice, and to irradiate the complex with a laser of carefully-chosen frequency. This can both vaporize the complex and ionize the DNA, possibly by first ionizing the matrix followed by charge transfer to the DNA. There is a great deal of art in applications of MALDI, which is considerably more difficult to use with DNA than with proteins and peptides. For example, problems arise with unwanted fragmentation of the (already-fragmented) DNA during the MALDI process. Moreover, this MALDI fragmentation process is different for different bases. It is now possible to generate DNA fragments up to 500 bases long with MALDI, with resolution at about the 10 base level (compared to the needed resolution of 1 base). Typically MALDI DNA fragments have one unit of charge for every several hundred base pairs.
Another promising method for ionization is electrospray ionization (ESI). Here the charge produced is much higher (but can be varied by changing the chemistry of the solution containing the DNA). For example, experiments using T4 phage DNA fragments up to 108 Da have shown charges up to 3×104. It is then necessary to determine both the mass per unit charge (as in conventional TOF MS) and the charge, in order to determine the mass. One potentially-important method introduces the accelerated ions into an open metal tube, where they induce an image charge that is measured; the charge-to-mass ratio is then measured by TOF.
MALDI-based methods are generally best for Sanger sequencing, but improvements are needed in the mass resolution and sensitivity (equivalently, DNA ion yield). ESI techniques lead to both higher mass resolution and higher mass accuracy, but because a great many charge states are created, it is not well-suited to analysis of a mixture of a large number of fragments (as is required in Sanger sequencing).
Looking toward the future, there are two ideas in MS that might someday reach fruition.
Arrays and multiplex MS sequencing. Several briefers discussed ideas for using large arrays of DNA fragments with MS. One scheme [Charles Cantor, briefing to JASON, July 3, 1997] involves using arrays with various laydowns of DNA fragments, for subsequent MALDI-MS, with the fragments on the MALDI array designed to have properties desirable for MS. Another [George Church, briefing to JASON, July 2, 1997] points out that multiplexing with arrays is feasible for MS sequencing at rates of possibly 103 b/sec. One uses large (~65000) arrays with electrophore-tagged primers on the DNA fragments, with each primer having an electrophore of unique mass attached. DNA primed with these primers is grown with dideoxy terminators, just as in Sanger sequencing. The four varieties are electrophoretically separated, then collected as droplets on an array. Finally, MALDI-TOF is used to remove the electrophores, ionize them, and identify them by MS. Each of the 400 different varieties of DNA is thus identified, yielding a multiplex factor which is the number of different electrophores (400 in this case). (Electrophore tagging of primers has been suggested as a means of increasing the ion yield from MALDI [P. F. Britt, G. B. Hurst, and M. V. Buchanan, abstract, Human Genome Program Contractor-Grantee Workshop, November ,1994].)
Single-molecule detection. It is not obvious that MS-DNA sequencing requires single-molecule detection, but it in any case can be cited as the ultimate in MS sensitivity. It has already been shown [R. D. Smith et al., Nature 369, 137 (1994)] that a single ESI-DNA ion (up to 25 kb long) can be isolated for many hours in an FTICR mass spectrometer cell, making it available for measurements during this time. In another direction, detecting a single DNA molecule after acceleration should be possible, thus increasing the sensitivity of MS methods. Methods used for detection might involve bolometric arrays of detectors similar to those used for searches for cosmic dark matter. Such bolometric arrays are made on a pitch of ~25 µm for use as sensitive IR focal-plane arrays. An ESI-ionized 30 kDa DNA fragment of charge 100 in a 30 keV potential drop will deposit some 3 MeV in a pixel, the same as 3×106 optical photons. The 25 µm spatial resolution can be used for resolving the mass and charge of the ion. It is intriguing to note that a single charged DNA fragment is something like the hypothesized magnetic monopoles of particles physics; both have masses of tens of kDa and large charges (of course, magnetic charge for the monopole). Considerable effort has gone into methods for detection of single monopoles, which are known to be very rare. [Subsequent to completing this study, we learned of very recent and promising work by Benner et al. at LBNL using superconducting tunnel junctions for single-molecule detection.]
2.3.3 Hybridization arrays
A new technology that has progressed considerably beyond the stage of laboratory research is the construction of large, high density arrays of oligonucleotides arranged in a two-dimensional lattice. [“DNA Sequencing: Massively Parallel Genomics,” S. P. A. Fodor, Science 277, 393 (1997)] In one scheme (termed Format 1), DNA fragments (e.g., short clones from cDNA libraries) are immobilized at distinct sites on nylon membranes to form arrays of 104-105 sites with spot-to-spot spacing of roughly 1 mm.[“DNA Sequence Recognition by Hybridization to Short Oligomers: Experimental Verification of the Method on the E. coli Genome,” A. Milosavljevic et al., Genomics 37, 77 (1996)] In a second scheme (termed Format 2), techniques of modern photolithography from the semiconductor industry or inkjet technology have been adapted to generate arrays with 400,000 total sites [Fodor, op cit.] and densities as high as 106 sites/cm2 [“DNA Sequencing on a Chip,” G. Wallraff et al., Chemtech, (February, 1997) 22], although the commercial state of the art appears to be perhaps 10 times smaller. For Format 2 arrays, distinct oligomers (usually termed the probes) are lithographically generated in situ at each site in the array, with the set of such oligomers designed as part of an overall objective for the array.
In generic terms, operation of the arrays proceeds by interacting the probes with unknown target oligonucleotides, with hybridization binding complementary segments of target and probe. For Format 2 arrays, information about binding of target and probe via hybridization at specific sites across an array is obtained via laser excited fluorescence from intercalating dyes which had previously been incorporated into either probe or target, while for Format 1 arrays, readout can be by either phosphor imaging of radioactivity or by fluorescence. Interrogation of the array via changes in conductivity is a promising possibility with potential for both high specificity and integration of the readout hardware onto the array itself. [T. Meade, private communication]
Typical probe oligomers are of length 7-20 base pairs, with single base-pair mismatches between target and probe having been detected with good fidelity. [“Mapping Genomic Library Clones Using Oligonucleotide Arrays,” R. J. Sapolsky and R. J. Lipshutz, Genomics 33, 445 (1996); “Accessing Genetic Information with High-Density DNA Arrays,” M. Chee et al., Science 274, 610 (1996)]. For lithographically generated arrays, an important point is that all possible oligomers of length L (of which there are 4L) can be generated in of order 4L processing steps, so that large search spaces (the number of probes) can be created efficiently.
Such large-scale hybridization arrays (with commercial names such SuperChips [Hyseq Inc., 670 Almanor Ave., Sunnyvale, CA 94086.] or GeneChips [Affymetric, http://www.affymetric.com/research.html] bring a powerful capability for parallel processing to genomic assaying. The list of their demonstrated applications is already impressive and rapidly growing, and includes gene expression studies and DNA sequence determination. While hybridization arrays are in principle capable of de novo sequencing [“DNA Sequence Determination by Hybridization: A Strategy for Efficient Large-Scale Sequencing,” R. Drmanac et al., Science 260, 1649(1993)], the combinatorics make this a formidable challenge for long segments of DNA, since an unknown string of length N base pairs is one of p=4N possibilities (e.g., for N=103, p~10600).
Some sense of the probe resource requirements for de novo sequencing can be understood by the following “reverse” strategy applied to an array of Format 2 type. Consider an array containing oligomers of total length J with nondegenerate cores of length L that is exposed to an unknown fragment of length N. A posteriori one must be left with a sufficient number of probes that have matched the target so that a tiling pattern of probes can be assembled to span the entire target. As a lower bound on the number of required probes, imagine butting a set of N/L probes representing the nondegenerate cores end to end to cover the target, with p=N/4L << 1 so that the conditional probability for two probes to match identical but disjoint regions of the target is small. For (L, N) = (7, 103), p~0.06, while for (L, N) = (10, 104), p~0.01. Since each probe has as its nondegenerate segment an arbitrary combination of base pairs, 4L distinct oligomers are required in the original array, which for L=7 is 2×104 elements (well within the realm of current capabilities), while L=10 requires about 106 elements (an array with 400,000 sites is the largest of which we are aware).
Unfortunately, this simple strategy does not allow one to deduce the ordering of the matching oligomer segments, of which there are approximately (N/L)! permutations. Hence, imagine augmenting the above strategy so that the matching probes are arranged one after the other with the nondegenerate regions overlapping but offset by k base pairs. That is, adjacent probes are identical to each other and to the target in their overlapping regions, but differ by k base pairs in the nondegenerate regions at each end to provide sufficient redundancy to determine the ordering of the segments with high confidence. The number of probe segments needed to tile the target is then 1+(N–L)/k. With the assumption of only pair-wise probe overlaps (i.e., k>L/2), the requirement for uniqueness in sorting then becomes r=4(L–k)/[1+(N–L)/k] >>1, which cannot be satisfied for (L, N)=(7, 103), while for (L, N)=(10, 103), r is at most 5. On the other hand, for sequencing applications with N=104, L must be increased (L=14 gives r~10 for k=7), with a concomitant explosion beyond current capabilities in the number of array elements required (414=3×108).
Note that these simple limits assume that target-probe hybridization and identification at each site are perfect and that N is a “typical” random sequence without perverse patterns such as multiple repeats (which would present a significant problem). Certainly in practice a number of processes are encountered that complicate the interpretation of the hybridization patterns presented by arrays (e.g., related to complexity of the thermodynamics of hybridization, of patterns from multiple mismatches, etc.) and that are currently being addressed in the research literature, with promising demonstrations of fidelity. Clearly in any real application somewhat larger arrays than those based upon simple combinatorics will be needed for de novo sequencing to maintain accuracy and robustness in the face of errors, with an optimum array size lying somewhere between the limits discussed above.
While there are undoubtedly many “niche” applications for high density hybridization arrays to de novo sequencing (e.g., increasing the read length from 500-700 bases to beyond 1 kb would be important in the assembly process), such arrays seem to be better suited to comparative studies that explore differences between probe and target. Indeed, for Format 1 arrays, previously non-sequenced biological materials can be employed. It is clear that hybridization arrays will profoundly impact comparative genetic assays such as in studies of sequence polymorphism [M. Chee et al., op cit.] and of gene identification and expression, as well as for understanding the relationship between genotype and phenotype. Beyond the research environment, one can imagine biochemical micro-laboratories for clinical applications [G. Wallraff et al., op cit.] with hybridization arrays as essential elements for (differential) sequence analysis.