I think that, when it can, scientific research should answer important practical questions, and I think it’s reasonable to ask that this be done with some degree of rationality and thoroughness.

The scientific literature on diet and health shows that it isn’t. I think we can do much better.

Diet, heath, and the void

Here’s an important question:

How can someone choose a diet that lowers the risk of early, slow, and horrible death?

This general, practical question leads to many specific, scientific questions about messy real-world causality, and some of these can be answered.

The answers are almost irrelevant to basic science (they fit in the lower left quadrant of the table in my previous post), but they are vitally important to human health.

Here’s what I find striking and disturbing:

1. Standard research methods can answer many of these questions.
2. No one did the work that would have answered these questions.
3. The system hasn’t changed, and I see no real effort to change it.

The scandal of the science of dietary fats

For example, consider the question of cardiovascular disease and isocaloric substitutions among saturated fats, polyunsaturated fats, and carbohydrates.

And the answers? Saturated fats have long been known to be deadly, so surely there is unequivocal science that shows… What’s that? — Shows nothing of the kind?

Here’s the title and abstract of a representative 2010 analysis of the literature to date:

Effects on Coronary Heart Disease of Increasing Polyunsaturated Fat in Place of Saturated Fat: A Systematic Review and Meta-Analysis of Randomized Controlled Trials

Reduced saturated fat (SFA) consumption is recommended to reduce coronary heart disease (CHD), but there is an absence of strong supporting evidence from randomized controlled trials (RCTs) of clinical CHD events and few guidelines focus on any specific replacement nutrient….

This study finds that replacing substantial amounts of SFAs with polyunsaturated fatty acids — but not carbohydrates — reduces cardiovascular disease mortality, but that this is almost precisely offset by increases in non-cardiovascular-disease mortality. Another meta-analysis suggests that switching from SFAs to carbohydrates increases the risk of coronary deaths. [17 June: Updated to correct confused references.]

These evaluations of the scientific literature are not outliers.

A huge warning flag about the way we do science

For decades, questions of diet and health have been among the most prominent scientific questions in the world, with millions of lives at stake. For decades, billions of dollars have been spent by foundations and public agencies on health-branded research, presumably in the public interest. And for decades, basic questions regarding diet and health have been unanswered, or worse, answered by studies so poorly conceived that they led to the wrong conclusions.

More billions of dollars — and priceless public attention and life-changing effort — were poured into promoting and responding to advice based on this needlessly inadequate science.

Obvious questions, obvious importance, accessible answers, a failure to respond in a remotely rational way.

For better results, repair institutions

What might be done to help correct this? Here’s a broad brush-stroke suggestion for avoiding huge gaps in science done in the public interest:

1. Obligate science funding agencies to establish institutional mechanisms that are responsible for registering (allegedly) open scientific questions, and from diverse, external sources.
2. Make it embarrassing for the responsible parties to ignore questions that obviously shouldn’t be ignored.
3. For each question worth attention, require that they state an explicit estimate of the importance and difficulty of answering it.
4. Obligate the agency to either fund research on the most important and answerable questions, or to explicitly state reasons for neglecting them.

It think that institutions along these lines would cause a perceptible shift toward rational priorities. There are many potential implementations. The financial cost would be negligible. It would be great sport

I urge reformers to ask a simple question: If practices along these lines had been in place for the last 20 years, wouldn’t we be better off? Looking forward, how much would this be worth?

How groups think
and how to do it better

A review of Infotopia

I’ve been discussing problems with public information and ways to improve it with Mike Nielsen, and on this topic, he recommended Infotopia: how many minds produce knowledge by Cass Sunstein. Having just finished reading it, I recommend it too.

With a solid grounding in experiments and studies of group behavior (and enlightened common sense), Sunstein explores how groups and societies succeed and fail in what is arguably their most vital task: drawing out and assembling pieces of knowledge that are scattered among many minds. When this process of knowledge integration succeeds, groups can understand, decide, and act with knowledge and wisdom that exceeds that of any of their members.

When knowledge integration fails or goes astray, however, groups can perform worse than even their average members, and sometimes worse than any member.

The results of Suntein’s exploration are sobering, but the opportunities for improvement are staggering. If the knowledge and recommendations that Sunstein offers us were widely known and applied, blunders in group decision making would become substantially less common. These blunders range in size from small to large, and by “large blunders” I mean disastrously wrong decisions at the trillion-dollar, fate-of-nations level.

Infotopia is broad, describing, comparing, and analyzing a range of processes that draw on the power of many minds:

• Group deliberation
• Polling
• Conventional markets
• Prediction markets
• Open source development
• Blogging
• Wikis and Wikipedia

Sunstein helps us understand how these process operate and why they work and don’t work under various circumstances. Perhaps the most disturbing result is that deliberative discussion by groups often doesn’t work — that it fails to elicit information from its members, and that, under typical conditions, deliberation is more likely to amplify errors than to correct them.

In addition to diagnosis, Sunstein offers recommendations for improvement, some that a reader can apply next afternoon at work, and others that would involve changing how organizations operate.

Knowledge matters. Decisions matter. If your work or interests involve either, I think you’d enjoy reading Infotopia, and forever after, be glad that you did.

(I recently reviewed another book by Cass Sunstein, Nudge, and with similar enthusiasm.)

• A Map of Science
• How to Learn About Everything
• How to Understand Everything (and Why)
• The Antiparallel Structures of Science and Engineering
• Science and Engineering: A Layer-Cake of Inquiry and Design
• A Telescope Aimed at the Future
• Exploratory Engineering:
  Applying the predictive power of science
  to future technologies

Perimeter defense
considered harmful

William Wulf and Anita Jones have written a brief, tantalizing, and important article in Science: “Reflections on Cybersecurity”. They point the way out of a tangle of security problems (not all, of course) that costs billions of dollars in losses billions in countermeasures, and billions more in opportunity costs — some known and some not even imagined.

With current countermeasures, we’ve been losing the war. Wulf and Jones argue that the fundamental approach is wrong:

The current model for most cybersecurity is “perimeter defense”….Hackers try to “break in,” “firewalls” protect the system, “intrusion” must be detected, etc. But is perimeter defense the right underlying model?

They note that

[1] perimeter defense does not protect against the compromised insider….

[2] once the perimeter has been breached, the attacker has free access….

and [3] …it has never worked. It did not work for ancient walled cities or for the French in World War II (at 20 to 25 km deep, the Maginot Line was the most formidable military defense ever built, yet France was overrun in 35 days). And it has not worked for cybersecurity. To our knowledge no one has ever built a secure, nontrivial computer system based on this model.

The authors note the reason for the success and scalability of protocols that form the foundation of the internet:

Dave Parnas, one of the early software engineers, made a provocative and, we think, deeply important observation that helps to explain the success of the TCP/IP protocols. He pointed out that, when doing a design, the hardest decision to change is the one you make first, because all the subsequent ones to some extent depend on it. The decision for the TCP/IP protocols to do so little never had to be reconsidered, because it precluded so little.

They recommend that we deploy a cryptographic protocol of similar generality, one that does little, but does it well, and thereby provides a foundation for systems as complex, diverse, and specialized as the vast superstructure built on TCP/IP:

Is such a minimal mechanism feasible? We think so. In particular, at the network level, an application can use any computable function to decide whether or not to provide its service to a client if it can be absolutely certain who is requesting it. There is a class of algorithms known as “cryptographic protocols” for doing this that require knowing the public key of an object—so we conjecture that by providing just a way of accessing the public key of an object, one could build an arbitrary end-to-end security policy.

Because this minimalist foundation precludes so little, its generality seems assured.

In addition to its other virtues, it would be smoothly compatible with computational methods that can solve crucial vulnerability problems at the level of applications on individual machines (with very lightweight implementations, by the way). Wulf himself led pioneering work in this area, and he is no lightweight in computer science or Federal technology policy.

This proposal deserves more than ordinary attention.

(And here, more than ordinary emphasis.)

Not quite the same

In the science press, Big News often turns out to be hyped trivia, but the current Big News in quantum computing is something else — a self-hyping mutant of genuine big news, the discovery of an algorithm that promises exponential speedup in a class of problems where the result depends on the solution to a large system of linear equations. Contrary to the mutant news, however, the algorithm does not provide solutions to systems of linear equations: It outputs scalars, not vectors, and this is not at all the same thing.

The paper, published in Phys. Rev. Letters, reports work done at University of Bristol and MIT (Seth Lloyd is a coauthor). The mutation appears in the press release from the MIT News Office, which headlines “A quantum algorithm that solves systems of linear equation”, and says that

The quantum algorithm…can solve systems with a trillion equations and a trillion variables.

Well, no, it can’t — the output is a trillion times smaller. [In the comments, Larry notes that the MIT News Office article later quotes Lloyd directly, explaining the key caveat: that the algorithm delivers a scalar measurement on a solution vector, which can be a function of the entire vector or, as a special case, any one of the trillion vector components.]

Science News, reasonably enough, follows the not-quite-right formulation in the MIT lead, speaking of the algorithm “quickly solving monster linear equations”, while a brief news item in Science reports that “Solving a set of linear equations is a generic and important problem…[and the authors have developed] a quantum algorithm that will solve the problem much more rapidly”.

The synopsis of the paper provided by the journal’s publisher, the APS, says that it describes “a quantum algorithm for solving a set of linear equations”, and then, to clarify, notes that the algorithm “does not find the solution to the linear equations”. (It actually does clarify, though it’s written a bit sideways.)

The paper, “Quantum algorithm for linear systems of equations” is available as a preprint [pdf]. Here’s the abstract, highlighting the part that tends to drop out of the press reports:

Quantum algorithm for linear systems of equations

Solving linear systems of equations is a common problem that arises
both on its own and as a subroutine in more complex problems: given a
matrix A and a vector b, and a vector x such that Ax = b. We consider the case where one doesn’t need to know the solution x itself, but rather an approximation of the expectation value of some operator associated with x, e.g., x^†Mx for some matrix M. In this case, when A is sparse and well-conditioned, with largest dimension N, the best classical algorithms can find x and estimate x^†Mx in O(N poly log(N)) time. Here, we exhibit a quantum algorithm for this task that runs in poly(log N) time, an exponential improvement over the best classical algorithm.

If providing an input vector of size N were itself a problem of size N, of course, then an algorithm that included the input operations wouldn’t scale so well.

This observation highlights an important difference between quantum algorithms of the present kind and algorithms in the other two known classes that enable exponential speedup on interesting problems, Shor’s factoring algorithm (and its relatives) and quantum simulations of of quantum systems themselves: Those algorithms provide speedups that could bring problems down from otherwise inaccessible exponential realms (or thereabout); this one would push a tractable problem down into log-land.

Afterthought, 11 November: I suppose the reason I’m intrigued by this sort of process (subtle-but-enormous distortions creeping into descriptions) is that I’ve seen it happen in second-, third-, and fourth-hand descriptions of studies of prospective advanced nanotechnologies.

A vivid example is the simplistic equation of atomically precise fabrication (for example, by using positional constraints to localize familiar kinds of chemical reactions) with grabbing and placing individual atoms (which would typically require something more like magic). Unfortunately, in this case, subsequent discussions centered on the impossibility of magic, rather than the implementation of technology.

• How Nanotubes Grow: A theory that has nothing to do with reality

Gordon Bell, a long-time leader and innovator in the world of computation, has immersed himself in a life-changing experiment. Bits and pieces of news about it have been circulating for years, and his new book, just published, gives a full picture. In brief, Gordon records and indexes what he sees, hears, and more — and not as a diary, but as a working tool for living. With the resources of Microsoft behind the experiment, Gordon is exploring a future years ahead of the anticipated mass-market technologies. He calls the system “MyLifeBits”.

As the laziest possible review, I’ll just quote my blurb from the book:

Gordon Bell and Jim Gemmell paint a vivid and personal picture of a revolution that is already in progress, a revolution that will transform our future by making our past transparent. Clear, detailed, and permanent knowledge of ourselves and others will change the fiber of our lives and societies, pervasively, from meal planning to constitutional law. If we are blind to the implications, we’ll be trying to solve the wrong problems with obsolete tools. Total Recall will open eyes, and the more, the better.

Total Recall is worth reading.

Talk slides [pdf]

I’ve posted the slides for my WORLDCOMP’09 keynote, “Advanced Nanotechnology: Advanced Computing on the Critical Path”: Click here to download.

My earlier talk for the 2009 Berkeley Nanotechnology Forum was directed to a non-specialized nanotechnology audience and surveyed near-term directions in framework-directed self assembly as a basis for next-generation nanosystems. The WORLDCOMP’09 talk is organized like my other presentations to industry conferences and corporate audiences, linking an overview of current directions and longer-term prospects for nanotechnology to a discussion of how these developments will affect the more specific interests of the group (photonics, cosmetics technology, venture capital — that sort of thing). I’ve been encouraged by the follow-on discussions that have resulted.

• Slides for Berkeley Talk on Molecular Nanosystems
• Macromolecular Modeling for Molecular Systems Engineering

A useful resource

Nir London of the Macromolecular Modeling Blog has invited me to offer my perspective on the field. After patiently waiting for me to complete it, he’s posted the resulting essay, which I have cross-posted below.

The Macromolecular Modeling Blog is hosted by the Rosetta Design Group, which offers molecular modeling services based on the Rosetta protein design software, and supported by the Furman Lab which is part of the Rosetta Commons, the academic consortium responsible for the ongoing development of the Rosetta software.
Along with other news, the blog offers ongoing updates on literature in the field.

Macromolecular systems engineering can help us meet some of the most important technological challenges in the world today, ranging from medicine to renewable energy, and the development of better-integrated computational design tools will accelerate progress. Macromolecular modeling capabilities are advancing rapidly, but much of their potential for supporting systems engineering has yet to be exploited. In this post, I’d like to describe the fundamental nature of the problems and outline some of what needs to be done to make their potential a reality.

Why are design tools important?

The state of the art in molecular fabrication would enable us to build amazing nanosystems, if only we could design them. Rich opportunities are emerging from the convergence of protein engineering, structural DNA nanotechnology, and the advances that have in recent years produced a vast range of nanoscale, non-biomolecular components. DNA origami has shown us how to construct precisely ordered, aperiodic, addressable frameworks on a scale of hundreds of nanometers; protein engineering and models from nature show us that engineered proteins can organize diverse functional components into smaller aperiodic structures, and zinc-finger technologies (for example) suggest how these could be plugged into larger DNA frameworks. The challenge is to put it all together, and for self-assembling systems, this task reduces to the problem of designing the parts.

Protein modeling for science and design

Protein engineering is central to this concept of composite nanosystems, and the history and practice of protein engineering illustrate basic issues in building a design technology on a foundation of science-oriented modeling.

The field has come a long way. I greatly enjoyed my first RosettaCon last year, and I’m impressed by the people, the software, and the reported research. Once upon a time (1981) I published a PNAS paper that’s now down at the bottom of the citation tree for de novo protein engineering. Progress since then has validated the central insight, which is that engineering design is fundamentally different from scientific inquiry, and that this difference has consequences for assessing when a modeling technology is adequate for design and how to apply it.

In the case of protein engineering, no distinction had been drawn between the two protein folding problems, prediction and design, and opinion held that successful design must await successful prediction. I argued that the problems are radically different, and that design is arguably easier. (Carl Pabo cited this and termed design an “inverse folding problem”, a concept that was subsequently turned around and used for fold prediction.)

The fundamental difference between scientific modeling and engineering design is that in a design process, the physical system isn’t a given entity, but is instead found through exploration guided by a functional objective. This has far-reaching consequences.

Supporting design exploration

A designer works by generating and testing concepts, making choices, and filling in increasing detail. The initial stages of the design process typically require rapid screening of low-resolution candidate designs; later stages focus on fewer candidates with more physical detail and accuracy.

The requirements of the design workflow make a range of software capabilities very valuable:

• Interactive display and editing of structures, constraints, spatially structured penalty functions, etc.
• Quick application of approximate physics and partial optimization (e.g., structural relaxation, side-chain repacking)
• Smooth application of more accurate but computationally expensive modeling methods
• Task-relevant summaries of modeling results and comparison of results from variant designs
• Integration with a database for storing and organizing sets of related designs, models, modeling results, and so forth
• An open, extensible framework for all of the above.

Widening design scope

The drive toward developing more advanced self-assembled nanosystems will demand software systems able to model physical systems that contain diverse materials and components. This will increase the importance of:

• Extensible force fields with good facilities for editing and evaluation
• Protein design code that performs combinatorial search in heterogeneous environments
• Adaptation of combinatorial search techniques to an increasing range of non-peptide foldamers (e.g. peptoids)
• Use of physics-based energy functions for materials that lack huge databases of empirical structures

Increasing design scale

Advancing methodologies for making large, heterogeneous molecular nanosystems will increase the need for multiscale modeling:

• Developing coarse-grain models from atomistic simulations
• Integrating coarse-grain and atomistic models
• Integrating solid-body models with atomistic and particle-based models

A crucial question: When is a model good enough?

Judging when a model is accurate enough can be a crucial part of deciding whether a design problem is ready for serious work. As with protein fold prediction vs. fold design, however, a design problem may become tractable before a seemingly similar scientific problem has been solved.

Consider the criteria for acceptable failure rates: In a science problem where the objective is prediction, a 90% probability of failure in a given instance is just what it seems — a 90% probability of failure to achieve the objective. In an engineering problem, by contrast, the objective is to provide a suitable design, and the same 90% probability of failure in prediction in each given instance may result in a 100% success rate in achieving the objective (with a mean of 10 trial designs per success). What fails for science may suffice for design.

Another difference has more complex ramifications: Designers can apply predictability as a criterion for the design itself. Nature has no reason to cater to our modeling limitations, but designers can restrict themselves to using relatively well understood building blocks (such as alpha-helix bundles) and to relatively predictable interactions. Further, where the objective is to maximize something (such as stability), a designer can aim to provide a margin of safety to compensate for a model’s margin of error (artificial proteins provide examples; aircraft provide others). Again, what fails for science may suffice for design (and some modeling techniques that have been shelved may deserve a second look).

For all these reasons, the state of the art in macromolecular modeling is ready to support a wider range of design applications that is commonly realized, but harnessing the science to support design will take work. We’re ready for a broader push into the world of macromolecular systems engineering, the potential rewards are enormous, and there are many opportunities to contribute to progress through software development.

• A Third Revolution in DNA Nanotechnology
• Self-Assembly for Nanotechnology
• From Self-Assembly to Mechanosynthesis
• Exploratory Engineering:
  Applying the predictive power of science to future technologies

One kind of
atomically precise CAD
(materials unavailable)

---

CAD for structural
DNA nanotechnology.

---

Images from
Nanoengineer-1

Computer-aided design of structures on an ordinary scale can ignore atoms, and this is a major simplification. A piece of steel, for example, can typically be treated as a homogenous and isotropic material. The dimensions and angles of a steel component can be chosen freely: With few limitations, a steel plate can be of any thickness, and a smooth groove can be cut across its surface at any angle.

On a length scale of nanometers, however, accurate control of the shape and properties of a component requires precise control of the arrangement of its constituent atoms, and atoms aren’t available in fractional sizes. A smooth plate of a particular crystalline material, for example, can have one of only a limited number of thicknesses, and these are determined by its lattice structure. Likewise, smooth surfaces can exist only at particular angles, the ones that align with lattice planes. The choice of size, shape, and material are interlinked.

In ordinary engineering, the designer chooses the shape of a component. In atomically precise engineering, the designer chooses the way atoms are to be linked (subject to many constraints!), and interatomic forces then choose the shape of the component. Design and modeling become more tightly linked: Rather than using physical models only to evaluate the behavior of a structure of an assigned shape, physical models must be applied earlier, when searching for a structure that will assume an acceptable shape. (I’ll have an update on Nanoengineer-1 in a few weeks.)

What I said above focused on atomically precise structures based on a materials having crystalline order, but these much easier to design than to make. The linkage between design and modeling becomes tighter for the easier-to-make, harder-to-design structures that are of practical interest today, as I’ll be discussing later.

See also: Design Software for Atomically Precise Nanotechnologies