[Implications for] chemistry

Chemical compounds are a little like cellular automata and other kinds of programs. For even though the basic physical laws relevant to chemical compounds have been known since the early 1900s, it remains extremely difficult to predict the actual properties of a given compound. And I suspect that the ultimate reason for this—just as in the case of simple programs—is computational irreducibility.

For a single molecule, the minimum energy configuration can presumably always be found by a limited amount of computational work—though potentially increasing rapidly with the number of atoms. But if one allows progressively more molecules computational irreducibility can make it take progressively more computational work to see what will happen. And much as in determining whether constraints like those on page 213 can be satisfied for an infinite region, it can take an infinite amount of computational work to determine bulk properties of an infinite collection of molecules. Thus in practice it has typically been difficult to predict for example boiling and particularly melting points (see note below). So this means in the end that most of chemistry must be based on facts determined experimentally about specific compounds that happen to have been studied.

There are currently about 10 million compounds listed in standard chemical databases. Of these, most were first identified as extracts from biological or other natural systems. In trying to discover compounds that might be useful say as drugs the traditional approach was to search large libraries of compounds, then to study variations on those that seemed promising. But in the 1980s it began to be popular to try so-called rational design in which molecules were created that could at least to some extent specifically be computed to have relevant shapes and chemical functions. Then in the 1990s so-called combinatorial chemistry became popular, in which—somewhat in imitation of the immune system—large numbers of possible compounds were created by successively adding at random several different possible amino acids or other units. But although it will presumably change in the future it remained true in 2001 that half of all drugs in use are derived from just 32 families of compounds.

Doing a synthesis of a chemical is much like constructing a network by applying a specified sequence of transformations. And just like for multiway systems it is presumably in principle undecidable whether a given set of possible transformations can ever be combined to yield a particular chemical. Yet ever since the 1960s there have been computer systems like LHASA that try to find synthesis pathways automatically. But perhaps because they lack even the analog of modern automated theorem-proving methods, such systems have never in practice been extremely successful.