What follows is a comment on the 2022 ICBO paper “What AlphaFold teaches us about deep learning with prior knowledge” by Jobst Landgrebe. It seeks to throw light on the sense in which the prior knowledge used by AlphaFold is to be understood in ontological terms.
The AI models developed in the life sciences
have a much lower predictive power than the
models developed in domains such as engineering
or physics. Why is this so? In a paper for this
conference, Jobst Landgrebe analyzes AlphaFold
[
For a cluster of such decoded sequences, Alphafold can be applied to identify certain patterns in each protein homologous to the cluster, which then allow it to make highly successful predictions about its structure. This is remarkable given that, before AlphaFold, a very low degree of success had been attained in making protein folding predictions.
As Landgrebe explains, the prior knowledge
about protein folding ingested as input into the
AlphaFold machine learning algorithm takes two
forms: 1) as protein structure data (CIF files); and
2; as knowledge about protein homology groups.
These form the decisive factors which enable the
predictive success of the algorithm, which uses
only the protein’s amino acid sequence as input
and the heavy atom angle information as output.
Like other prediction algorithms in the field, the
ability of the AlphaFold model to predict protein
structure can be applied only for proteins
homologous to those with established structures.
This ability depends on a part of the implicit
model capturing the relationship between these
known folding structures and sequence clusters.
This is why the model succeeds; but also why it
fails to predict structures for those molecules
which are not homologous to proteins for which
the structures had been already determined using
classical protein crystallography. The model can
therefore not create new protein folding
knowledge – this must still be obtained from
experiments, which can take several years per
protein or fail altogether after unsuccessful efforts
(as is often the case, for example, for
transmembrane domains of proteins). On the other
hand AlphFold stands out as compared to other
structure prediction algorithms because it
achieves high accuracy even for sequences with
fewer homologous sequences [
There are two meanings of the term “ontology”: (i) as a branch of classical metaphysics dealing with the fundamental structure of the world, and (ii) as a scientific discipline that developed over the last 30 years, and which deals with organizing data and information about the world in a structured form to enable various sorts of data exploitation, for example in what is called ‘data science’. A significant fraction of the work carried out today under the label of ontology in sense (ii) is influenced by our understanding of ontology in the more traditional sense (i), above all in its use of the distinction between universals, organized in taxonomical hierarchies, and their respective instances – typically entities such as cells and molecules, which exist in time and space.
When Landgrebe claims that the prior
knowledge that was used by AlphaFold is a form
of ontology, he is referring to the CIF files, each
of which represents the structure of the protein it
describes. In what sense, then, is a CIF file an
ontology, or a part of an ontology? This is a deep
question, which takes us to the very foundations
of sense (ii) ontology, namely to the distinction
between universals and instances. The terms
protein, amino acid chain, histone, chordin, and so
forth, are unquestionably ontological – they are all
terms from the Protein Ontology [
PR:P06733-2: alpha-enolase isoform hMBP-1 (human) whose position in the PRO hierarchy is illustrated in Figure 1. This term is defined in PRO as: An alpha-enolase (human) that is a translation product of some mRNA whose exon structure and start site selection renders it capable of giving rise to a protein with the amino acid sequence represented by UniProtKB:P06733-2.
The mentioned UniProt sequence is an example sequence for a certain class of molecules (briefly: molecules having the same translation site and exon structure, where ‘same’ means ‘belong to the same class’).
Why, now, do we regard PRO as an ontology, and UniProtKB as a database? There are a number of answers to this question. Most importantly, as Figure 1 makes clear, the content of PRO is organized in terms of hierarchies of representations of universals of greater and lesser generality; something which is not the case for the content of UniProtKB.
Indeed, there is a sense in which UniProtKB comprises instance data as its content: almost all of the protein sequences it provides are derived through translation of the coding sequences (CDS) submitted to public nucleic acid databases on the basis of analysis of biological samples, i.e. of instances. In this sense each sequence is, given its provenance, itself a specific instance (it is the sequence of a corresponding specific sample).
Yet at the same time each such sequence is found in (is the sequence of) many trillions of molecules. It is for this reason that UniProtKB is useful to biological and biomedical research. Each UniProt sequence in fact represents a universal with these many corresponding instances. UniProtKB, it is true, lacks an explicit hierarchy for these universals, though one could infer an implicit hierarchy from the information in the entry. (We know, for example, that all information about the sequences is derived from the indicated gene.) UniProt, as contrasted with PRO and other ontologies, also lacks explicit definitions – though, again, these are implied. PRO is explicit in its representation of the molecules themselves and--for those cases that derive from UniProtKB-makes explicit those implied hierarchies and definitions.On the other hand, all the entries in UniProtKB (and in practically all other putative databases maintained by biological scientists), consist of representations of universals. Each of the protein sequences contained in UniProtKB, for example, almost certainly exists in some trillions of instances.
In the same way, each of the mmCIF files in
the Protein Data Bank (PDB) represents a protein
structure that, again, almost certainly exists in
trillions of instances. [
Why, then, are there not already more
predictive models in biology? Because organisms
are complex systems and it is only certain aspects
of such systems that can be modelled using
mathematics [
Thanks are due to Jobst Landgrebe, Darren Natale and Asiyah Lin for helpful comments.