Motivation
Our session will address recent methodological advances in exploring biological molecular function. What is biological molecular function? We doubt you’d be able to get two people in the room to agree on what precisely this entails. By relying on textbook definitions the term can be interpreted to mean the specific biochemical activity performed by any molecule. Within a cell this is typically a protein or RNA. Activity includes enzymatic activities, binding interactions, and structural roles among an infinity of other functions. Here, however, the ease of definitions ends: Can function be described without consideration of molecular, cellular, organismal, and even community/ecosystem context? Are parts of the same pathway considered to be interacting even if they are never in the same location? Does the human version of a yeast enzyme have the same function?
Insensitive to the problems of definition, we, as a field, have been building methods for describing protein function for many years. Computational algorithms, and lately, specifically, deep learning advances, have tried hard to describe and predict functions of all molecules that we could find. Curiously, the latest generative AI advances have rushed ahead focusing on making new, synthetic molecules that carry out the specific function of interest – however defined. All this without being able to assign labels to the already existing variety.
In microbial communities and viruses existing methods fail to annotate a large fraction of the proteins present, often failing to find even putative homologs. RNA sequences, metabolites and other small molecules are even less well-characterized with only a fraction of identified species having any kind of known functional role. Furthermore, even the existing functional labels are limited. That is, once a new molecule is identified, it is unclear if it has some known functionality that it can be assigned or if it does something brand new.
We emphasize that function must depend on the context in which it takes place. High-throughput data generation methods, along with automated platforms for experimental investigation, offer ways to provide contextual information by measuring many thousands of molecular species in biological samples under different environmental conditions that cause changes in system interactions. Until now, however, the potential of these data streams to inform learning models for functional annotation has remained largely untapped for a variety of reasons.
The recent revolution in AI/ML methods represents a significant opportunity to make in-roads into the problem of known and unknown functions. We hope to bring to light new thinking about the following questions: How can we define (label) a function of a particular molecule? How can this label be propagated to other molecules? What is the relationship of these labels? How can labeling systems (ontologies or otherwise) describe context at different levels of detail, and link to phenotypic outcomes?
