Breno Livio Silva de Almeida — Computational Biologist

My current research interests, in a more technical version:

My research is about a representation problem: how much biologically relevant information can alignment-free methods recover from a DNA or protein sequence, and what remains beyond their reach? I compare handcrafted descriptors (which encode explicit compositional, physicochemical and mathematical properties) and embeddings learned by foundation models (which learn higher-order patterns from large collections of sequences). I look at what each representation encodes, where it breaks down and how useful it is across different datasets and evolutionary distances.

I use automated machine learning to select the appropriate representations and predictive models for individual biological tasks, thereby reducing the amount of expertise and resources required to apply these methods. I'm also using protein language models and hierarchical classification to examine proteins that resist annotation with traditional alignment-based methods in environmental metagenomes, and I'm building the infrastructure to enable reproducible public deposition of the underlying genomes and associated ecological metadata. The end game is to work out the representational limits of sequence-based prediction and to determine when additional context or experimental validation is required.

Now, in a less technical version that I could explain to my mom and impress her:

Proteins are molecules that do all sorts of things in humans, in other organisms and in the environment. A lot of them still have no known function, though. They're sometimes referred to as “functional dark matter” (although I'm told some people are quite against that term!).

The standard way of working out what a protein does is to compare its sequence to that of proteins whose function we already know. That works pretty well for proteins that are very similar to the ones we have a handle on, but it can fail to pick up proteins which have changed significantly through evolution. To the comparison tool they may appear entirely unrelated even though they might have similar structure or function.

My thesis is about how we can turn a biological sequence into information that a computer can learn from. One approach is to describe the properties that we already know how to measure (e.g. composition, physicochemical properties). Another approach is to use language models which learn the patterns in the data from millions of sequences (in much the same way that language models learn the relationships between words). Neither of these approaches is complete so I look at what each one tells us and where each one breaks down.