Breno Livio Silva de Almeida — Computational Biologist

My current research interests, in a more technical version:

My research centers on 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, with embeddings learned by foundation models, which capture higher-order patterns from large sequence collections. I study what each representation encodes, where it fails, and how its usefulness changes across datasets and evolutionary distances.

I use automated machine learning to select representations and predictive models for individual biological tasks, reducing the expertise and resources required to apply these methods. I also use protein language models and hierarchical classification to investigate functional dark matter in environmental metagenomes, while developing infrastructure for the reproducible public deposition of the underlying genomes and ecological metadata. Ultimately, I want to understand the representational limits of sequence-based prediction and when additional context or experimental validation is needed.

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

Proteins are molecules that perform many essential functions in humans, other organisms, and throughout the environment. However, many proteins still have no known function. Researchers sometimes call these uncharacterized proteins "functional dark matter" (although some are pretty much against using the term "dark matter").

The usual way to infer a protein's function is to compare its sequence with proteins we already understand. This works well for close matches, but it can miss proteins that have changed greatly through evolution. To a comparison tool, these proteins may look unrelated even when they share structural or functional properties.

My thesis asks how we can translate a biological sequence into information a computer can learn from. One option is to describe properties that scientists already know how to measure, such as composition and physicochemical characteristics. Another is to use language models that learn patterns from millions of sequences, much as language models learn relationships among words. Neither view captures everything, so I investigate what each one reveals and where each one stops being reliable.