Jo Faraguna

I'd be interested in supervising projects using amino acid and structural space motifs and signals to facilitate pathogen detection from wastewater sequencing data. Depending on the direction this work takes, it could be slightly more traditionally bioinformatics oriented rather than deep learning / AI x bio oriented. Projects would likely leverage our ++long-read pooled nasal swab data++ or our ++short-read wastewater sequencing data++.

​

Some potential projects I'd be excited about, roughly in priority order:

​

1. Benchmark models for their ability to accurately fold and search protein fragments
2. We can extract 50-200aa protein fragments from our sequencing data, which are often not full-length proteins. It would be useful for us to assess these fragments for concerning properties (structural similarity to known toxins, binding potential to human receptors), but folding fragments outside the context of their larger structures tends to be tricky, both because short protein sequences are themselves ++degenerate in structural space++ and because even well-defined peptide folding ++can++ be ++hard.00479-8)++
3. This work would involve
   1. benchmarking existing folding models on the Pareto frontier of speed and accuracy (AlphaFold3, SimpleFold, ProstT5) for their ability to predict cropped protein regions in isolation,
   2. benchmarking existing structure-search models (FoldSeek) for their ability to accurately search using crops, and
   3. potentially fine-tune folding models using synthetic crop databases to improve their performance on this task.
4. Other approaches might involve e.g. constructing a custom MSA via protein search prior to fragment folding.
5. Classifying metagenomic dark matter with bio-foundation models
6. A large fraction of our metagenomic data is often unclassified by our ++existing tooling++ and is attributable to adapter dimers, PCR artifacts, or fragments too short to be classified by Kraken. However, some of the data does not fall into these categories.
7. Gold-standard approaches like BLAST could potentially be used to assign these reads, but would be computationally intractable. Furthermore, sometimes these unknown sequences do not yield high-quality BLAST hits.
8. Instead, it would be interesting to develop deep learning / genome/protein bio-foundational model based classifier methods to assign these reads to taxids or meaningful functional categories. Precedents for this work could include more broad approaches like ++METAGENE-1++ or more tailored, class-specific approaches like ++DeepARG++.
9. It would be important to benchmark these tools against more traditional bioinformatics approaches, such as very large Kraken2 databases, potentially augmented with assembled genomes derived from wastewater sequencing (++preprint++).
10. Biological language model sequence attribution
11. In our ongoing attempts to detect ++engineered pathogens++, we sometimes leveraged tools like ++chimera detection++, which rely on identifying substantial organism engineering via reads that span chimeric junctions. However, subtler forms of engineering are also worth trying to detect.
12. Similar to how there are methods for LLM text detection, it might also be possible to design methods to detect bio-foundation-model-generated nucleotide and amino acid generated strings.
13. There is ongoing academic work to explore watermarking (++example++), but given the availability of existing open-source, open-weights models and of public training data, it would be important to assess sequence attribution independent of watermarking. It might be tractable to develop classifiers for a set of commonly-used sequence models, such as ESM, Evo2, HyenaDNA. Tools such as ProteinMPNN might be in-scope as well.
14. In practice, it's likely extremely difficult to attribute sequences to bio-foundation models, even using our long-read pooled nasal swab ++data++, because these models are trained to reproduce natural sequence distributions. In practice, it seems likely that these models would be used to guide selection of a relatively small number of point mutations (which are often represented in the protein's ancestral sequences or protein family MSAs), so there would be low signal relative to WT proteins. Additionally, signal in generated sequences would likely be further noised using traditional host-codon-usage-based codon optimization techniques.
15. Therefore, this project would probably first involve a feasibility assessment prior to starting deep implementation work.