Scientific Computing Associate II - AI Methods for Genomics

What We Provide:

• A competitive compensation package, with comprehensive health and welfare benefits.

• The opportunity to collaborate with skilled scientists and software engineers and work alongside computational and experimental enthusiasts.

• The ability to work as an independent scientist.

• An exciting and inspiring work environment at HHMI Janelia

What You'll Do:

We are seeking a talented and motivated candidate with machine learning experience to develop a deep learning classifier to identify the ancestors of genes of unknown origin, which is sometimes called the dark matter of the genome. On this project, you will be working in Scientific Computing Software reporting to Stephan Preibisch and collaborate with the Stern Lab . You will receive additional mentorship and guidance from Dr. Srini Turaga. Knowledge of genome and protein structure is a plus, but not necessary.

Many functional elements of the genome evolved so rapidly that their ancestral DNA sequences (remote homologs) can no longer be identified using standard DNA sequence similarity methods (e.g., BLAST). Many genes that parasites introduce into hosts, such as the so-called bicycle genes that small insects called aphids use to control plant physiology and development, are in this category. The Stern lab showed that remote homologs of bicycle genes can be found using a linear classifier that exploits gene structure features (specific DNA sequence elements within a gene such as exon size, number, phase, etc.) rather than only gene sequences. However, they also found that gene structure is evolving within the bicycle gene family and that the classifier loses power with more distantly related species.

Cells transcribe and translate gene sequences into proteins that carry out cellular functions and protein structure tends to be more highly conserved than the underlying gene sequence. A recent break-through in artificial intelligence, AlphaFold, which was recently awarded a Nobel prize, now allows researchers to predict protein structures of any gene. There is now the opportunity to use this abundant protein structure information together with gene structure and sequence information to search for remote homologs.

You will build a deep learning classifier that will exploit (1) genome sequence, (2) gene structure, and (3) predicted protein structure simultaneously both to identify remote homologs of bicycle genes and genes of unknown function across the tree of life.

If the classifier proves generally useful, there is an option to apply for support to develop it into a user-friendly and developer-friendly tool supported by the Janelia Open Science Software Initiative.