Placeholder text
We built four pieces of software while searching for plastic-degrading enzymes. Each is useful outside our own project, and each is described below with a small demo that runs in this page.
LOBSTER
To maximize our Wet Lab activity validation efficiency, we developed the LOBSTER (Library Optimization for Bioprospecting Sequences Through Economical Recombination) algorithm. Rather than naively purchasing full-length DNA sequences for each PETase core, LOBSTER utilizes golden gate fragment assembly and degenerate codons to create efficient fragment pool designs that can be combinatorially assembled to form a library of full-length cores. LOBSTER seeks to maximize the number of PETadex natural PETase cores which are generated in the resulting full-length library, which we define as coverage. To keep the generated libraries experimentally feasible, this coverage optimization is run subject to nucleotide quantity and library size constraints. Additionally, since the full cores are combinatorially assembled, golden gate assembly will create cores not found in the PETadex. While these chimeras of naturally-occurring mutations were not identified through natural sequencing, it is still plausible that they show plastic-degrading activity. This further expands our Wet Lab validation space of potentially active PETases.
Placeholder text
This section will also explain how the algorithm chooses fragments, how we checked its output against enzymes with known activity, and how to install and run it on your own sequences.
Placeholder demo
LOBSTER on one enzyme cluster: adjust the degeneracy and see the fragment library, the number of variants it encodes, and its cost against ordering every sequence in full.
Sharded DIAMOND search
Placeholder text
The search tool lets anyone compare a protein sequence against our full corpus from a web browser. The database is split into shards that are searched in parallel and merged, so a query over the whole corpus returns in under a minute, and each result has a permanent link that can be shared.
This section will explain how the search is built and how to use it.
Owner: Dennis
Placeholder: 32-shard zstd, Lambda fan-out, ~44 s end to end, shareable URLs.
Placeholder demo
Search results for a handful of well-known enzymes: a real hit table from the live tool that can be sorted and filtered.
Sequence record card
Placeholder text
Every sequence in the corpus has a record that gathers everything we know about it in one place: its domains, its predicted structure, the environment it was sampled from, the cluster it belongs to and its predicted activity.
This section will explain how a record is assembled and how to look one up.
Owner: Dennis / Sara / Lisa
Placeholder: data model, Postgres RDS, byte-offset retrieval, Next.js rebuild.
Placeholder demo
One candidate enzyme's record, with each annotation layer shown in turn.
ESM Atlas
To allow for exploration of PETadex enzymes in a more intuitive manner than alignment search, we built an Atlas of UMAPed embeddings of PETadex enzymes, modelled off of https://biohub.ai/esm/protein/atlas. This atlas provides an accessible way for users to see the scale of PETadex, even for those not familiar with biology. The embeddings capture rich information about the proteins themselves: colouring along metadata terms shows clustering along CATH domains, domains, and phylum, showcasing the diversity of PETadex. Clicking on any point on the atlas takes the user to a page for that family (for example https://petadex.net/family/23828/) providing a quick summary glance of that 30% cluster, including metadata terms associated with its centroid.
Availability and licensing
Placeholder text
This section will list where the source code for each tool lives, the open-source licence it is released under, and where the datasets and trained model can be downloaded.
Owner: Purav
Placeholder: S3, Hugging Face, bulk downloads, licensing.