News from the lab

There are a couple of posts on the topic of large language models and the way we write about research that I have been thinking about for some time. Now that both the physics and chemistry Nobel prizes have just been awarded to this topic and its applications I figured it would be a good time to dust them off. Here’s the first one!

Anyone who is working in education at any level has probably been thinking on how to respond to the advent of Large Language Models (LLMs), which really entered in the daily practice of millions of students when ChatGPT was released in early 2023. Regardless of which side one takes on the use of LLMs for writing essays, manuscripts and other kinds of assignments or writing-based contents, we have to recognize that they will be used extensively for the foreseable future. Given how basically impossible it is to buld a reliable detector of text written by a LLM, there is no easy way to outright ban its use. Once that realization has sinked in, we can actually start focusing on how these tools can be used as a way to become better at our job.

Even though these LLMs have progressively gotten better at various tasks, they still tend to produce very long pieces of texts, and as a result they lack a good “punch”. This is maybe why everyone has fun prompting these models but almost nobody likes to read the text they produce. This also means that one way to differentiate ourselves from a LLM is to be very concise and to the point when describing something. But this is not an easy task! This famous quote from Pascal sums it up nicely: “I apologize for such a long letter - I didn’t have time to write a short one.”.

With these thoughts in mind we came up with a little “game” during our last lab retreat. Could we all come up with very concise descriptions of our own research? We also are mindful of the need to reduce the amount of jargon as much as possible to make our research understandable by as large an audience as possible. We therefore started with the following task: “describe your research using only the 1’000 most common words in the English language”. This task is of course inspired by the Up Goer Five xkcd comic, and we used the great online text editor made by Theo Sanderson.

Here’s the results:

• We think that different kinds of this tiny animal that lives in our stomach become stronger to stuff more quickly than others. We are checking this by taking these different kinds of tiny animals and giving them more and more of the stuff until they die. We think that those tiny animals that live longer then the others are faster, and so more interesting.
• I am working with the things that remembers how very small animals change over time, which all animals carry inside themselves. My job is to make a computer find positions on these things that are possible attack points for new things that make people not die as fast. I am focusing on the small animals that are very bad for people and try to not hurt the good ones.
• To use computers to understand how really little animals live and move around people and the world, and recognize which things are more important for them to do so.
• Tiny little things have tiny blocks that help them to live in their houses. Computers help to study what those tiny blocks do and that is important to study. Each little thing has its own blocks that do different jobs, and we do not know a lot about them. We can better understand how little things live and do their job if we study their blocks.
• Small living things can grow in different places. One way they can do this is by causing changes within themselves sometimes in areas that are not very obvious. I am looking into these places to see if they can explain why these tiny things act the way they do.
• I made a group of tiny things, and put them with some drinking things to see how tiny things can grow inside group
• During my time joining the group I started with learning how to use a computer to find possible reasons why some things are worse than others for feeling well. After a while, I also started to do this by using a faster thing for that. I was trained me on how to check it in real life.

Notice how hard we had to work to come up with a way to write “bacteria”; we all came up with a different term: “tiny animal”, “small animal”, “little animal”, “tiny little things”, “small living things”, “tiny things”. The most accurate definition was Bamu’s “small living things”. We then relaxed the requirements and rewrote these descriptions to use any word from the English language, but still trying to be as accessible as possible:

• We think that different types of this bug that lives in our intestine become resistant to antibiotics with different speeds, meaning that some types become resistant faster than others. We are checking this by giving these different types of bug more and more antibiotics, until they die. We think that those types that live for longer are able to develop resistance faster, which is interesting.
• My project focuses on the DNA of bacteria and how we can find areas on the DNA that are possible targets for new drugs. We have a need for these new drugs, because we used the old ones too much and now the bacteria have gotten used to them. My computer program will find very specific areas on the DNA that (hopefully) only occur in a few bad bacteria, so we can target them instead of a wide range of bacteria, which is also bad for us.
• I work with my computer and use the information of sequenced DNA - as it was done with the covid variants, to understand which bacteria infect more than others, and use this to prevent further infections.
• Essential genes are required for bacteria to survive and grow under different conditions. Bacteria can die or cannot grow properly if they loss those genes. Using ML* we can predict which gene is essential for bacteria but it is not an easy task. Genes essentiality is context-dependent; we do not have a big amount of labelled data (correct answers for computer) and we do not know which properties of genes make it essential. I build model that can learn genes properties and predict whether they are essential. *ML (black box for fortune telling)
• Bacteria can fine-tune their behaviour to adapt to different environments. But how do they do this? Most research have focused on changes within specific regions known to carry information about these “behaviour”. But what if there is more to the story? My work looks into the surprising role overlooked regions of the bacterial DNA have in regulating how their genes are expressed and how it might affect their response to some drugs used to kill them.
• Antibiotic resistance is the most burning issue in public health, people try to combat this issue by several method, including by trying to find the mechanism. One of them is how bacteria can spread antibiotic resistance genes. My project aims to create synthetic bacterial communities, which can detect antibiotic resistance genes spread among the communities
• In my internship, I learned how to find sections in the DNA of bacteria, that are responsible for making them more likely to cause infections. With this method, one could also try to understand, why infections with some bacteria are worse than others. I also learned how to evaluate the findings by performing experiments in the laboratory.

This exercise was definitely easier than the previous one, and hopefully easier to follow as a reader. Dilfuza’s definition of machine learning (“black box for fortune telling”) was particularly funny! However, you may notice how we all ended up writing a longer piece of text now that we could use any word we liked. We therefore tried one more time with the requirement being that we needed to write a single sentence of maximum 100 characters (spaces excluded):

• We study how different bugs develop resistance to drugs by giving them more until they die and comparing how long it took
• I use a computer to find specific areas on DNA of bad bacteria, for which we can make new drugs that target these areas
• Use genome data to understand how bacteria behave and prevent transmission and infection
• I make a computer to learn which genes of bacteria are important for their survival that help us to kill them efficiently
• I study how bacteria induce changes within their DNA to adapt to environments and cause antimicrobial resistance
• By evolving synthetic bacterial communities with antibiotics, I aim to detect antibiotic resistance genes spread
• I learnt computational and laboratory methods to find changes in DNA of germs that alter their potential in infections

This last excersize is probably still difficult for a LLM to master, and we think we did a pretty good job! The machines are not getting us fired yet, and we have an incentive to be more concise and direct.

Believe it or not, exactly one year after our preprint on predicting epistatic interactions in SARS-CoV-2 went online we have published a revised version in the journal Genome Biology. Talk about good timing!

What we wrote one year ago is still a good primer on what epistatic interactions are and why SARS-CoV-2 is a good test case for developing methods that are able to work on large datasets.

What did we change with respect to the preprint? The main change is very obvious: we have six (6!) new authors, which have contributed to run some experiments aimed at validating some of the predicted interactions.

We decided to focus on interactions that had emerged after the March 2023 cutoff we had from the preprint, as a way to have a relatively “blind” validation. Our colleagues from Twincore and HZI (with a special mention to Maureen and Henning!) then built and tested pseudoviruses with the single and double mutations for their ability to infect human cells and escape antibodies. We were able to validate all of our hand-picked interactions, which is a good indication that our predictions are sound and that in theory the method we propose could become part of the genomic epidemiology “arsenal”.

Another addition of note is a visual representation of how our samples weighting method works, which explains how we were able to identify the epistatic interaction that gave Omicron its “super-powers” with as little as seven (7!) Omicron samples (the light orange dot below).

Lastly, a couple of notes on the process of getting this paper published: given that the main author (Gabriel) did not need to have this work published in the usual way for the benefit of his own career (he had already an offer for a graduate student position), we could experiment a bit more freely. We decided to use Review Commons to solicit reviews in a journal-agnostic way. We got very stimulating comments from two anonymous reviewers (attached to the preprint); after we were done replying to them we could pick where to submit the revised manuscript from a list of affiliated journals. I wish I could tell you which journals are currently affiliated, but my google-fu came back empty-handed. I really liked the whole idea and process, and deciding which journal might be a good fit after being done with the first round of peer review feels much easier. We are convinced that Genome Biology has the right readers for the main points we tried to make. The only downside to the choice of journal would have been its steep APC of 4290 Euros. Luckily for us we did not have to directly cover this cost, as our institution participates in Projekt DEAL, of which I only had a very vague idea about. Not only did we use SARS-CoV-2 as a way to peek into the future of genomic epidemiology, but also in that of the research publishing system!

We have just published a research preprint describing microGWAS, a software pipeline to facilitate microbial GWAS studies. This effort was led by Judit, Bamu, and Jenny, who worked as a team to make my old chaotic code a “production-ready” tool. It was a real pleasure to see this communal effort take shape!

What we wrote last time we published a method related to microbial GWAS remains a good simple primer on the subject:

As everyone in the field of genomics has heard ad nauseam, we now have an abundance of genome sequences available; when that is combined with phenotypic measurements the obvious question is then “which gene is responsible for this phenotype?”. Statistical genetics (i.e. genome-wide association studies, GWAS) would be one way to answer that question, or rather the more correct one “which genetic variant is associated, and hopefully causal, for the variation in phenotype, across this collection of genomes?”.

The complex nature of microbial genetic variability means that in practice one has to use a number of software tools to preprocess genomes prior to the actual statistical association analysis. These tools have each a number of quirks and informal best practices, and very often one needs to write a small script to connect the output of one tool to the input of the next one in the so-called “software pipeline”. On top of these diffilcuties, very rarely the “raw” output of the association analysis provides the information that user needs. Most commonly a user would want to know: 1) which genes are associated with a given phenotype, and 2) is there a biological process that is overrepresented in the gene list?

These are the problems that our pipeline hopes to solve! We used the popular Snakemake workflow management system to connect each individual step of the typical end-to-end microbial GWAS analysis, including a number of downstream analyses to provide the users with an annotated gene list.

As you can see from the simplified scheme above, the pipeline carries out all of the work needed to go from annotated genome assemblies and a phenotype table to annotated results and diagnostic plots. We even leverage Snakemake’s support for conda to automate the cumbersome (and frankly irritating) process of installing and insulating individual tools.

As we want to make this pipeline sustainable in medium term, we have also added a small test dataset to speed up the developing process; we hope that young and eager researchers in the microbial bioinformatics community will be interested in contributing to maintain the pipeline and implement new features.

More information about the four (4!) sets of genetic variants that are used in five (5!) distinct associations tests can be found in the preprint, as well as in the online documentation.

Last month Dilfuza successfully defend her PhD dissertation from the questions of her two examiners: Ana Rita Brochado and Dan Depledge.

Congratulations to her for pulling this off and be the first PhD student to graduate from our lab!

We are very happy to report that the first PhD thesis from the lab has been submitted this month! With just one day to spare before the deadline (as it should be 😅), Dilfuza has submitted her thesis to the ZIB office. Now we wait for the public defense in June.

In other news, last month Adam has officially left the lab to take an exciting new job as a postdoc in the lab of Craig MacLean at the University of Oxford. Luckily Adam made a big push before leaving and finished some large scale experiments thanks to his usual stamina, which will be missed!

Congratulations to Dilfuza and Adam for these exciting news!