lysozyme

Former graduate student here. The most important thing to remember is that every paper has a main point, and everything in it is reinforcing that main point. Figure out what the point is. The whole story is in the figures.

Start with the title. Then look at each of the figures carefully. Figure out the limits of the paper from the methods. Certain methods can only give certain types of results. Again, look at the figures. The figures tell the whole story and are worth way more than the text. What is the story the figures tell?

Ignore the rest of the text. If you must read something, read the text of the results and look at the supplement. If you have a lot of spare time, skim the citations and identify the couple of sentences in the discussion that express the author’s methodological regrets. Under no circumstances should you read an introduction.

Discussed last year too (329 comments)

https://news.ycombinator.com/item?id=25854523

MMseqs2 stands out as great software in the world of biology. It’s a useful, fast, powerful, flexible command line tool for searching and clustering biological sequences.

Unlike a lot of free software in biology, MMseqs2 has a decent user experience, with a fast and clean command line app and great documentation

That’s true, software for biology tends to be terribly, almost comically bad, with one-off file formats, brittle data interchange, and impossible-to-maintain code being the norm. With user interface and ergonomics being the most neglected aspect. Why do you think that is?

Surely there is plenty of money in biology these days to hire a good designers to design good user experiences. Surely better user experience for biology software would lead to better understanding of biological systems and better outcomes in bioengineering.

Where are the polished, powerful design tools for biology like those that exist for other fields like online advertising that routinely process and distill huge amounts of lightly-structured data?

I’ve tried a lot of ad blockers over the years. Browsing mainstream news on the internet, in particular, highlights the absolute cesspool of pop ups and auto-playing videos that characterize the contemporary web.

None, least of all Stop the Madness, have worked as well and on as many devices as this simple bookmarklet, whose author I don’t know but who has saved me countless aggravations

    javascript:(function%20()%20%7B%20var%20i,%20elements%20=%20document.querySelectorAll('body%20*');%20%20for%20(i%20=%200;%20i%20%3C%20elements.length;%20i++)%20%7B%20if%20(%5B%22sticky%22,%20%22fixed%22%5D.includes(getComputedStyle(elements%5Bi%5D).position))%20%7B%20elements%5Bi%5D.parentNode.removeChild(elements%5Bi%5D);%20%7D%20%7D%20%7D)();

Edit: credit to the original author [1] and thank you!

1. https://alisdair.mcdiarmid.org/kill-sticky-headers/

There are many different types of plastic and they have surprisingly different molecular structures. Enzymes are remarkably specific catalysts, so an enzyme that eats PET [1] can’t eat another type of plastic, like HDPE [2]. So one solution could be to engineer an enzyme that can’t eat the liner material.

I’m not sure if HDPE would make a good liner material or not, but just as an example, the differences in molecular structure between PET and other plastics are big enough that it would take an array of different enzymes to eat them all.

1. https://en.m.wikipedia.org/wiki/Polyethylene_terephthalate 2. https://en.m.wikipedia.org/wiki/Polyethylene

In terms of the enzyme engineering challenge, it turns out to have been pretty straightforward to improve upon the starting enzyme.

The designed enzyme sequence

>contains five mutations compared to wild-type [1]

In this case, the enzyme used as the starting point for engineering turns out to need only a few small changes to improve its ability to efficiently depolymerize PET.

Choosing which of those small changes to make is the aim of the machine learning algorithms the authors used. The authors provide a visualization of exactly where those changes are inside the protein on their website

>Interactive visualizations of MutCompute for Fig. 1 are available at https://www.mutcompute.com/petase/5xjh and https://www.mutcompute.com/petase/6ij6

1. https://www.nature.com/articles/s41586-022-04599-z

The technique used here to get pictures of the folded proteins (cryo-electron microscopy, or cryo-EM) has seen steady improvements over the past couple of decades. [1] It’s analogous to room temperature light microscopy like you might have done in school, but with electrons instead of photons:

>Imaging biological objects in an electron microscope is, in principle, analogous in some respects to light-microscopic imaging of cell and tissue specimens mounted on glass slides. In light microscopy, visible photons serve as the source of radiation; once they pass through the specimen, they are refracted through glass optical lenses to form an image. In electron microscopy, the radiation is electrons, emitted by a source that is housed under a high vacuum, and then accelerated down the microscope column

The number of protein structures determined via cryo-EM is growing fast. Why is cryo-EM exciting for protein structure determination?

The primary technique for determining protein structures is X-ray crystallography, which, as the name implies, requires you to first produce, purify, and crystallize the protein. In contrast, cryo-EM allows determination of the protein structure without having to crystallize it.

In a typical cryo-EM experiment for protein structure determination, the protein molecules are imaged sparsely on a thin film, and many 2-D images are taken. These 2-D images are used to reconstruct the 3-D structure using a variety of computational techniques (including, recently, deep learning)

>images of the object, each with a different orientation, have 2D Fourier transforms that correspond to sections (indicated by red arrows) through the 3D Fourier transform of the original object. Thus, once the 3D Fourier transform is built up from a collection of 2D images spanning a complete range of orientations, Fourier inversion enables recovery of the 3D structure

In the analogous X-ray crystallography experiment, you have to grow crystals of your protein before imaging. The conditions that provide nice crystals are unknown and crystallization itself is sometimes completely out of the question, such as in this article where the authors are imaging a tissue sample

1. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3537914/

Thank you for posting this link. The video does a very good job of connecting the molecular details (even down to the geometry of a transition state structure) to the macro-scale movements of ATP synthase

Myosin is a very cool protein because it’s a physical and direct link between the microscopic world of atoms and electrical charges and the macroscopic world of moving bodies.

How myosin and other protein machines work at the atomic level is actually pretty well-known! [1] Myosin and its partner actin use the same biophysical principles, such as hydrogen binding and protein conformational change, as other proteins. But whereas other proteins typically act on the scale of atoms (doing things we think of as “chemistry”, such as making or breaking chemical bonds), myosins and other molecular machines are directly and physically responsible for the macro-scale movements of beating wings, walking legs, and beating hearts.

On the subject of molecular machines: it’s sometimes asked why biology doesn’t use the wheel. Protein machines like ATP synthase [2] (which creates the ATP used by myosin to create movement) provide an answer

1. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6618170/

2. https://pubmed.ncbi.nlm.nih.gov/7704582/

This tracks with your comment history

https://news.ycombinator.com/item?id=567736

Facebook may bring “a lot of good to the world”, but it brings a lot of bad, too. Is it worth the tradeoff?

The strain of rat used in these experiments, the Sprague Dawley, is commonly used to study cancer because it has a higher baseline incidence of cancer and grows solid tumors faster than other strains [1].

Interestingly, it’s the same strain that was used in the now-discredited study purportedly showing that RoundUp causes cancer [2].

1. https://en.wikipedia.org/wiki/Laboratory_rat#Sprague_Dawley_...

2. https://en.wikipedia.org/wiki/Séralini_affair

A related work, “MSA Transformer” [1], contrasts the strengths and weaknesses of language models using single sequences as input (as here) against language models using alignments as input. The authors of “MSA Transformer” perform ablation studies and various kinds of feature randomization, and compare directly against the state of the art (Potts models) in predicting residue-residue couplings.

One interesting note from “MSA Transformer”

> Potts models and single-sequence language models predict protein contacts in fundamentally different ways. Potts models are trained on a single MSA; they extract information directly from the covariance between mutations in columns of the MSA. Single-sequence language models do not have access to the MSA, and instead make predictions based on patterns seen during training. The MSA Transformer may use both covariance-based and pattern-based inference

1. https://www.biorxiv.org/content/biorxiv/early/2021/02/13/202...

That’s a really interesting way of looking at the difficulty of the problem that’s being solved. I’m curious, how do you arrive at the number 10^9?

Pretty cool to see the success of PacBio (Pacific Biosciences) technology for long-read sequencing. PacBio is one of the few successful sequencing technologies using nature’s canonical high-fidelity DNA reading tool (DNA polymerase). Other successful sequencing tech uses approaches that are (ingeniously) different from how DNA is read in a living cell.

PacBio circular consensus sequencing (used here) is a clever way of performing extremely accurate single-molecule reads: the target linear DNA is joined into a circle, which is read over and over again, enabling high accuracy of each base by consensus

If we say it’s nine rolls of two dice, and the probability of getting any particular pair of numbers in each roll is 1/36 (underestimate since we sometimes only care about the sum), then getting any specified sequence would be like (1/36)^9, around 1e-14. Of course then we’d have to get the gameplay right too