Just gonna leave this here.

https://www.biorxiv.org/content/10.1101/2024.02.29.582810v1

Tl;dr: DNA is NOT all you need.

most of the examples in that paper (a single paper) show that DNA is nearly all you need, with the rest being RNA.

RNA is an obvious example. The examples and benchmarks they give in the paper are not the straw men the DNA LLMs are beating the stuffing out.

Also CRE activity is highly cell type specific. This article is a pretty awesome demonstration of model guided design of cell type specific cis regulatory elements.

https://www.biorxiv.org/content/10.1101/2023.08.08.552077v1

An LLM would not be able to do this because DNA itself contains no contextual information about cell type - every cell has a copy of the full genome. Epigenetic tracks however contain a lot of information germane to the cellular context - ex which parts of the genome are being transcribed.

but epigenetics is just DNA. it's state information stored directly in the DNA, or in directly attached machinery. from the perspective of learned models, those are just other features.

But realistically, the right source for transcription is the RNA in the cell, not the epigenetics. Nearly all cell type profiling is based on RNA. It's far easier and more reliable to interrogate the transcriptome than to try to gain info from epigenetic states.

Epigenetics is not just DNA, think of it more like the (hidden) state of DNA. Histone modifications and open chromatin and other epigenetic readouts are like emissions / indicators of the hidden state.

The relationship is like that between the words in a book and the page that is actively being read. I know that is a hackneyed analogy; but coffee is wearing off :)

Those are all readable using standard DNA sequencing techniques, so again, it's just state attributes of the DNA.

(I've worked in genomics for 30+ years. I'm not just spitballing here).

DNA libraries are read out using sequencing techniques. But I doubt anyone would say they are just measuring DNA. It’s kind of like saying a luciferase based assay is just measuring the intensity of light.

yes, that is the terminology we use.

I tend to agree (the cell being in control and all the 4D interactions and epigenetics mechanisms etc), but out of curiosity, what would you say we also need?

For starters, chemical environment modeling. But also cells differentiate, so in any system you need to understand the differentiation, and how those differentiated cells will change the environment of other cells, based on the environment they encounter.

That's not to say you can't glean a ton from DNA, but there are some external inputs we may simply never know enough about to incorporate into the model. Ultimately DNA IS all you need...if you have perfect environmental information.

The article I posted shows what is working better - the Olga Troyanskaya / David Kelley style models. There was another one (Kundaje group?) recently that used Hi-C data.

I think you are missing what the Evo project is trying to do -- create a new prokaryotic genome through a generative model. How this would work would be like the earlier hand-made synthetic genomes like Synthia (Gibson et al, 2010).

In such a system you would take an existing bacterial cell and replace its genome with the newly synthesized version. The proteins and other molecules from the existing cell would remain (before eventually being replaced) and serve to "boot" the new genome.

It's an interesting endeavor, but there are some obvious safety concerns.

Within Prokaryotes, there is a lot of horizontal gene transfer. What if some of the synthetic sequences get into other organisms and spread out?

Those genes would have to confer an evolutionary advantage or they would immediately be discarded/selected against. The chances of that happening are nil… we’re not going to come up with something more useful to bacteria than billions of years of natural selection. Synthetic biology with organisms produced to generate small molecules useful for humans is widely practiced but has the opposite problem- all of the engineering changes to the cells are constantly being selected against, and revert on their own.

Sounds cool, but how you define success for something like that? I can copy a prokaryotic genome mutated based as some non-zero rate and it would probably be viable. Is that synthetic enough to count? Are they going for a minimal genome?

How about something more useful, lucrative, and easy to define success for like engineering a morphine synthesis pathway into E. coli or something.

Imo, if you are talking about synthetic biology, then their training data is insufficient. Synthetic bio explores a lot of design space that is far outside of anything you would see in nature. There the secret sauce would not be in the generative pretraining, but in the RL. Unfortunately bio experiments are noisy, slow, and expensive so good luck getting enough data before the heat death of the universe.

https://onlinelibrary.wiley.com/doi/10.1002/bies.201300153 tl;dr: metabolism is all you need.

while potentially interesting work, very shortsighted and premature to say this is a "GPT" moment in biology. ML people in bio need to think hard not only about what they are doing, but why are they are doing it (other than this is cool and will lead to a nice Nature publication). Their basic premise (learning from DNA is the next grand challenge in biology) is shaky. Imo, the grand challenge in biology is determining what the grand challenge is, and that is a deep scientific/philosophical question.

DNA is all you need? In the future generative AI will generate You!

Would be interesting to see what comes of it.

As you progress along the following chain: genomics-->proteomics->interactomics->metabolomics, our understanding becomes blurrier and challenges harder.

I built the wrapper/playground [0] linked in the article. Feel free to give feedback here or by the email in my bio

[0] https://evo.nitro.bio/

Hi Nishant. Great work, as always.

Thanks for the kind words :)

I'm a little disappointed that their linked preprint doesn't appear to include any molecular biology; i.e. they don't actually try to synthesize any of their predicted sequences and test function. It wouldn't be an outrageous synthesis task to make some of the CRISPR-Cas sequences they generated.

Also interesting that AlphaMisense is omitted from Figure 2B; it substantially outperforms the ESM-based ESM1b in our hands. But I guess the idea is that this is a general-purpose DNA language model whereas AlphaMissense is domain-specific for variant effect prediction?

Strong second for wishing they tried physically testing some model output. The importance of "model that makes outputs AlphaFold thinks look like Cas" is very different from "model that makes functional Cas variants".

For design tasks like in this paper, I think computational models have a big hill to climb in order to compete with physical high-throughput screening. Most of the time the goal is to get a small number of hits (<10) out of a pool of millions of candidates. At those levels, you need to work in the >99.9% precision regime to have any hope of finding significant hits after multiple-hypothesis correction. I don't think they showed anything near that accurate in the paper.

Maybe we'll get there eventually, but the high-throughput techniques in molecular biology are also getting better at the same time.

You are correct that it is dangerous to rely on the results of a model being an oracle for another model, extremely good models (say F=ma) are used all the time.

My immediate thought. Big Claims without backing.

Your model makes predictions. Prove they’re worth salt.

This should really be a requirement for bio type related generative methods rather than a nice-to-have. A very high percentage of compounds generated by genai type methods have been shown not to work as intended. Anything without wetlab validation should really be taken with a large grain of salt