This is a challenge, even for someone who has professionally used the breadth of proteins. I really like the test. I'm actually kind of surprised at my own pulling on knowledge to make a guess - it's an orthogonal way to think about the question than is usually posed.
I wonder if there's a way to ease the difficulty by filling in 'correct' features of the guesses: if your guess is a 'transmembrane' then it reveals that as a property. On the other hand, I don't think the annotations are clean enough - and are often designed for 'at all' rather than 'primary' features. For one of the examples, once I noticed it was an adhesion protein, it would have been interesting to sift through classes or cell types as opposed to just continuing to shoot in the dark based on the structure alone.
I presume you're showing even the 'low confidence' portions of the predicted structure? Please do.
You could also show the primary amino acid sequence too - there's a weird familiarity with those given how often the structures themselves have historically not been so accessible. BLASTING each of the guesses would be another interesting thing to see.
> I wonder if there's a way to ease the difficulty by filling in 'correct' features of the guesses
Rather than allowing players to guess individual features, I opted for the "highlight" system where all hidden features that match your guessed protein's features get auto-revealed. This way, if you suspect a transmembrane protein, you can just guess a known transmembrane protein and see which features auto-reveal.
> it would have been interesting to sift through classes or cell types
You're welcome to suggest databases with good coverage over the proteome that I could use for these.
> I presume you're showing even the 'low confidence' portions of the predicted structure?
Yes, any residues in the files I fetch get rendered. I rank by coverage before fetching.
> You could also show the primary amino acid sequence too
Great work! I really like the interface here, I think you put a lot of taste into the way this all came together. It's quite hard (appropriately so!) I would love to see a experienced structural biologist go at this.
Thank you. I started the development with the interface first, basically making a mockup of how the finished game should look like, and then prodding LLMs with a stick to make a backend that would support this interface without crashing.
Making a nice-looking web GUI without knowing relevant vocabulary was a very clunky process in comparison to code pipelines, basically just pasting screenshots into the chat window and asking LLMs to "line up stuff properly", which they still couldn't manage to do in places.
Huh. Is this kind of manifold something that a group of humans can now identify on sight and associate with a part of a DNA sequence? That's pretty spectacular. I have a friend who worked for years on AlphaFold, but wasn't aware that people had gotten to this level of confidence in visually identifying proteins.
My dream is forming a team of Tetlock-style superbiologists who can identify gene names on sight like Rainbolt, beat prediction markets on shorting biotech stocks, and smell out pre-cancerous cells like it's the final round of amongus.
Hahahaa ... I kinda love this, along with humans who can sit at a piano and play a song they've only heard once, or cook a complicated dish they only tasted once. That seems like the great Turing test... zero-shot humans.
But seriously, there probably are a few people who can see genes from proteins like that, faster than a whole datacenter of GPUs. Putting together such a brain trust could be invaluable.
Cool game. Harder for me. I have background in most of the undergrad and some grad molecular biology. Still, the hints didn't help much. The most useful hints were on function of the gene and cellular components. Also, I have been out of touch with gene names from the last few years. I tried searching using a gene's function but didn't work as much. At the end I got 81% close but far from the answer.
You can play more than once in the practice mode. It turns on after you complete the daily puzzle.
There's almost 20k entries in my human gene database. You can choose any of them as your guess, but the "gene of the day" is chosen from a subset that has decent 3D structure coverage (meaning, not falling back to alphafold2 for visualization).
I wonder if there's a way to ease the difficulty by filling in 'correct' features of the guesses: if your guess is a 'transmembrane' then it reveals that as a property. On the other hand, I don't think the annotations are clean enough - and are often designed for 'at all' rather than 'primary' features. For one of the examples, once I noticed it was an adhesion protein, it would have been interesting to sift through classes or cell types as opposed to just continuing to shoot in the dark based on the structure alone.
I presume you're showing even the 'low confidence' portions of the predicted structure? Please do.
You could also show the primary amino acid sequence too - there's a weird familiarity with those given how often the structures themselves have historically not been so accessible. BLASTING each of the guesses would be another interesting thing to see.
> I wonder if there's a way to ease the difficulty by filling in 'correct' features of the guesses
Rather than allowing players to guess individual features, I opted for the "highlight" system where all hidden features that match your guessed protein's features get auto-revealed. This way, if you suspect a transmembrane protein, you can just guess a known transmembrane protein and see which features auto-reveal.
> it would have been interesting to sift through classes or cell types
You're welcome to suggest databases with good coverage over the proteome that I could use for these.
> I presume you're showing even the 'low confidence' portions of the predicted structure?
Yes, any residues in the files I fetch get rendered. I rank by coverage before fetching.
> You could also show the primary amino acid sequence too
I'll consider it.
Making a nice-looking web GUI without knowing relevant vocabulary was a very clunky process in comparison to code pipelines, basically just pasting screenshots into the chat window and asking LLMs to "line up stuff properly", which they still couldn't manage to do in places.
But seriously, there probably are a few people who can see genes from proteins like that, faster than a whole datacenter of GPUs. Putting together such a brain trust could be invaluable.
You could start with popular gene names.
GeneGuesserIt
There's almost 20k entries in my human gene database. You can choose any of them as your guess, but the "gene of the day" is chosen from a subset that has decent 3D structure coverage (meaning, not falling back to alphafold2 for visualization).