A new paper challenges how we have used protein language models (pLMs) for protein engineering. It suggests that using pLMs as antipredictors (try mutations the model thinks are bad) can generate the best leads for engineering novel function.
blog.genesmindsmachines.com/p/protein-la...
blog.genesmindsmachines.com/p/protein-la...
Protein language models are overly constrained by covariation
Why good performance on one set of goals can lead to poor performance on another
blog.genesmindsmachines.com
August 12, 2026 at 2:34 PM
A new paper challenges how we have used protein language models (pLMs) for protein engineering. It suggests that using pLMs as antipredictors (try mutations the model thinks are bad) can generate the best leads for engineering novel function.
blog.genesmindsmachines.com/p/protein-la...
blog.genesmindsmachines.com/p/protein-la...
We've been investing heavily in better protein language models (PLMs), but relatively little work addresses how to best generate with them. We present a new search-based method for PLMs and exhaustively benchmark models and methods, including with in vitro data from antibody therapeutics campaigns.🧵
March 13, 2026 at 1:24 PM
We've been investing heavily in better protein language models (PLMs), but relatively little work addresses how to best generate with them. We present a new search-based method for PLMs and exhaustively benchmark models and methods, including with in vitro data from antibody therapeutics campaigns.🧵
Quero sumir por sla, 3 meses plms
September 6, 2024 at 10:48 PM
Quero sumir por sla, 3 meses plms
send pet pics plms <3 long hard week already lol
November 25, 2025 at 10:50 PM
send pet pics plms <3 long hard week already lol
o yuki merecia tanto essa vaga 💔
espero q plms o liam mostre resultados
espero q plms o liam mostre resultados
December 19, 2024 at 7:04 PM
o yuki merecia tanto essa vaga 💔
espero q plms o liam mostre resultados
espero q plms o liam mostre resultados
Protein Language Models (PLMs) could transform outbreak response
New research from @kieranlamb.bsky.social & colleagues show that PLMs can identify mutation hotspots and key features of viral proteins - even from a single sequence.
www.gla.ac.uk/research/az/...
New research from @kieranlamb.bsky.social & colleagues show that PLMs can identify mutation hotspots and key features of viral proteins - even from a single sequence.
www.gla.ac.uk/research/az/...
March 30, 2026 at 10:00 AM
Protein Language Models (PLMs) could transform outbreak response
New research from @kieranlamb.bsky.social & colleagues show that PLMs can identify mutation hotspots and key features of viral proteins - even from a single sequence.
www.gla.ac.uk/research/az/...
New research from @kieranlamb.bsky.social & colleagues show that PLMs can identify mutation hotspots and key features of viral proteins - even from a single sequence.
www.gla.ac.uk/research/az/...
🧵
[1/n] Does AlphaFold3 "know" biophysics and the physics of protein folding? Are protein language models (pLMs) learning coevolutionary patterns? You can try to guess the answer to these questions using mechanistic interpretability.
[1/n] Does AlphaFold3 "know" biophysics and the physics of protein folding? Are protein language models (pLMs) learning coevolutionary patterns? You can try to guess the answer to these questions using mechanistic interpretability.
December 12, 2024 at 10:50 PM
🧵
[1/n] Does AlphaFold3 "know" biophysics and the physics of protein folding? Are protein language models (pLMs) learning coevolutionary patterns? You can try to guess the answer to these questions using mechanistic interpretability.
[1/n] Does AlphaFold3 "know" biophysics and the physics of protein folding? Are protein language models (pLMs) learning coevolutionary patterns? You can try to guess the answer to these questions using mechanistic interpretability.
Awesome paper. Simple post-hoc trick (averaging over homologs) with elegant evolution theory dramatically improves zero shot coding variant effect prediction in pLMs & actually delivers better results from the larger models (inverting the trend with raw likelihoods). 1/
From Likelihood to Fitness: Improving Variant Effect Prediction in Protein and Genome Language Models https://www.biorxiv.org/content/10.1101/2025.05.20.655154v1
May 26, 2025 at 9:40 AM
Awesome paper. Simple post-hoc trick (averaging over homologs) with elegant evolution theory dramatically improves zero shot coding variant effect prediction in pLMs & actually delivers better results from the larger models (inverting the trend with raw likelihoods). 1/
Queria morar nesse prédio. Plms teria entretenimento
July 16, 2025 at 12:29 AM
Queria morar nesse prédio. Plms teria entretenimento
There it is again: using PLMs to predict antigen-epitope interactions from sequence alone yields a prediction accuracy of 0.65, in line with a proposed upper limit from a previous study QTed below (from doi.org/10.1101/2025.02.12.637989; I deleted an older version of this post due to typos/errors)
March 3, 2025 at 5:45 AM
There it is again: using PLMs to predict antigen-epitope interactions from sequence alone yields a prediction accuracy of 0.65, in line with a proposed upper limit from a previous study QTed below (from doi.org/10.1101/2025.02.12.637989; I deleted an older version of this post due to typos/errors)
🧬 What are protein language models (PLMs) actually learning about biology? Our paper introduces InterPLM - a framework that reveals interpretable features in PLMs using sparse autoencoders, giving us a window into how these models represent protein structure and function.
🧵(1/8)
🧵(1/8)
November 19, 2024 at 7:36 PM
🧬 What are protein language models (PLMs) actually learning about biology? Our paper introduces InterPLM - a framework that reveals interpretable features in PLMs using sparse autoencoders, giving us a window into how these models represent protein structure and function.
🧵(1/8)
🧵(1/8)
Apparently this tendency of PLMs to cut and paste across multiple sequnces during de novo design mimics what we see in natural sequences. Not sure what to make of it yet
July 31, 2025 at 2:16 PM
Apparently this tendency of PLMs to cut and paste across multiple sequnces during de novo design mimics what we see in natural sequences. Not sure what to make of it yet
Acaba essa corrida plms
November 23, 2025 at 5:11 AM
Acaba essa corrida plms
•introduced “zero shot prediction” as a question of guessing a bioassay’s outcome by likelihoods of pLMs
•commented on biases in evolutionary signals from Tree of life used to train pLMs (a favorite paper I read in 2024: shorturl.at/fbC7g)
•commented on biases in evolutionary signals from Tree of life used to train pLMs (a favorite paper I read in 2024: shorturl.at/fbC7g)
December 16, 2024 at 6:29 AM
•introduced “zero shot prediction” as a question of guessing a bioassay’s outcome by likelihoods of pLMs
•commented on biases in evolutionary signals from Tree of life used to train pLMs (a favorite paper I read in 2024: shorturl.at/fbC7g)
•commented on biases in evolutionary signals from Tree of life used to train pLMs (a favorite paper I read in 2024: shorturl.at/fbC7g)
This plot is quite the indictment of fine-tuned PLMs, showing how performance is entirely data-dependent and, at the upper end of performance, equally achievable with randomized model weights
February 26, 2026 at 6:24 PM
This plot is quite the indictment of fine-tuned PLMs, showing how performance is entirely data-dependent and, at the upper end of performance, equally achievable with randomized model weights
Fui escrever um desabafo e fiz O poema(eu achei plms)
January 10, 2025 at 8:03 AM
Fui escrever um desabafo e fiz O poema(eu achei plms)
plms isso aqui tapa um buraco😭
September 4, 2024 at 11:01 PM
plms isso aqui tapa um buraco😭
keirajosh plms 🥺
October 6, 2025 at 7:45 PM
keirajosh plms 🥺
Have we hit a "scaling wall" for protein language models? 🤔 Our latest ProteinGym v1.3 release suggests that for zero-shot fitness prediction, simply making pLMs bigger isn't better beyond 1-4B parameters. The winning strategy? Combining MSAs & structure in multimodal models!
May 8, 2025 at 12:29 AM
Have we hit a "scaling wall" for protein language models? 🤔 Our latest ProteinGym v1.3 release suggests that for zero-shot fitness prediction, simply making pLMs bigger isn't better beyond 1-4B parameters. The winning strategy? Combining MSAs & structure in multimodal models!
🧬 What do deep-learning models for proteins actually learn?
Our new review looks at how model predictions relate to fitness, folding stability and function.
With @cwjpugh.bsky.social, Mafalda Dias & @jonnyfrazer.bsky.social
🔗 chemrxiv.org/doi/full/10....
Our new review looks at how model predictions relate to fitness, folding stability and function.
With @cwjpugh.bsky.social, Mafalda Dias & @jonnyfrazer.bsky.social
🔗 chemrxiv.org/doi/full/10....
From sequences and structures to fitness, folding and function: challenges in the age of AI | ChemRxiv
Deep-learning models have transformed our ability to predict the phenotypic effects
of sequence perturbations. Protein language models (pLMs) score the evolutionary propensity
of any amino-acid substi...
chemrxiv.org
September 25, 2026 at 1:01 PM
🧬 What do deep-learning models for proteins actually learn?
Our new review looks at how model predictions relate to fitness, folding stability and function.
With @cwjpugh.bsky.social, Mafalda Dias & @jonnyfrazer.bsky.social
🔗 chemrxiv.org/doi/full/10....
Our new review looks at how model predictions relate to fitness, folding stability and function.
With @cwjpugh.bsky.social, Mafalda Dias & @jonnyfrazer.bsky.social
🔗 chemrxiv.org/doi/full/10....
plms preciso de moots de f1
August 29, 2024 at 3:43 AM
plms preciso de moots de f1
Through homology search & pLMs, we identified an effective kynureninase that degrades a key immunosuppressor in cancer, reducing tumor weight in mice by 3.4x.
📄 www.biorxiv.org/content/10.1...
Seek & rank your own protein based on only a handful of measures.
⛵ seekrank.steineggerlab.com
📄 www.biorxiv.org/content/10.1...
Seek & rank your own protein based on only a handful of measures.
⛵ seekrank.steineggerlab.com
January 25, 2024 at 2:42 AM
Through homology search & pLMs, we identified an effective kynureninase that degrades a key immunosuppressor in cancer, reducing tumor weight in mice by 3.4x.
📄 www.biorxiv.org/content/10.1...
Seek & rank your own protein based on only a handful of measures.
⛵ seekrank.steineggerlab.com
📄 www.biorxiv.org/content/10.1...
Seek & rank your own protein based on only a handful of measures.
⛵ seekrank.steineggerlab.com
Key observation: If you compare prediction performance to a naive model that simply predicts the mean (known) fitness value at each site, pLMs barely win out, and on viral data they do worse! In other words, taking the mean value at each site does better than using the 650 million parameter pLM. 2/
March 11, 2026 at 3:18 PM
Key observation: If you compare prediction performance to a naive model that simply predicts the mean (known) fitness value at each site, pLMs barely win out, and on viral data they do worse! In other words, taking the mean value at each site does better than using the 650 million parameter pLM. 2/