They just updated their arXiv paper in May 6, 2024, and still not show code repo in github. Will be time-consumed if we want to work bottom up, and may not be worthing it if it is not any better than RFDiffusion All-Atom.
This could be used for Ribosomopathies _(abnormalities in rRNA genes, ribosomal component proteins)_ in anaemia and bone marrow, to pre-emptively reverse engineer them by computer. So one makes a new, custom, ribosome organelle computationally. Comparison across genus and species could be for hypothetical empirical emulation _(and as an aside, sexual selection for immune system traits could be postulated, for inbreeding and outbreeding or phenotype)._ Cancer susceptibility and mitigating factors such as premature cell death _(not solely mutation from RNA interference)_ could be predicted. Extension to sickle cell prediction would be a worthwhile investigation. For instance: Narrow down and isolate the Mutation Preference Inference from RNA interference by means of singular value decomposition used in _(including Gaussian)_ noise elimination with a logistics fit, also by linear regression and derive if it is in a sparse matrix or dense matrix. Express the molecular Schrödinger electron density _(Gaussian probability estimation, in a Rosenblatt-Parzen window treating the kernel as its hypercube)_ as kissing-spheres _(or sphere packing via combinatorial optimisation)._ For learning _(educationally but also for machine-learning),_ that could be done in R (cran-project) for better graphics fo the polynomials but other than that, python would be similar enough language. OpenCL for heterogeneous low power computation _(low power implant medical devices for instance)_ might be an option although DLib and Eigen maths libraries for C++ would augment it. As an (extra) aside, discovery of the mechanisms of diploidisation in the (sans meiosis) parasexual cycle _(fungi and prokaryote, in mitosis)_ could be researched using singular value decomposition for theoretical soil emulation and fungi creation (or a plant) for antibiotic discovery. Then apply that to neofunctionalisation, subfunctionalisation and genome downsizing (post-polyploidisation diploidisation), for instance, as relevant to DNA repetition and gene deletion _(and extraneous gene copies with alleles in Eukaryota's taxonomic groups)._ My comment has no hate in it and I do no harm. I am not appalled or afraid, boasting or envying or complaining... Just saying. Psalms23: Giving thanks and praise to the Lord and peace and love. Also, I'd say Matthew6.
"Unreasonable usefulness of self-supervised" is right. It's actually weird that simply a higher likelihood per AA is enough to improve various proteins
As cutting edge as it gets. Amazing work. Very cool to hear about how by essentially solving an engineering problem and evolving PACE to PRANCE, Erika is more readily able to understand how and why the final protein variants are evolved. I read Kevin Esvelt's PACE paper in my first or second year of grad school and have been following the developments in this area ever since and it never fails to inspire me. The synthetic phylogenies that the PRANCE system can help generate will surely provide some incredible insights.
In your opinion the viewers : In what sense would it be interesting to train EvoDiff-seq (without modifying its architecture) on the PDB dataset of 200K sequences, these sequences that have been used to train RFDiffusion and the state of the art structure based generative models.
38:23 this is amazing work! I am very new to this. A naive question, for the last example, what info do you need to provide to the model to get the binder? Do you provide any guidance in terms of backbone info or any other minimal info?
Very nice -- does anyone know what the 'rep 1.2' means in her sampling method? I'm guessing 950 tokens with the highest probability were sampled 'randomly' (multinomially) and the rest were omitted? What does the 'rep 1.2' mean here?
A question: if I want to design 10000 sequences with ProteinMPNN, is it better to use one seed for all, or one seed per sequence to increase diversity? what do you think?
Great work. Very helpful to the community to encompass these exciting new evolutionary methods and combine with learning from the assay labeled data (so that we don't have to re-live the ordeal that you went through). Thanks also for sharing the person behind the work. The picture of one researcher looking at a big sea around them probably resonates with many. People can be more at ease after hearing about what it was like for you :)
Interesting talk. Do you think this method could be extended to predict peptide-protein or protein-protein interactions? If that's not the case could you explain what are the main limitations in going in that direction?
For the calculations to all work out and to be able to project the new sequence into space, the new sequence should have the same length as the sequences in the MSA. There are some tricks you can do depending on whether or not the new sequence is longer or shorter than the MSA length to “change” the sequence length… and that is definitely case-dependent.
Great talk! Love this series of seminars. It's indeed quite helpful to know that even very little filtering of the first batch of data or guidance by zero-shot methods can improve the frequency of hits so much.
