Sequence generation

Generating and optimizing sequences with sequence models, evolutionary profiles, or hallucination, including general sequence-search strategies. Fixed-structure sequence design goes under design/inverse-folding.

Quartz 4

Explorer

Sequence generation

Prefusion stabilization often requires multiple mutations, but single substitutions can suffice

Unconditional design does not consistently improve with model scale

Distograms can be used as loss functions for protein design by hallucination

Stronger diffusion guidance reduces diversity of generated outputs

Inversion of protein folding neural networks

Selectivity, specificity, and conformational constraints can be encoded in loss functions during protein design

Joint sequence-structure diffusion or flow matching models show superior performance when designing structure at the beginning and sequence near the end

Base PLMs must usually be fine-tuned to generate functionally active sequences

Conformationally stabilizing mutations can be identified from sequence clusters that are predicted to fold into that conformation when passed to AlphaFold2

Coupling sidechain and main chain prediction or design does not always lead to improvements

Guided sequence-only diffusion outperforms autoregressive fine-tuned PLMs on designing high-fitness sequences

Larger PLMs generate more novel sequences from more sparsely populated protein families

PLM-designed sequences match the distribution of fitness values, lengths, and structure prediction confidence of natural sequences

PLMs generate more divergent functional proteins than Potts models

ProGen

Protein backbones designed by diffusion, but not by language models, have more secondary structure

Protein backbones designed using diffusion, but not sequence-based models, have fewer beta sheets

Protein language models and PLM-based structure prediction generalize to de novo designed proteins

Protein language models can be steered to design proteins with specific properties

Sequence perplexity

De novo designed proteins (from language models) and naturally occurring proteins borrow heavily from other functionally similar sequences

Gibbs sampling is less effective than simpler algorithms for sequence design using PSSMs

High-pLDDT designs can be insoluble

Including optimal growth temperature during pre-training of PLMs improves prediction and design of thermostability

Rosetta