Published version

This version corresponds to the published version of DeepBLAST

Version 1.0.1 minor release

We improved model loading to make this software more portable. Minor typos have been fixed

Version 1

This is the latest release of DeepBLAST that enables strongly comparable structural alignments compared to existing structural aligners, without needing structural inputs. Multiple changes happened since the last iteration

• We swapped out the Bepler LSTM pretrained model with ProTrans
• CNN intermediate layers to estimate the BLOSUM matrices are used instead of MLP layers
• Support for differentiable Smith-Waterman is in place
• Compatibility with TM-vec to enable scalable remote homology search is in place
• Distributed computing across multiple nodes and GPUs is enabled

Alpha Release

We are excited to announce our first release of DeepBLAST. Here, we provide the first implementation of the differentiable Needleman-Wunsch algorithm to enable learning protein structural similarity.

We have provided command line scripts on how to pretrain these models and load them.