Msp Remote Alignment
Shanae Maerz
I bumped into a few posts from people who run remote observatories, where they claim that they don't mess with PA for weeks, sometimes months! I understand that guiding will eventually (try to) correct most (all?) PA imperfections, but how is that even remotely optimal? How can one get away with such horrible PA when doing long exposures?
On one hand you have countless threads where people strive for "perfect" PA using Polemaster and/or SharpCap Pro and on the other hand you have people with remote observatories who don't bother with good PA.
The whole point about having an observatory is that the equipment doesn't move. You get the polar alignment right to begin with, and it doesn't ever change. If you are not packing the mount up and storing it at the end of a session, there is no reason why the alignment would change.
No. When you are accurately polar aligned, the polar axis is not pointed at Polaris (which is about 40 arcmins from the pole), it is pointed at a stationary point in the sky (the North Celestial Pole). If you are well polar aligned initially, the mount's polar axis will still be pointing at the stationary NCP six months later, even though Polaris will be making a small 40 arcminute circle about the NCP during that time. Guiding should be just as hard or easy as when you initially set-up assuming your polar alignment is unchanged.
Really, you are assuming that polar alignment drifts a lot more than it in fact does. Yes, frost can shift things, and some things settle. But, in a balanced mount, properly polar aligned, on a big, well mounted pier and all that, there is really very little, if any, shift.
My area is prone to earthquakes -- so even minor ones might shift things a bit at some point, requiring re-alignment. Long-term the ground may settle a bit more (over more than 1.5 years), or I may have to remove the mount for cleaning or lubricating, at which point I'll have to re-align.
I have a small law firm of 4 computers. on the receptionist computer, when I remote in with the default syncro remote control feature, the cursor clicks about half an inch to the right of where the cursor is on the screen. very annoying. I have contacted syncro support, they didnt have an answer. any one seen this? the only thing different on this computer is a java applet that lets the receptionist see voicemails. everything else is identical.
The FRAS is an alignment system comprising almost one thousand sensors distributed along the 200 metres of new magnets that are installed on either side of the ATLAS and CMS detectors. The sensors, along with their electronics and software programs and a system of motorised jacks on which some of the components will sit, will be used to adjust the relative positions of the components remotely, in real time, with no human intervention needed in the cavern. This is a crucial requirement in the irradiated environment of the HL-LHC. Although a similar remote-controlled system has already been deployed along 50 metres of magnets inside the current accelerator, the FRAS, which will cover a greater distance, is innovative in several respects. The alignment tolerance will be the same as it is now (+/- 0.15 mm) but the system will cover a much greater distance (200 m instead of 50 m).
When protein structures are available, structural alignment tools such as TM-align15, Dali18, FAST19 and Mammoth20 can provide a measure of structural similarity by aligning protein structures via superposition15,18,20,21,22. Although this approach can provide a measure of structural similarity in low-sequence-similarity scenarios, there are two major limitations. First, protein structures are not available for most proteins. Despite the rapid advances made by AlphaFold2, there remains a large gap between known protein sequences and predicted protein structures23. In metagenomics samples alone, 2.4 billion24 to 68 billion25 unique proteins have been observed, highlighting the small percentage of proteins with known structures. Furthermore, AlphaFold2 has limited utility in the context of predicting structures for proteins with short sequences26. Work on structure prediction that uses single or few homologous sequences is ongoing, but most methods exhibit reduced accuracy and take substantial time and memory resources per sequence, limiting scaling to genomic protein databases.
Given the rapid growth of protein structure databases, most existing structural alignment tools are far too computationally intensive to run at scale, requiring brute-force all-versus-all comparisons to query structurally similar proteins. Although there are emerging tools for scalable homology search on structural databases27, as well as for embedding proteins for either search or alignment28,29 (Table 1), tools that perform explicit structural similarity search and alignment on large protein sequence databases are also needed.
We showcase the merits of TM-Vec models in the context of CATH37 and SWISS-MODEL38 to show how our tool can scale with regard to database size while maintaining high precision in identifying structurally similar proteins. Our benchmarks suggest that TM-Vec can extrapolate beyond known fold space, and we contrast TM-Vec with AlphaFold2 (ref. 10), OmegaFold39 and ESMFold40 in a case study where TM-Vec can distinguish between bacteriocin classes more accurately than AlphaFold2, OmegaFold and ESMFold in combination with TM-align11. We also showcase the merits of DeepBLAST on different remote homology benchmarks, demonstrating that language model embeddings can capture more of the structural basis for alignment than purely sequence-based alignment. TM-Vec and DeepBLAST are broadly applicable tools that have the potential to enable the structural (and structural-similarity-based) annotation of proteins and their functions in the vast and growing biodiversity contained in protein sequence collections.
Our contributions are twofold: (1) we introduce a framework to perform scalable structure-aware search, TM-Vec, that affords substantial improvements in speed and sensitivity41 (Fig. 1 and Supplementary Fig. 1); and (2) we introduce a differentiable sequence alignment algorithm, DeepBLAST, that performs structural alignments (Supplementary Fig. 2).
The basis of DeepBLAST is to predict the structural alignments of proteins by training models on proteins with both sequences and structures available. Our alignment strategy uses recent developments in differentiable dynamic programming and protein language models to predict the structural alignments given by TM-align for pairs of protein sequences (Supplementary Fig. 2).
We showcase the ability of DeepBLAST to extract structural alignments from remote homologs on the Malidup43 and Malisam44 structure databases compared with existing alignment algorithms. Furthermore, we evaluate the ability of TM-Vec to perform remote homology search on the CATH37, SWISS-MODEL38, Malidup43 and Malisam44 structure databases. Finally, we showcase the merits of using TM-Vec in tandem with DeepBLAST in the context of the BAGEL bacteriocin database45.
The challenge of applying our proposed structural alignment algorithm to large-scale protein databases is the demanding runtime requirements. Each DeepBLAST structural alignment takes on the order of milliseconds and scales linearly with database size, making structural alignment searches on large databases impractical. To mitigate this issue, we developed TM-Vec, a model that is designed to efficiently query structurally similar proteins. Our strategy relies on the construction of twin neural networks, whose purpose is to provide per-protein vectors for fast indexing. The cosine distance of these vectors approximates the TM-score between pairs of proteins. This model can then be applied to entire protein databases to create an index over all the protein vectors. The resulting database can be efficiently queried in \(O(\log ^2n)\) time for n proteins41, providing sublinear scaling to retrieve structurally similar proteins based on their TM-score.
Our method DeepBLAST uses sequence alone; we do not supply the atomic coordinates of either protein to the algorithm after training it. To form a common reference for an optimal alignment, we focused on two gold-standard benchmark sets comprising manually curated structural alignments, named Malisam44 and Malidup43. Manual structure alignment is intuitive human assessment, typically emphasizing 3D overlap and topology preservation, as those features are easier to visualize than a plethora of local alignments and contacts53,54,55. All methods tend to agree when the problem is trivial owing to near sequence identity and near structural identity. Therefore, the most valuable gold-standard alignment benchmark set is where the dataset members have low sequence identity as well as varied degrees of structural similarity. Our benchmarks were performed on the curated Malisam44 and Malidup43 protein structural alignment benchmarking datasets (which are heavily skewed towards difficult-to-detect, low-sequence-identity remote homology).