Sudheshna_week11

week11.md

@@ -0,0 +1,20 @@
+https://www.nature.com/articles/s41586-021-03819-2
+
+
+Highly accurate protein structure prediction with AlphaFold
+
+Proteins are essential to life and understanding their structure can help us understand the functions they perform. The alpha is a deep novel neural network architecture to predict the 3D proetin structure computationally. It uses physical, evolutionary and geomteric constraints of protein strcutures that can predict a protein structure highly accurately
+
+The alpha fold directly predicts the 3D coordinates of all heavy atoms for a given protein using the primary amino acid sequence and aligned sequences of homologues as inputs
+ The trunk called the Evoform in which the repeated neural network produce an (No of sequences * No of residues) array that presesnts MSA and (No of residues * No of residue ) array that represents residue pairs.
+
+The network comrpises two stages: 
+First stage: The trunk called the Evoform in which the repeated neural network produce an (No of sequences * No of residues) array that presesnts MSA and (No of residues * No of residue ) array that represents residue pairs. The operation operates on a concrete 3D backbone structure, prioritizes back bone such that side chains are highly and constrained and peptides are not. Next they generate inavriant point attention (a geometry of 3D positions) and assign key value pairing with 3D points that are produced in local frame of each residue such that the final value is invariant to global rotations and translations.
+Predictions of side-chain χ angles as well as the final, per-residue accuracy of the structure (pLDDT) are computed with small per-residue networks on the final activations at the end of the network.
+The resulting Nframes × Natoms distances are penalized with a clamped L1 loss. This creates a strong bias for atoms to be correct relative to the local frame of each residue and hence correct with respect to its side-chain interactions, as well as providing the main source of chirality for AlphaFold
+They use supervised learning technique
+to train to enhance accuracy
+
+To test the network, they trained a seperate structure module and ran the 48 evoformer blocks in the network which provided trajectory of 192 intermediate strcutures. The resulting trajectories are smooth and indicating alphafold makes constant incremental improvements to the structure until it no longer can improve. Alphafold functions far more efficiently compared to other algorithms from a limited data in the PDB and also is able to cope with the complexity and variety of structural data.
+
+PS: I might be wrong with my understanding of how the evoform is working but I'm still trying to understand the concept in depth.

week12.md

@@ -0,0 +1,17 @@
+COVID-view: Diagnosis of COVID-19 using Chest CT
+
+The researchers have developed a COVID-19 specific 3D view of the chest CT in order help radiologist get an second level opnion to diagnose and observe other problems in the chest area along with COVID diagnosis.
+They provide the following results through their pipeline
+    1) Automatic segemntation of lungs: They incorporated Hofmanniger et al. lungs segmentation model that was trained on a lrge variety of dieseased lungs, encompassing different lesions and abnormalities with air pockets, tumors, and effusions, and includes COVID-19 data. The algorithm seemed to work and did not show any deviations from true broundary of lungs even with the outliers.
+
+    2) localization of lung abnormalities: The system provides segmentation outlines in 2D views for identifying and examining regions of abnormalities, which helps in drawing radiologists attention to regions that need more attention 
+
+
+    3) Classification: The pipeline uses multiple instance learning, the goal is to train a model to predict labels of unseen bags. The algorithm uses the training sets contain bags where ecah bag is composed of set of instances. In MIL, only the bag level is known and the instance level label is unknown and it suits for medical imaging becuase here only a patient level or imager level label is known
+
+    4) Visual interface:  This interface allows the radiologists to visualize chest CT in both 2D and 3D views that enhances the radiologists workflow by providing them with a contemporary discrement of lesions and their morphology for a better diagnosis
+
+    5) Evaluation: The data of 580 CT scans, out of which 343 positivi ethrough RT PCR and 237 were negative were collected. The model was implemented using pytorch and the performance of the model was evaluated using 5 fold cross validation. For the detetction of accuracy, the ROC curve and AUC curve were plotted with a 95% confidence interval. Th results were (ROC) curve (AUC) are 0.952( 95% confidence interval (CI): 0.938, 0.966) and 0.985 (95% CI: 0.981, 0.989), respectively. The sensitivity and specificity are 0.953 (95% CI: 0.932, 0.974) and 0.949 (95% CI: 0.928, 0.97), respectively
+
+The researchers, I think were successful in developing a model that high accuracy and reliability to diagnose COVID-19 and helps radiologists as second reader. 
+