James X. Li

VisuMap Technologies.
Discovering knowledge by visualization.

[James X. Li @VisuMap]

@TripInChina 80

[James X. Li @VisuMap]

@timfduffy Against commonly accepted "wisdom", t-SNE normally converges to the same embedding up to rotation, regardless of initialization/seed. The key issue is that you need to reduce the exaggeration gradually to 1.0 during the training.

[James X. Li @VisuMap]

Their main findins: they measured the folding time of 8 proteins, is took 1 to 4 milliseconds.

[nature @Nature]

Scientists say they have made some of the first direct measurements of how long it takes an individual, ordinary protein to fold – and the results were surprising.

https://t.co/gs7UUNBLke

[James X. Li @VisuMap]

@hxiao The perplexity parameter for t-SNE is likely too small in this demo. If you choose the value 500, instead of the default value 50, the resulting embedding will have better cluster separation.

[James X. Li @VisuMap]

Ideally t-SNE embedding grows, during its learning process, from a tine spot to a short bar that then elongates, splits and expands to a cluster of various shapes with gradient boarder. The following is an embedding of 6K proteins based on their 3D coordinates:

[James X. Li @VisuMap]

t-SNE & Initialization. Because the use of exaggerated learning phase its embeddings normally converge to a single final map (upto rotaion). This makes dedicated initialization largely irrelevant. For this reason, VisuMap's implementation only supports random initialization.

[James X. Li @VisuMap]

t-SNE & learning-rate. A key feature of t-SNE is its use of adaptive-gains for each optimized variables. This make the learning-rate parameter largely irrelevant. For this reason, VisuMap's implementation has a fixed learning-rate (500), users don't have to care about it.

[James X. Li @VisuMap]

@old_crone_code @mudscryer t-SNE should converge to an unique embedding up to a rotation, if you choose a large enough perplexity; and reduce the exaggeration slowly from say 5.0 to 1.0. You can try the implementation of t-SNE in VisuMap.

[James X. Li @VisuMap]

Protein Atlas with t-SNE: We can vectorize amino-acid chains with Fourier-Transform and embed them with t-SNE into 2D space. Here is a map of ca. 100K AA-chains. The map reveals the similarity between the 3D structures.

[James X. Li @VisuMap]

https://t.co/nNGaI8XXBo

[The Nobel Prize @NobelPrize]

This year’s chemistry laureates Demis Hassabis and John Jumper have developed an AI model, AlphaFold2, to solve a 50-year-old problem: predicting proteins’ complex structures.

Check out two examples of protein structures determined using AlphaFold2. First up, a bacterial enzyme that causes antibiotic resistance. The structure is important for discovering ways of preventing antibiotic resistance.

Animation: ©Terezia Kovalova/The Royal Swedish Academy of Sciences

#NobelPrize

[James X. Li @VisuMap]

UnFolding protein with t-SNE: The following video clip shows the 3D structure of a protein complex (6EMK, https://t.co/kdtUf451qH). The t-SNE embedding shows clearly the C2 symmetry between upper and lower halves.

[James X. Li @VisuMap]

t-SNE for un-folding protein 3D structure: When we add the sequential index as the forth dimension to the 3D coordinates of polypeptides, t-SNE can produce 2D maps which unfold their 3D structures; and reveal more sub-clusters with their distinguished shapes.

[James X. Li @VisuMap]

Apart from an overview about the whole 3D structure, a 2D map also captures local structures like alpha helixes; and large structures like symmetries and replicating patterns.

[James X. Li @VisuMap]

t-SNE for 2D protein maps: In order to facilitate the exploration of 3D folding structure of protein polymers, we can use t-SNE to create 2D maps for protein polymers. Here are the steps:

[James X. Li @VisuMap]

@MicTott With approperite parameters, t-sne mostly convergs to an unique embedding, upto rotations. If you get different embeddings with different initionalizations, it indicats rather that the settings of t-sne are not properly choosen.

[James X. Li @VisuMap]

@wats_updog @datepsych The main purpose of t-SNE is to preserve local and global distances/similarity, and it worked better than most other MDS algorithms. The sentence "it proves nothing" or "distance is meaningless" is meaningless itself; and not even-wrong.

[James X. Li @VisuMap]

@datepsych Most critics here are pretty unsubstantiated. It would be however helpful to find 3 to 5 features, which contribute mainly to the cluster separation. You can just examine the average values of small selected region; and move the region from on cluster to the another one.

[大漠孤鹰⛓ @LCA176232]

新华网在B站关于姜萍的视频“探索数学世界只是姜萍的PlanB”已经被删除了。

[James X. Li @VisuMap]

T-sne was the true breakthrough from 3 dozens of nonlinear dim. reduction methods in 2008. bhSne, fastSne and umap are sad "optimizations" which ignore large distances, and made it more or less useless for complex data. https://t.co/cvzwYOonXC