Latest mind-shifting thoughts, updates, ideas, and more
Viet Phan
This is how you can apply reinforcement learning for stock picking:
Simulate synthetic data from normal distribution
Normalize features
Put features to agent
Observe state action and receive feedback
8:39 AM · Jun 3, 2023
000
Viet Phan
Step up from word2vec are
Transformers that GPT models uses
Unfortunately, today we will explore only encoders (Bert), not decoders.
Good for predicting fixed output.
7:53 AM · Jun 2, 2023
000
Viet Phan
NLP approach = basically word2vec
Calc cosine similarity and find current molecule given surrounding molecules (context)
Sounds pretty naive to me. Does it really work?
7:47 AM · Jun 2, 2023
000
Viet Phan
How molecules can be transformed to text embeddings.
Fascinating 🧐
7:44 AM · Jun 2, 2023
000
Viet Phan
Protein sequences are like language words
Hence NLP models can be trained
Like CC(C)C=C1
Need to rehash my chemistry knowledge 😅
7:40 AM · Jun 2, 2023
000
Viet Phan
Find you own drug with these public datasets.
Unlabeled and unstructured.
7:38 AM · Jun 2, 2023
000
Viet Phan
Basic moat of Machine learning drug discovery.
- Find molecule
- Score molecules
- Optimize molecule combinations
Seems simple and brutr force applicable
7:36 AM · Jun 2, 2023
000
Viet Phan
Current state of the art model: AlphaFold from google.
Using only NVIDIA 200 GPUs.
No wonder NVIDIA rally is real.
7:31 AM · Jun 2, 2023
000
Viet Phan
Disease = protein malfunction
Malfunction = misbalance in production
Cure = drug
Drug = molecules binding to protein and tries to change its behaviour
🤯
7:27 AM · Jun 2, 2023
000
Viet Phan
90 pct chance of failure on drug discovery trials in 4-5 years timeframe
Not bad odds.
Will rethink strategy on betting on biotech🧐
7:21 AM · Jun 2, 2023
000
