Add param_inference command to Docker commands

Author: niklases

README.md

@@ -40,7 +40,7 @@ Written by Niklas Siedhoff and Alexander-Maurice Illig.
 </p>
 
 Protein engineering by rational or random approaches generates data that can aid the construction of self-learned sequence-function landscapes to predict beneficial variants by using probabilistic methods that can screen the unexplored sequence space with uncertainty *in silico*. Such predictive methods can be applied for increasing the success/effectivity of an engineering campaign while partly offering the prospect to reveal (higher-order) epistatic mutation effects. Here we present an engineering framework termed PyPEF for assisting the unsupervised optimization, supervised training, and testing of protein fitness models for predicting beneficial combinations of (identified) amino acid substitutions using machine learning approaches.
-As training input, the developed framework requires the variant sequences and the corresponding screening results (fitness labels) of the variants as CSV files (or FASTA-Like ("FASL") data files following a self-defined convention). Using linear or nonlinear regression methods (partial least squares (PLS), Ridge, Lasso, Elastic net, support vector machines (SVR), random forest (RF), and multilayer perceptron (MLP)-based regression), PyPEF trains on the given learning data while optimizing model hyperparameters (default: five-fold cross-validation) and can compute model performances on left-out test data. As sequences are encoded using amino acid descriptor sets taken from the [AAindex database](https://www.genome.jp/aaindex/), finding the best index-dependent encoding for a specific test set can be seen as a hyperparameter search on the test set. In addition, one-hot and [direct coupling analysis (DCA)](https://en.wikipedia.org/wiki/Direct_coupling_analysis)-based feature generation are implemented as sequence encoding techniques, which often outperform AAindex-based encoding techniques. While one-hot encodings fail for positional extrapolation, DCA-based sequence encoding offers positional extrapolation capabilities and is hence better suited for most generalization tasks. In addition, a hybrid, combined model of the unsupervised and supervised DCA model provides even better performance and robust predictions, even when training with only a few data points (e.g. 50-100 variant fitness labels). Furthermore, a mixed hybrid DCA model combined with LLM models  predictions show even increased overall performance across the [ProteinGym](https://proteingym.org/) datasets tested. 
+As training input, the developed framework requires the variant sequences and the corresponding screening results (fitness labels) of the variants as CSV files (or FASTA-Like ("FASL") data files following a self-defined convention). Using linear or nonlinear regression methods (partial least squares (PLS), Ridge, Lasso, Elastic net, support vector machines (SVR), random forest (RF), and multilayer perceptron (MLP)-based regression), PyPEF trains on the given learning data while optimizing model hyperparameters (default: five-fold cross-validation) and can compute model performances on left-out test data. As sequences are encoded using amino acid descriptor sets taken from the [AAindex database](https://www.genome.jp/aaindex/), finding the best index-dependent encoding for a specific test set can be seen as a hyperparameter search on the test set. In addition, one-hot and [direct coupling analysis (DCA)](https://en.wikipedia.org/wiki/Direct_coupling_analysis)-based feature generation are implemented as sequence encoding techniques, which often outperform AAindex-based encoding techniques. While one-hot encodings fail for positional extrapolation, DCA-based sequence encoding offers positional extrapolation capabilities and is hence better suited for most generalization tasks. In addition, a hybrid, combined model of the unsupervised and supervised DCA model provides even better performance and robust predictions, even when training with only a few data points (e.g. 50-100 variant fitness labels). Furthermore, a mixed hybrid DCA model combined with LLM models  predictions show even increased overall performance across the [ProteinGym](https://proteingym.org/) datasets tested.
 
 Finally, the selected (un-) trained (pure or hybrid) model can be used to perform directed evolution walks *in silico* (see [Church-lab implementation](https://github.com/churchlab/UniRep) or the [reimplementation](https://github.com/ivanjayapurna/low-n-protein-engineering)) or to predict natural diverse or recombinant variant sequences that subsequently are to be designed and validated in the wet-lab.
 
@@ -83,6 +83,7 @@ Pull from Docker Hub or build the image using the stored [Dockerfile](./Dockerfi
   ```bash
   docker run --gpus=all -v ./datasets/:/datasets --workdir /datasets/AVGFP niklases/pypef:0.4.2 /bin/bash -c \
       "python /app/run.py mklsts --wt P42212_F64L.fasta --input avGFP.csv --ls_proportion 0.01 && \
+       python /app/run.py param_inference --msa uref100_avgfp_jhmmer_119.a2m --wt P42212_F64L.fasta && \
        python /app/run.py hybrid --ls LS.fasl --ts TS.fasl --params GREMLIN --llm prosst --wt P42212_F64L.fasta --pdb GFP_AEQVI.pdb"
   ```
 - building image from Dockerfile
@@ -93,6 +94,7 @@ Pull from Docker Hub or build the image using the stored [Dockerfile](./Dockerfi
   ```bash
   docker run --gpus=all -v ./datasets/:/datasets --workdir /datasets/AVGFP pypef /bin/bash -c \
       "python /app/run.py mklsts --wt P42212_F64L.fasta --input avGFP.csv --ls_proportion 0.01 && \
+       python /app/run.py param_inference --msa uref100_avgfp_jhmmer_119.a2m --wt P42212_F64L.fasta && \
        python /app/run.py hybrid --ls LS.fasl --ts TS.fasl --params GREMLIN --llm prosst --wt P42212_F64L.fasta --pdb GFP_AEQVI.pdb"
   ```