update-docs

DOCS.md

@@ -1,97 +1,47 @@
-# Example Instance Datasets
+# Demo Instance Lab Dataset
 
-This folder contains three example datasets that would be available to the user for downstream model training. The derived data is calculated from
-log probabilities of each library candidate in an assay sample. An assay sample is defined by the library sequenced at different stages of the assay. The different stages are:
+This repo contains three example datasets that would be available to the user for downstream model training. The derived data is calculated from log probabilities of each library candidate in an assay sample. The different data modalities are:
 
-- Bacterial Outgrowth
-- Phage Expression
-- 1 nM Dose of Target
-- 10 nM Dose of Target
-- 100 nM Dose of Target
-- 1000 nM Dose of Target
-- Blank Dose of Target
+- Expression
+- Binding
+- Specificity
 
-The dosing stages are specific to a target molecule, so the dataset will include rows for each target that was tested for the library.
+Expression scores are log enrichments of the candidate translated protein frequency relative to the initial DNA pool candidate frequency. The binding score is defined as the ratio of the relative enrichment of a candidate over the enrichment of the weakest candidate in the dataset represented on a log scale. The specificity score is defined in terms of binding scores of off-target and on-target antigens, and is the ratio of the targeting binding score divided by the average to the binding scores of all off-target interactions.
 
-The base dataset contains the log probabilities used to derive the relative dissociation constants and expression levels contained in the other datasets.
+## Common Metadata
 
-## Base Log Probability Dataset
+There are columns that are common to each of the three datasets, which facilitate simple associations if one wants to compute features over multiple different scores. The columns are:
 
-This dataset contains metadata columns that are useful for filtering and grouping the data. The columns are:
+- `candidate_library_id`: This ID refers to the specific sample that contains all the translated proteins expressed together. In binding and specificity contexts, this ID represents pool of binder proteins that is mixed with different antigens to generate binding and specificity scores.
+- `candidate_name`: The user-provided name of the candidate sequence.
+- `candidate_sequence`: The amino acid sequence of the candidate binder.
+- `candidate_id`: A cryptographic hash of the candidate sequence.
+- `candidate_type`: An enum field representing the type of candidate protein. In this demo, it's either `miniprotein` or `vhh`.
+- `target_pdb_id`: The PDB ID of the target protein that conditioned the model to generate the candidate sequence.
 
-- `experiment_name`: The name of the experiment/library you ordered.
-- `target_name`: The name of the target molecule tested.
-- `candidate_name`: The name of the generated protein candidate in the library.
-- `candidate_id`: A unique identifier that is specific to the candidate sequence.
-- `experiment_replicate_num`: Any experiment/library can be replicated resulting in repeat samples that are run through the same assay workflow. This value is used to distinguish between experimental replicates.
-- `clone_num`: The index associated with a clone of a library candidate. This gets used to find means and variances for each library candidate.
+## Expression Dataset
 
-The dataset contains sequence information for each library candidate clone. The columns are:
+The unique identifiers for each row in this dataset is `["candidate_library_id", "candidate_id"]`. So, the number of rows are equal to the number of candidate proteins in each distinct pool. The scoring columns are:
 
-- `candidate_sequence`: The designed sequence of the library candidate from the order form.
-- `clone_sequence`: The consensus sequence of the library candidate clone validated by NGS.
+- `expression_score`: The log enrichment of the expression of the translated candidate relative to the double stranded pool.
+- `expression_score_std`: The standard error of the expression score estimate calculated from each of the replicates measured.
 
-Finally, the dataset contains the log probabilities of each library candidate clone at each stage of the assay. The columns are:
+## Binding Dataset
 
-- `bacterial_outgrowth_log_prob`: The log probability of the library candidate clone being expressed after transformation and bacterial outgrowth.
-- `phage_expression_log_prob`: The log probability of the library candidate clone being expressed in phage.
-- `dose_1nM_log_prob`: The log probability of the library candidate clone being expressed at 1 nM.
-- `dose_10nM_log_prob`: The log probability of the library candidate clone being expressed at 10 nM.
-- `dose_100nM_log_prob`: The log probability of the library candidate clone being expressed at 100 nM.
-- `dose_1000nM_log_prob`: The log probability of the library candidate clone being expressed at 1000 nM.
-- `dose_blank_log_prob`: The log probability of the library candidate clone being expressed at a blank dose.
+The unique identifiers for each row in this dataset is `["candidate_library_id", "candidate_id", "antigen_pdb_id"]`. So, the number of rows are equal to the number of candidate protein + antigen combinations in each distinct library. The columns specific to the binding dataset are:
 
-We also know the standard deviation of each of the log probability estimates. The columns are:
+- `antigen_pdb_id`: The PDB ID of the antigen protein used in the binding assay.
+- `targeting`: Boolean value that specifies whether the antigen used in the binding assay matches the target for the candidate protein.
+- `binding_score_50nM`: The mean binding score from the assay where the candidate library was tested with 50 nM of the antigen protein.
+- `binding_score_500nM`: The mean binding score from the assay where the candidate library was tested with 500 nM of the antigen protein.
+- `binding_score_50nM_std`: The standard error of the mean 50 nM binding score estimate calculated from each of the replicates measured.
+- `binding_score_500nM_std`: The standard error of the mean 500 nM binding score estimate calculated from each of the replicates measured.
 
-- `bacterial_outgrowth_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed after transformation and bacterial outgrowth.
-- `phage_expression_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed in phage.
-- `dose_1nM_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed at 1 nM.
-- `dose_10nM_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed at 10 nM.
-- `dose_100nM_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed at 100 nM.
-- `dose_1000nM_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed at 1000 nM.
-- `dose_blank_log_prob_std`: The standard deviation of the log probability of the library candidate clone being expressed at a blank dose.
+## Specificity Dataset
 
-## pKD Dataset
+The unique identifiers for each row in this dataset is `["candidate_library_id", "candidate_id"]`. So, the number of rows are equal to the number of candidate proteins in each distinct pool. This columns specific to the specificity dataset are:
 
-This dataset contains the relative dissociation constants for each library candidate clone. The dissociation constants are relative
-to the average dissociation constant of the entire library. Knowledge of absolute dissociation constants of any library candidate
-would set the scale and allow derivation of absolute dissociation constants for each library candidate. The relative dissociation
-constants are averaged over all clones of a library candidate producing a single value for each library candidate.
-
-The metadata columns are nearly the same as the base dataset. The `clone_num` column is removed as it was used to calculate the relative
-dissociation constants. The metadata columns columns are:
-
-- `experiment_name`
-- `target_name`
-- `candidate_name`
-- `candidate_id`
-- `experiment_replicate_num`
-
-The rest of the dataset would be values that are directly consumed by model training. The columns are:
-
-- `candidate_sequence`
-- `pK_D_relative`
-- `pK_D_relative_std`
-
-## Candidate Expression Dataset
-
-This dataset contains the expression levels for each library candidate. The metadata columns are the same as the pKD dataset. The columns are:
-
-- `experiment_name`
-- `target_name`
-- `candidate_name`
-- `candidate_id`
-- `experiment_replicate_num`
-
-The rest of the dataset would be values that are directly consumed by model training. The columns are:
-
-- `candidate_sequence`
-- `log_enrichment_expression`
-- `log_enrichment_expression_std`
-
-## Note on dataset size
-
-The base dataset is going to be larger than the pKD or candidate expression datasets because it contains rows for each clone of each library candidate. There are typically ~100 clones of each library candidate.
-So, the number of rows in the base dataset is N_candiates * N_targets * N_replicates * N_clones.
-
-The pKD and expression datasets are smaller because they average over the clones of each library candidate. So, the number of rows becoms N_candidates * N_targets * N_replicates.
+- `log_specificity_score_50nM`: The mean specificity score from the assay where the candidate library was tested with 50 nM of the antigen protein.
+- `log_specificity_score_500nM`: The mean specificity score from the assay where the candidate library was tested with 500 nM of the antigen protein.
+- `log_specificity_score_50nM_std`: The standard error of the mean 50 nM specificity score estimate calculated from each of the replicates measured.
+- `log_specificity_score_500nM_std`: The standard error of the mean 500 nM specificity score estimate calculated from each of the replicates measured.