Duplicate from aalphabio/open-alphaseq

Co-authored-by: David Noble <dnoble@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,68 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text
+YM_005.csv filter=lfs diff=lfs merge=lfs -text
+YM_1068.csv filter=lfs diff=lfs merge=lfs -text
+YM_549.csv filter=lfs diff=lfs merge=lfs -text
+YM_693.csv filter=lfs diff=lfs merge=lfs -text
+YM_852.csv filter=lfs diff=lfs merge=lfs -text
+YM_985.csv filter=lfs diff=lfs merge=lfs -text
+YM_988.csv filter=lfs diff=lfs merge=lfs -text
+YM_989.csv filter=lfs diff=lfs merge=lfs -text
+YM_990.csv filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,3 @@
+scratch/
+archive/
+.DS_Store

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 A-Alpha Bio, Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,194 @@
+---
+tags:
+- chemistry
+- biology
+pretty_name: A-Alpha Bio open source data
+size_categories:
+- 10K<n<10M
+configs:
+  - config_name: YM_0005
+    data_files: "data/YM_0005/data.parquet"
+  - config_name: YM_0549
+    data_files: "data/YM_0549/data.parquet"
+  - config_name: YM_0693
+    data_files: "data/YM_0693/data.parquet"
+  - config_name: YM_0852
+    data_files: "data/YM_0852/data.parquet"
+  - config_name: YM_0985
+    data_files: "data/YM_0985/data.parquet"
+  - config_name: YM_0988
+    data_files: "data/YM_0988/data.parquet"
+  - config_name: YM_0989
+    data_files: "data/YM_0989/data.parquet"
+  - config_name: YM_0990
+    data_files: "data/YM_0990/data.parquet"
+  - config_name: YM_1068
+    data_files: "data/YM_1068/data.parquet"
+---
+# Open Protein–Protein Interaction Affinity Datasets with AlphaSeq
+
+Protein-protein interactions (PPIs) are fundamental to countless biological processes. One of the most informative biophysical properties of a PPI is the binding affinity: the strength of how two proteins interact. Yet, despite its importance, publicly available affinity data remains limited, constraining the development and benchmarking of protein modeling methods.
+
+Our high-throughput yeast mating assay, [AlphaSeq](https://www.pnas.org/doi/10.1073/pnas.1705867114), enables the quantitative measurement of PPIs at scale (often generating libraries up to 1M interactions per experiment!).
+
+To help bridge the gap between experimental affinity measurements and computational protein models, we’re open-sourcing a selection of our datasets.
+
+Each dataset captures the results of a yeast mating experiment between two protein libraries—one of binders and one of targets. Detailed experimental context and metadata are provided in the accompanying data cards.
+
+## Quickstart
+
+Load one dataset:
+
+```python
+# Replace with the specific experiment ID you want to load
+from datasets import load_dataset
+ds = load_dataset("aalphabio/open-alphaseq", "YM_0549", split="all")
+# From here, cast it to pandas / polars dataframe for analysis if desired
+df = ds.to_pandas()
+
+# ...or load directly with pandas
+import pandas as pd
+ym_id = "YM_0549"
+df = pd.read_parquet(f"hf://datasets/aalphabio/open-alphaseq/data/{ym_id}/data.parquet")
+print(df)
+```
+
+Fetch multiple experiments at once:
+
+```python
+from datasets import load_dataset, concatenate_datasets
+
+repo = "aalphabio/open-alphaseq"
+cfgs = ["YM_0549", "YM_1068"]
+
+# Load datasets and add experiment column
+datasets_list = [load_dataset(repo, cfg, split="all") for cfg in cfgs]
+datasets_labeled = [
+    ds.add_column("experiment", [cfg] * len(ds))
+    for ds, cfg in zip(datasets_list, cfgs)
+]
+combined = concatenate_datasets(datasets_labeled)
+print(combined)
+```
+
+Use different experiments for modeling splits:
+
+```python
+from datasets import load_dataset, DatasetDict
+
+repo = "aalphabio/open-alphaseq"
+ds = DatasetDict({
+    "train": load_dataset(repo, "YM_0693", split="all"),
+    "test": load_dataset(repo, "YM_0988", split="all"),
+})
+print(ds)
+```
+
+## Experiment glossary
+
+| YM ID | Description | Details |
+|:--------|:-------------|:---------|
+| YM_0005 | Anti-CoV Ab Panel x CoV2-RBD Mutagenesis | [details](./data/YM_0005/README.md) |
+| YM_0549 | VHH72 Optimization Variants Iter0 x CoV2-RBD | [details](./data/YM_0549/README.md) |
+| YM_1068 | VHH72 Optimization Variants Iter1 x CoV2-RBD | [details](./data/YM_1068/README.md) |
+| YM_0693 | PP489 Optimization Variants Iter0 x TIGIT | [details](./data/YM_0693/README.md) |
+| YM_0988 | PP489 Optimization Variants Iter1 x TIGIT | [details](./data/YM_0988/README.md) |
+| YM_0852 | Pembrolizumab-scFv Optimiziation Variants Iter0 x PD-1 | [details](./data/YM_0852/README.md) |
+| YM_0985 | Pembrolizumab-scFv Optimiziation Variants Iter1 x PD-1 | [details](./data/YM_0985/README.md) |
+| YM_0989 | Trastuzumab-scFv CDR3 Optimization Variants Iter0 x HER-2 | [details](./data/YM_0989/README.md) |
+| YM_0990 | Trastuzumab-scFv CDR+FW Optimization Variants Iter0 x HER-2 | [details](./data/YM_0990/README.md) |
+
+## References
+
+This dataset collection builds on the AlphaSeq high-throughput yeast mating assay technology and includes data from multiple publications:
+
+### Core Technology
+
+- Younger, D., Berger, S., Baker, D., & Klavins, E. (2017). High-throughput characterization of protein-protein interactions by reprogramming yeast mating. Proceedings of the National Academy of Sciences of the United States of America, 114(46), 12166–12171. [https://doi.org/10.1073/pnas.1705867114](https://doi.org/10.1073/pnas.1705867114)
+
+### Dataset Sources
+
+**YM_0005:**
+
+- Engelhart, E., Lopez, R., Emerson, R., Lin, C., Shikany, C., Guion, D., Kelley, M., & Younger, D. (2022). Massively multiplexed affinity characterization of therapeutic antibodies against SARS-CoV-2 variants. Antibody therapeutics, 5(2), 130–137. [https://doi.org/10.1093/abt/tbac011](https://doi.org/10.1093/abt/tbac011)
+
+**YM_0549, YM_0693, YM_0852, YM_0985, YM_0988, YM_0989, YM_0990, YM_1068:**
+
+- Agarwal, A. A., Harrang, J., Noble, D., McGowan, K. L., Lange, A. W., Engelhart, E., Lahman, M. C., Adamo, J., Yu, X., Serang, O., Minch, K. J., Wellman, K. Y., Younger, D. A., Lopez, R. M., & Emerson, R. O. (2025). AlphaBind, a domain-specific model to predict and optimize antibody-antigen binding affinity. mAbs, 17(1), 2534626. [https://doi.org/10.1080/19420862.2025.2534626](https://doi.org/10.1080/19420862.2025.2534626)
+
+## Dataset schema
+
+| Column | Description |
+|:--------|:-------------|
+| **mata_description** | Unique description of each MATa library element. Also contains negative control strains (ANeg1, ANeg2, ANeg3). Often supplemented with assay-specific information such as nomenclature or replicate labeling. |
+| **matalpha_description** | Unique description of each MATα library element. Also contains negative control strains (AlphaNeg1, AlphaNeg2, AlphaNeg3). Often supplemented with assay-specific information such as nomenclature or replicate labeling. |
+| **mata_sequence** | Amino acid sequence corresponding to the MATa library element. The sequence excludes linkers, epitope tags, and the cell wall anchor Aga2, and is empty for negative controls. |
+| **matalpha_sequence** | Amino acid sequence corresponding to the MATα library element. The sequence excludes linkers, epitope tags, and the cell wall anchor Aga2, and is empty for negative controls. |
+| **alphaseq_affinity** | Pairwise interaction affinity, reported as log₁₀(estimated Kd in nM). Values map as –1 = 0.1 nM, 0 = 1 nM, 1 = 10 nM, 2 = 100 nM, etc. Missing values indicate no observed cell fusions between the protein pair (see Younger et al., 2017). |
+| **affinity_upper_bound** | Upper (stronger) bound of the 95 % confidence interval for `alphaseq_affinity`, computed with the Wilson score interval. Provided even when no point estimate is available. |
+| **affinity_lower_bound** | Lower (weaker) bound of the 95 % confidence interval for `alphaseq_affinity`, computed with the Wilson score interval. |
+| **normalized_affinity** | Z-score comparing each interaction’s affinity to the distribution of all interactions involving the same MATa or MATα protein. Computed by successive normalization of the affinity matrix (Olshen & Rajaratnam 2010). More negative values indicate higher specificity (e.g., –2 ≈ 2 SD below the mean). Strongly negative values are not expected when many proteins bind broadly. |
+| **above_background** | Boolean flag indicating whether the interaction affinity is significantly stronger than the assay background (q < 0.05, Benjamini–Hochberg corrected t-test). |
+| **sufficient_replicate_observations** | Boolean flag indicating whether the interaction was observed in > 50 % of replicate samples. A value of False suggests a potentially spurious or unreliable interaction. |
+
+## FAQ
+
+Please see below for some clarifying details:
+
+### Where can I find experimental details for each dataset?
+
+Each dataset has a corresponding README.md in its subfolder summarizing the experiment's goals, library composition, and citation info. See `./data/YM_0005/README.md` for an example.
+
+### What kind of sequences are in the library?
+
+While not a strict rule, the A-libraries typically contain designed sequences, while the Alpha-libraries contain corresponding targets of interest. Historically, we’ve used VHHs or scFvs in the A-library and antigen targets in the Alpha-library. Each dataset will have a card that details specific information of the individual assay run. When building or training models, note that PPIs can generally be treated as symmetric. However, members within the same library may share sequence, functional, or structural similarities. Also, some models are sensitive to input order — so ensure that (A, Alpha) pairs are treated consistently between training and testing.
+
+### Why are there duplicate PPIs in the dataset?
+
+Some datasets include technical replicates, often for the wild-type (“WT”) or parent sequence in mutation studies. Replicates help capture the experimental and biological variation in measured affinities. This can be useful for analyses that assess the statistical significance of observed affinity difference, such as identifying how much a vaiant changes binding strength relative to a parent protein.
+
+### What is considered a strong or good binder?
+
+Affinity measurements are reported in log10 Kd nM (a value of 0 indicates 1 nM, 3 is 1 uM, 5 is 100 uM). Lower values indicate stronger binding. In practice, we often compare relative affinities - for example, assessing differences in binding strength as a target interface is mutated, or comparing variant binders to their parent.
+
+### The dataset has NaN values in the affinity, why?
+
+Not all PPIs form detectable interactions; weak or non-binding interactions may result in no paired barcode reads, yielding NaN values. For these cases, it may be more useful to look at the lower or upper bound affinities to help interpret the range of possible affinity within the assay.
+
+### What do Iter0 / Iter1 mean?
+
+Iter0 and Iter1 (abbreviated for Iteration) are our nomenclature for describing antibody variant libraries designed for antibody optimization campaigns. Iter0 libraries are generated in a zero-shot fashion (ie random mutations or selections by ESM) while Iter1 libraries are generated with models trained on Iter0 datasets.
+
+### How should I cite this dataset?
+
+To properly acknowledge this work, please cite:
+
+1. **This dataset repository:**
+
+   ```text
+   A-Alpha Bio (2025). Open Protein–Protein Interaction Affinity Datasets with AlphaSeq.
+   https://huggingface.co/datasets/aalphabio/open-alphaseq
+   ```
+
+2. **The AlphaSeq technology paper:**
+
+   ```text
+   Younger, D., Berger, S., Baker, D., & Klavins, E. (2017).
+   High-throughput characterization of protein–protein interactions by reprogramming yeast mating.
+   Proceedings of the National Academy of Sciences, 114(46), 12166-12171.
+   https://doi.org/10.1073/pnas.1705867114
+   ```
+
+3. **The specific experiment paper(s) for the dataset(s) you use:**
+   - **If using YM_0005:** Cite the CoV epitope mapping paper (Engelhart et al., 2022) - see [References](#references) section
+   - **If using any other experiments (YM_0549, YM_0693, YM_0852, YM_0985, YM_0988, YM_0989, YM_0990, YM_1068):** Cite the AlphaBind paper (2025) - see [References](#references) section
+
+For complete citation details and DOI links, see the [References](#references) section above.
+
+### Can I use this dataset for model training or benchmarking?
+
+Yes — the dataset is released fully open source, and is suitable for both academic and commercial use.
+
+### Who can I contact with questions or feedback?
+
+Feel free to email maintainers [Natasha Murakowska](mailto:nmurakowska@aalphabio.com) or [David Noble](mailto:dnoble@aalphabio.com). We will host a discussions tab for open discourse as well.

data/YM_0005/README.md

@@ -0,0 +1,25 @@
+# Anti-CoV Ab Panel x CoV2-RBD Mutagenesis (YM_0005)
+
+## Overview
+
+YM_0005 is a Covid receptor-binding domain (RBD) single-site mutagenesis (SSM) library against a panel of 33 ScFvs.
+
+## Experimental details
+
+We studied the effects of a panel of ScFvs against COVID RBD. In this dataset, our panel of ScFvs are tested in 2 orientations: LH for light-heavy chains and HL heavy-light chains. We explore the local landscape of the RBD in tandem with Covid for epitope mapping by mutating each position. We include a control of ACE2, as the protein of most therapeutic antibodies disrupt.
+
+This dataset includes 62 unique scFvs and 2431 unique RBD sequences.
+
+A more extensive methods section can be found in our publication [here](https://academic.oup.com/abt/article/5/2/130/6584706).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+- Binder descriptor. HL / LH denotes orientation of the scFv as expressed on the yeast surace.
+
+### Alpha-library
+
+- Any `matalpha_description` that contains the term "WT" can be assumed as a replicate of the WT target. Other sequences will be indicated by their original residue, position of mutation, and mutated residue (i.e. S23A is Ser -> Ala mutation in position 23).

data/YM_0005/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1581ed43276e35fa548e78c3f1f05ca35897e0fe2ae4ddf155c2b1875f97b260
+size 3207875

data/YM_0549/README.md

@@ -0,0 +1,33 @@
+# VHH72 Optimization Variants Iter0 x CoV2-RBD (YM_0549)
+
+## Overview
+
+YM_0549 consists of VHH72 designs against SARS CoV-2 RBD. We wanted to perturb the local landscape of VHH72, so we performed mutagenesis to observe the sensitivity of VHH-72 to different RBD targets. This dataset is great to study the local sensitivity of mutations relative to parent "WT" VHH72.
+
+## Experimental details
+
+This dataset includes 29765 unique VHHs and 8 unique RBD sequences. The alpha-library is the same as YM_1068.
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `wt_<i>`: These are WT replicates.
+- `candidate_`: Various mutations of VHH-72
+
+### Alpha-library
+
+- `SARS-CoV2_RBD_(6LZG)`
+- `WIV1`
+- `LYRa11`
+- `MERS_RBD`
+- `OMICRON_BA2_TStarr_Seq&Trunc`
+- `SARS-COV-2_GAMMA_RBD_LLNL_Trunc`
+- `SARS-CoV-2_Delta_RBD_L452R_T478K`
+- `SARS-CoV1_RBD_(6LZG_Mimic)`

data/YM_0549/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ea7338eaba293661ba920d422f53292365bcb90c9a59db1daaa87a01d1af8b9
+size 7286376

data/YM_0693/README.md

@@ -0,0 +1,27 @@
+# PP489 Optimization Variants Iter0 x TIGIT (YM_693)
+
+## Overview
+
+YM_693 is a dataset of anti-TIGIT designs against TIGIT. This dataset contains only 2 targets, but they are species homologs of human and mouse. The designs offer a few mutations to study the local interaction between  and TIGIT. This is a dataset to explore relative affinities to the parent for antibody optimization.
+
+## Experimental details
+
+We studied the efficacy of generating binders with different model hyperparameters. This dataset includes 26726 unique scFvs and 2 unique target sequences.
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `wt_<i>`: These are WT replicates.
+- `candidate_`: Various mutations of PP489
+
+### Alpha-library
+
+- `TIGIT_22-137_POI-AGA2`: Human TIGIT
+- `TIGIT_Mouse`: Mouse TIGIT

data/YM_0693/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b82f9bb39c835ec160ca192ef0e1ba165a0d6651c520791081e800c485c907b
+size 3124450

data/YM_0852/README.md

@@ -0,0 +1,29 @@
+# Pembrolizumab-scFv Optimiziation Variants Iter0 x PD-1 (YM_0852)
+
+## Overview
+
+YM_852 are Pembrolizumab mutations against its native target; we introduce mutations that include deletions, insertions, double and single mutations to quantify the sensitivity of the local landscape in engaging the native target.
+
+## Experimental details
+
+This dataset includes 29883 unique scFvs and 1 unique native target sequence. A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `WT_<i>`: These are WT replicates.
+- `del_<i>`: A deletion from the WT
+- `ins_<i>`: An insertion from WT
+- `pair_<i>`: A double mutation (2 residues mutated from WT)
+- `single_<i>`: A single mutation
+
+To get the mutations of interest relative to the parent, we recommend an alignment to the WT sequence.
+
+### Alpha-library
+
+There is only 1 sequence, which is the native target.

data/YM_0852/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:caf6b0903e944b9f102bd9ff4d3b81e764ad5e5b669fa058f155f414b3972ce0
+size 3679609

data/YM_0985/README.md

@@ -0,0 +1,31 @@
+# Pembrolizumab-scFv Optimiziation Variants Iter1 x PD-1 (YM_0985)
+
+## Overview
+
+YM_0985 includes Alphabind designs against PD-1. We explored several model hypothesis: (i) Does pre-training aid predicitivity and (ii) does the featurization of the input sequences matter. To test pretraining, we refer to `mata_descriptions` with the term **warm** to include pretraining, and **cold** to start from a randomly initialized seed. For featurization, we explored **label-encoded** sequences with a one-hot-encoder of amino acid identities, versus an **ESM-featurized** embedding to represent each sequence in the PPI.
+
+## Experimental details
+
+We studied the efficacy of generating binders with different model hyperparameters. This dataset includes 34890 unique VHHs and 1 unique RBD sequences.
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `Pembro144_WT_<i>`: These are WT replicates.
+- `Pembro144_label_encoded_cold`: Label encoded sequences with no pretraining
+- `Pembro144_label_encoded_warm`: Label encoded sequences with pretraining
+- `Pembro144_esm_cold`: ESM featurized sequences with no pretraining
+- `Pembro144_esm_warm`: ESM featurized sequences with pretraining
+
+To get the mutations of interest relative to the parent, we recommend an alignment to the WT sequence.
+
+### Alpha-library
+
+There is only 1 sequence, which is the native target.

data/YM_0985/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95a95293298ab05fffa0e8e7120987b070175946b24257edfd120e364a18699c
+size 4153247

data/YM_0988/README.md

@@ -0,0 +1,30 @@
+# PP489 Optimization Variants Iter1 x TIGIT (YM_0988)
+
+## Overview
+
+YM_0988 includes ABC001 against 2 TIGIT homologs. We explored several model hypothesis: (i) Does pre-training aid predicitivity and (ii) does the featurization of the input sequences matter. To test pretraining, we refer to `mata_descriptions` with the term **warm** to include pretraining, and **cold** to start from a randomly initialized seed. For featurization, we explored **label-encoded** sequences with a one-hot-encoder of amino acid identities, versus an **ESM-featurized** embedding to represent each sequence in the PPI. Optimization was performed on the human ortholog.
+
+## Experimental details
+
+We studied the efficacy of generating binders with different model hyperparameters. This dataset includes 26726 unique scFvs and 2 unique target sequences.
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `ABC001_WT_<i>`: These are WT replicates.
+- `ABC001_label_encoded_cold`: Label encoded sequences with no pretraining
+- `ABC001_label_encoded_warm`: Label encoded sequences with pretraining
+- `ABC001_esm_cold`: ESM featurized sequences with no pretraining
+- `ABC001_esm_warm`: ESM featurized sequences with pretraining
+
+### Alpha-library
+
+- `TIGIT_22-137_POI-AGA2`: Human TIGIT
+- `TIGIT_Mouse`: Mouse TIGIT

data/YM_0988/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f40be4e87167cf8cc2e154196d7b8e4076542ab51d3ac1bdc87febcda3b46fe9
+size 5372635

data/YM_0989/README.md

@@ -0,0 +1,29 @@
+# Trastuzumab-scFv CDR3 Optimization Variants Iter0 x HER-2 (YM_0989)
+
+## Overview
+
+YM_0989 are Trastuzumab designs against HER2. We explored several model hypothesis: (i) Does pre-training aid predicitivity and (ii) does the featurization of the input sequences matter. To test pretraining, we refer to `mata_descriptions` with the term **warm** to include pretraining, and **cold** to start from a randomly initialized seed. For featurization, we explored **label-encoded** sequences with a one-hot-encoder of amino acid identities, versus an **ESM-featurized** embedding to represent each sequence in the PPI.
+
+## Experimental details
+
+We studied the efficacy of generating binders with different model hyperparameters. This dataset includes 20828 unique VHHs and 1 unique sequences. All designs are limited to the CDRs of the proteins of interest
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `TrastuzumabCDR_WT_<i>`: These are WT replicates.
+- `TrastuzumabCDR_label_encoded_cold`: Label encoded sequences with no pretraining
+- `TrastuzumabCDR_label_encoded_warm`: Label encoded sequences with pretraining
+- `TrastuzumabCDR_esm_cold`: ESM featurized sequences with no pretraining
+- `TrastuzumabCDR_esm_warm`: ESM featurized sequences with pretraining
+
+### Alpha-library
+
+There is only 1 target of interest.

data/YM_0989/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f39740cd68213b970d1ada3fb71f82d6e35db5957f6c1df0bca37aed7848f6b5
+size 3132363

data/YM_0990/README.md

@@ -0,0 +1,28 @@
+# Trastuzumab-scFv CDR+FW Optimization Variants Iter0 x HER-2 (YM_0990)
+
+## Overview
+
+YM_0990 are Trastuzumab designs against HER2. We explored several model hypothesis: (i) Does pre-training aid predicitivity and (ii) does the featurization of the input sequences matter. To test pretraining, we refer to `mata_descriptions` with the term **warm** to include pretraining, and **cold** to start from a randomly initialized seed. For featurization, we explored **label-encoded** sequences with a one-hot-encoder of amino acid identities, versus an **ESM-featurized** embedding to represent each sequence in the PPI.
+## Experimental details
+
+We studied the efficacy of generating binders with different model hyperparameters. This dataset includes 20828 unique VHHs and 1 unique sequences. In this experiment, we enable designs to span a window that encompasses the frameworks and CDR regions.
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (scFvs)
+
+There are several terms you can filter by:
+
+- `TrastuzumabCDR_WT_<i>`: These are WT replicates.
+- `TrastuzumabCDR_label_encoded_cold`: Label encoded sequences with no pretraining
+- `TrastuzumabCDR_label_encoded_warm`: Label encoded sequences with pretraining
+- `TrastuzumabCDR_esm_cold`: ESM featurized sequences with no pretraining
+- `TrastuzumabCDR_esm_warm`: ESM featurized sequences with pretraining
+
+### Alpha-library
+
+There is only 1 target of interest.

data/YM_0990/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7534c38c36b2d2ea6fa22964ce7d37922a7ffcd9ca314eff3b2aee71233b89fd
+size 4751930

data/YM_1068/README.md

@@ -0,0 +1,36 @@
+# VHH72 Optimization Variants Iter1 x CoV2-RBD (YM_1068)
+
+## Overview
+
+YM_1068 consists of VHH72 variants measured against SARS CoV-2 RBD. We explored several model hypothesis: (i) Does pre-training aid predicitivity and (ii) does the featurization of the input sequences matter. To test pretraining, we refer to `mata_descriptions` with the term \textbf{warm} to include pretraining, and \textbf{cold} to start from a randomly initialized seed. For featurization, we explored \textbf{label-encoded} sequences with a one-hot-encoder of amino acid identities, versus an \textbf{ESM}-featurized embedding to represent each sequence in the PPI.
+
+## Experimental details
+
+We studied the efficacy of generating binders with different model hyperparameters. This dataset includes 36862 unique VHHs and 8 unique RBD sequences.
+
+A more extensive methods section can be found in our publication [here](https://pmc.ncbi.nlm.nih.gov/articles/PMC12296056/).
+
+## Misc dataset details
+
+We define the following binders:
+
+### A-library (VHHs)
+
+There are several terms you can filter by:
+
+- `VHH_WT_<i>`: These are WT replicates.
+- `VHH_label_encoded_cold`: Label encoded sequences with no pretraining
+- `VHH_label_encoded_warm`: Label encoded sequences with pretraining
+- `VHH_esm_cold`: ESM featurized sequences with no pretraining
+- `VHH_esm_warm`: ESM featurized sequences with pretraining
+
+### Alpha-library
+
+- `SARS-CoV2_RBD_(6LZG)`
+- `WIV1`
+- `LYRa11`
+- `MERS_RBD`
+- `OMICRON_BA2_TStarr_Seq&Trunc`
+- `SARS-COV-2_GAMMA_RBD_LLNL_Trunc`
+- `SARS-CoV-2_Delta_RBD_L452R_T478K`
+- `SARS-CoV1_RBD_(6LZG_Mimic)`

data/YM_1068/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd459ff8a4e076c0cc71a779c2cc7229c585201e970ff04efdf103f153b4c0ac
+size 9583963