Model type
Unified RNN-CNN: GRU seq2seq encoders with a convolutional regression head
Model
DeepAffinity
DeepAffinity predicts compound-protein affinity from a compound SMILES string and a structure property-annotated protein sequence (SPS), using pretrained recurrent seq2seq encoders followed by convolutional layers.
Sources
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 abstract, p.1; §2.2.2, p.3; §2.4, p.4Overview
Inputs
Compound SMILES string and protein SPS string
Sources
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.2, PDF pp.2-3Output
Affinity on a logarithmic scale (pIC50, pKi or pKd)
Sources
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.1, p.2limited source coverage · Automated source review, 2026-09-24. All specifications and missing details
Evaluations and results
3 evaluations · 3 metric rows. Different protocols are not a single leaderboard.
Filter evaluations
Applied filters: All linked evaluations
| Tested configuration | Protocol and dataset | Finding | Evidence and details |
|---|---|---|---|
| Configuration: DeepAffinity (unified RNN/RNN-CNN; DSSP-derived SPS) (cited as [Karimi et al., 2019]) | Task: ATOM3D LBA-RMSE: Ligand binding affinity, root mean squared error Dataset subset: ATOM3D LBA (ATOM3D split) | 1.89 rmse error · lower Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceTrained and scored under the ATOM3D split for this task. Asterisks in the paper mark a run whose training data differed. Aggregation: Not reported ATOM3D: Tasks On Molecules in Three Dimensions · Table 5, row(LBA RMSE), column([Karimi et al., 2019]) |
| Configuration: DeepAffinity (unified RNN/RNN-CNN; DSSP-derived SPS) (cited as [Karimi et al., 2019]) | Task: ATOM3D LBA-RP: Ligand binding affinity, global Pearson correlation Dataset subset: ATOM3D LBA (ATOM3D split) | 0.415 pearson_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceTrained and scored under the ATOM3D split for this task. Asterisks in the paper mark a run whose training data differed. Aggregation: Not reported ATOM3D: Tasks On Molecules in Three Dimensions · Table 5, row(glob. RP), column([Karimi et al., 2019]) |
| Configuration: DeepAffinity (unified RNN/RNN-CNN; DSSP-derived SPS) (cited as [Karimi et al., 2019]) | Task: ATOM3D LBA-RS: Ligand binding affinity, global Spearman correlation Dataset subset: ATOM3D LBA (ATOM3D split) | 0.426 spearman_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceTrained and scored under the ATOM3D split for this task. Asterisks in the paper mark a run whose training data differed. Aggregation: Not reported ATOM3D: Tasks On Molecules in Three Dimensions · Table 5, row(glob. RS), column([Karimi et al., 2019]) |
Source checking is not independent reproduction. Release 2026-09-24-eb3fb1cb4c7f.
Use this model
How it works, versions and access
Versions and evaluated configurations
How it works
How it works
GRU seq2seq auto-encoders are first trained without labels on compound SMILES and protein SPS strings. In the unified RNN-CNN model, a 1D convolution and max-pooling layer is added after each encoder, the two outputs are concatenated and passed through two fully connected layers, and the whole pipeline is trained end to end with the pretrained encoders as initialisation. A separate RNN-CNN baseline keeps the encoders fixed.
Sources
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §§2.3-2.4, p.4Protein representation
In the original paper, SPS strings group residues into secondary structure elements using secondary structure and solvent accessibility predicted from sequence by SSpro/ACCpro.
Sources
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.2.2, p.3Variants
The paper also describes separate, marginalized and joint attention mechanisms trained jointly with the encoders and CNN, and a unified RNN/GCNN-CNN variant that replaces the compound RNN with a graph CNN.
Sources (2)
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2); DeepAffinity repository README (Shen-Lab/DeepAffinity) · arXiv:1806.07537v2 §2.5, p.4 and PDF p.8; README Table of contentsStrengths, limitations and unresolved questions
Strengths and limitations
Strengths and considerations
No source-reviewed explanatory claims are recorded here yet.
Limitations and conditions
- In the original paper the protein structure annotations are predicted from sequence rather than taken from solved structures.
Sources
DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.2.2, p.3
Profile review details
AI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction.
Stable record: identity-model-deepaffinitySpecifications
Inputs, training, access and other details
Explanatory profile: limited source coverage · Automated source review, 2026-09-24. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.
| Property | Description and evidence |
|---|---|
| Model type | Unified RNN-CNN: GRU seq2seq encoders with a convolutional regression headSourcesDeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §§2.3-2.4, p.4 |
| Inputs | Compound SMILES string and protein SPS stringSourcesDeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.2, PDF pp.2-3 |
| Output | Affinity on a logarithmic scale (pIC50, pKi or pKd)SourcesDeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.1, p.2 |
| Original training data | BindingDB IC50, Ki and Kd labels, with four protein classes held out of IC50 trainingSourcesDeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 §2.1, p.2 |
| Code | github.com/Shen-Lab/DeepAffinitySourcesDeepAffinity repository README (Shen-Lab/DeepAffinity) · README at debca4c9f01991a37594e1b67f3dbb46a29e82d1 |
| Parameter count | Not extracted · Needs further source reviewSourcesDeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · arXiv:1806.07537v2 |
| Known versions | Not extracted or verified for this record. |
| Context limits | Not extracted or verified for this record. |
| Access | Not extracted or verified for this record. |
| Code licence | Not extracted or verified for this record. |
| Weights licence | Not extracted or verified for this record. |
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
12 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Model type Unified RNN-CNN: GRU seq2seq encoders with a convolutional regression head Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §§2.3-2.4, p.4 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Inputs Compound SMILES string and protein SPS string Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §2.2, PDF pp.2-3 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Output Affinity on a logarithmic scale (pIC50, pKi or pKd) Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §2.1, p.2 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Original training data BindingDB IC50, Ki and Kd labels, with four protein classes held out of IC50 training Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §2.1, p.2 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Code github.com/Shen-Lab/DeepAffinity Individual claims | DeepAffinity repository README (Shen-Lab/DeepAffinity) README at debca4c9f01991a37594e1b67f3dbb46a29e82d1 Version: debca4c9f01991a37594e1b67f3dbb46a29e82d1 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Parameter count Not extracted Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 Version: 1806.07537v2 | unextracted automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Limitation In the original paper the protein structure annotations are predicted from sequence rather than taken from solved structures. Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §2.2.2, p.3 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| How it works GRU seq2seq auto-encoders are first trained without labels on compound SMILES and protein SPS strings. In the unified RNN-CNN model, a 1D convolution and max-pooling layer is added after each encoder, the two outputs are concatenated and passed through two fully connected layers, and the whole pipeline is trained end to end with the pretrained encoders as initialisation. A separate RNN-CNN baseline keeps the encoders fixed. Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §§2.3-2.4, p.4 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Protein representation In the original paper, SPS strings group residues into secondary structure elements using secondary structure and solvent accessibility predicted from sequence by SSpro/ACCpro. Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §2.2.2, p.3 Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
| Variants The paper also describes separate, marginalized and joint attention mechanisms trained jointly with the encoders and CNN, and a unified RNN/GCNN-CNN variant that replaces the compound RNN with a graph CNN. Individual claims | DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) arXiv:1806.07537v2 §2.5, p.4 and PDF p.8; README Table of contents Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 1806.07537v2 | source checked automated source review · 2026-09-24 Audit detailsAI-assisted review of the cited claims against pinned primary sources. No human scientific review and no independent reproduction. Field: Source artifact SHA-256: Hash scope: SHA-256 of the retrieved original artifact bytes |
Sources and history
View linked audit checks and correction history
Release 2026-09-24-eb3fb1cb4c7f · Record review: source checked
2 source records and release history
- DeepAffinity: interpretable deep learning of compound–protein affinity (arXiv:1806.07537v2) · Original source · 1806.07537v2
- DeepAffinity repository README (Shen-Lab/DeepAffinity) · Original source · debca4c9f01991a37594e1b67f3dbb46a29e82d1
Technical metadata and extraction receipts
Stable ID: identity-model-deepaffinity
- areas
- molecular-interactions
- entity level
- method
- reported name
- DeepAffinity
- version
- Not reported
- entity classification
- review date: 2026-09-24; rationale: The cited paper names DeepAffinity as a learned compound-protein affinity predictor. This record covers the method family described there; no checkpoint or variant is implied.; source ids: source-label-deepaffinity-arxiv-97d6e3ae; source locator: arXiv:1806.07537v2 title and abstract, p.1; §§2.4-2.5, p.4; ambiguities: The paper reports separate and unified RNN-CNN models, three attention mechanisms and a graph (GCNN) variant. Configurations must state which variant they used.