DeepAffinity: Interpretable Deep Learning of Compound-Protein Affinity through Unified Recurrent and Convolutional Neural Networks

Mostafa Karimi; Di Wu; Zhangyang Wang; Yang Shen

Source-linked AI summary

DeepAffinity: Interpretable Deep Learning of Compound-Protein Affinity through Unified Recurrent and Convolutional Neural Networks

Mostafa Karimi, Di Wu, Zhangyang Wang, Yang Shen

arXiv:1806.07537v2Jun 20, 2018q-bio.BMcs.LGstat.ML

TL;DR

The paper addresses the lack of accurate, applicable, and interpretable compound–protein affinity prediction from sequences alone. It combines structurally annotated representations with semi-supervised unified RNN-CNN models and attention, achieving reported IC50 relative errors within 5-fold on comprehensive tests and within 10-fold for unseen protein classes.

Problem
Existing methods lack high-applicability, accurate, and interpretable prediction of quantitative compound–protein affinity from sequences alone.
Method
The method combines SMILES and structurally annotated protein-sequence representations with unlabeled-data seq2seq pre-training, labeled-data unified RNN-CNN learning, and attention mechanisms.
Results
Relative error for IC50 is within 5-fold on a comprehensive test set and within 10-fold on generalization sets containing protein classes excluded from training.
Takeaways & Limitations
Attention analyses support predicting protein binding sites and explaining binding selectivity, while target selectivity was predicted successfully in the reported case studies.
Takeaways & Limitations
Protein-sequence representations required substantially more computation and produced test perplexity over 12, compared with 1.001 for SPS, making them inadequate for subsequent semi-supervised learning.

Abstract

from arXiv · show

Motivation: Drug discovery demands rapid quantification of compound-protein interaction (CPI). However, there is a lack of methods that can predict compound-protein affinity from sequences alone with high applicability, accuracy, and interpretability. Results: We present a seamless integration of domain knowledges and learning-based approaches. Under novel representations of structurally-annotated protein sequences, a semi-supervised deep learning model that unifies recurrent and convolutional neural networks has been proposed to exploit both unlabeled and labeled data, for jointly encoding molecular representations and predicting affinities. Our representations and models outperform conventional options in achieving relative error in IC$_{50}$ within 5-fold for test cases and 20-fold for protein classes not included for training. Performances for new protein classes with few labeled data are further improved by transfer learning. Furthermore, separate and joint attention mechanisms are developed and embedded to our model to add to its interpretability, as illustrated in case studies for predicting and explaining selective drug-target interactions. Lastly, alternative representations using protein sequences or compound graphs and a unified RNN/GCNN-CNN model using graph CNN (GCNN) are also explored to reveal algorithmic challenges ahead. Availability: Data and source codes are available at https://github.com/Shen-Lab/DeepAffinity Supplementary Information: Supplementary data are available at http://shen-lab.github.io/deep-affinity-bioinf18-supp-rev.pdf

1 Introduction

The paper targets quantitative compound–protein affinity prediction from sequence data alone, seeking broader applicability and interpretability than prior structure- or sequence-based approaches. It introduces compact representations and a semi-supervised unified RNN-CNN framework that uses unlabeled and labeled data.

Motivation and prior methods: Sequence-based methods are broadly applicable and fast, but commonly classify binding outcomes and offer limited interpretability.Structure-based methods provide quantitative, interpretable affinity estimates but depend on available 3D structures and costly docking.
Research goal: The paper aims to predict quantitative affinity measures such as IC50, Ki, and Kd from sequences alone.The goal is to balance applicability, throughput, and interpretability while addressing a more difficult regression problem than prior sequence-based classification.
Proposed approach: The proposed representations encode compounds as SMILES and proteins with structural and physicochemical alphabets.The authors describe these representations as lower-dimensional and more informative than substructure fingerprints or Pfam domains.
Proposed approach: A semi-supervised framework uses unlabeled data to pre-train bidirectional RNN representations before end-to-end supervised affinity prediction with a unified RNN-CNN model.The pipeline combines unsupervised representation learning with labeled-data training.

2 Materials and Methods

The method combines structurally annotated sequence representations, unsupervised RNN pre-training, supervised CNN prediction, and attention mechanisms for interpretable compound–protein affinity modeling. It assumes only one-dimensional sequence data and uses public labeled and unlabeled datasets.

Data: The study uses BindingDB affinity labels, STITCH compound SMILES, and UniRef protein sequences.These sources provide the labeled and unlabeled data used in the framework.
Data: Only 1D sequence data are assumed available; 3D structures of proteins, compounds, and complexes are not used.This defines the method’s structural-data scope.
Representations: SMILES strings encode compound structures compactly, while SPS encodes protein secondary structure, solvent accessibility, physicochemical properties, and segment length.SPS uses four small alphabets and is described as providing higher-resolution structural detail than Pfam features.
Unsupervised pre-training: Seq2seq GRU auto-encoders map SMILES or SPS sequences to fixed-dimensional thought vectors learned from unlabeled data.The encoder produces the representation and the decoder reconstructs the target sequence; pre-training captures long-range dependencies.
Affinity prediction: Supervised prediction appends 1D convolution and max-pooling layers to the compound and protein RNN encoders, then combines their outputs.The unified pipeline is trained end to end with pre-trained RNN initialization.
Attention mechanisms: Separate, marginalized, and joint attention mechanisms quantify focus on protein SSEs, compound atoms, or their pairwise interactions.Joint attention models pairwise interactions and offers the best interpretability among the three mechanisms, with the most parameters.

3.1 Compound and protein representations

The authors compare seq2seq representation-learning variants and select bidirectional GRUs with attention for subsequent compound and protein representation learning. Perplexity is used to assess auto-encoding quality.

Representation-learning variants: The study compares vanilla seq2seq with bucketing, bidirectional GRU, attention, and combined bidirectional-GRU-plus-attention variants.These variants are evaluated using supplementary tables.
Evaluation: Perplexity, related to the entropy of the modeled probability distribution, is the evaluation metric for auto-encoding performance.The paper gives Perp(P) = 2^H(P) ≥ 1.
Model selection: The bidirectional-GRU model with attention is selected as the most appropriate variant for learning compound and protein representations.The selected variant is used thereafter.

3.2 Compound-protein affinity prediction

The study compares learned representations and model architectures for compound-protein affinity regression, then evaluates attention mechanisms and transfer learning for new protein classes. Unified RNN-CNN models and deep transfer learning improve prediction relative to the tested alternatives.

Comparing novel representations to baseline ones: Novel SMILES/SPS representations outperform k-hot molecular/Pfam baselines in affinity regression while using less training time and memory.For the best random forest models, training time decreases by 46% and memory use by 24%; default-test performance is unchanged, while RMSE improves on two generalization sets.
Comparing shallow and deep models: Unified RNN-CNN models outperform both random forest and separate RNN-CNN models.The comparison uses the novel representations and evaluates single models and ensembles with separate attention.
Comparing shallow and deep models: Unified RNN-CNN models predict IC50 with relative errors below 10^0.7 = 5-fold on the test set and around 10^1.3 = 20-fold for unseen protein classes.The models use relatively small labeled datasets, while representations learned from abundant unlabeled data are tuned for the task.
Comparing attention mechanisms: Separate attention edges joint attention in generalization, especially for receptor tyrosine kinases, while all attention mechanisms show similar training and test performance.The comparison uses the same dataset and unified RNN-CNN models for pIC50 prediction.
Deep transfer learning for new classes of protein targets: Deep transfer learning increasingly improves prediction for new protein classes as labeled-data coverage increases, with significant improvement at 1% coverage.Transferred deep models outperform random forest models retrained specifically for each new protein class.

3.3 Predicting target selectivity of drugs

The study evaluates whether unified RNN-CNN models can predict drug-target selectivity across thrombin/Xa, COX, and PTP examples. The models generally identified selective interactions, with joint attention improving some selectivity-margin estimates but not all cases.

Overall selectivity prediction: Unified RNN-CNN models successfully predicted target selectivity for 6 of 7 drugs, whereas baseline representations and random forest failed for most drugs.The evaluation used three sets of drug-target interactions with increasing prediction difficulty.
Factor Xa versus thrombin: Unified RNN-CNN models correctly predicted DX-9065a’s preference for factor Xa over thrombin.Separate and joint attention models predicted pKi differences of 2.4 and 3.9, respectively, while the known selectivity margin is at least 4.7.
Cyclooxygenase (COX) protein family: Unified RNN-CNN models predicted all three NSAID selectivities between COX-1 and COX-2.Random forest with novel representations correctly predicted two of three, while baseline representations made COX-1 and COX-2 equally favorable for each drug.
Protein-tyrosine phosphatase (PTP) family: Within the PTP family, unified RNN-CNN models correctly predicted selectivity for all but one of three compounds.The exception was compound 1, OBA.
Protein-tyrosine phosphatase (PTP) family: Joint attention improved predicted selectivity margins for PTP compounds 2 and 3 to 0.58 and 0.82 pKi units.Separate attention predicted smaller differences of 0.09 and 0.03, described as likely insignificant.

3.4 Explaining target selectivity of drugs

The unified RNN-CNN model uses attention scores to explain compound-protein interactions, including binding-site prediction and target selectivity. Case studies connect high-attention regions with known structural contacts and selectivity determinants.

Binding-site interpretation: Attention scores from interpretable representations identify protein SSEs and compound atoms that may contribute to compound-protein interactions.The analysis uses interpretable SPS and SMILES representations with attention between RNN encoders and convolution layers.
Binding-site interpretation: Joint attention enriched binding-site prediction 1.7–5.8 fold over random ranking across three compound-protein interactions.Joint attention also performed better than separate attention, and binding sites received significantly higher attention scores than nonbinding sites.
Binding-site interpretation: For factor Xa–DX-9065a, warmer attention colors concentrated near the ligand, while the known binding-site loops ranked second.The top-ranked false-positive SSE was an immediate sequence neighbor near the ligand; relaxing the binding-site definition made all top 10% SSEs correct.
Binding-site interpretation: The top-ranked compound atom was nitrogen forming a hydrogen bond with factor Xa Asp189, although more cases are needed before drawing a conclusion.Many compound atoms received high attention, consistent with compounds fitting largely within protein pockets or grooves.
Target selectivity: The selectivity analysis compares normalized attention ranks across aligned factor Xa and thrombin sequence segments to locate the origin of differential binding.The method was designed for selective compounds such as factor Xa-selective DX-9065a.
Target selectivity: For DX-9065a, aligned segment 42 contains position 192, where Xa has Gln192 and thrombin has Glu192, the known selectivity determinant.The carboxylate forms a favorable hydrogen bond with Xa Gln192 but unfavorable electrostatic repulsion with thrombin Glu192.

4 Discussion

The discussion evaluates alternative protein and compound representations. It finds that residue-level protein sequences and graph-based compounds introduce computational or architectural challenges relative to the paper’s SPS and SMILES choices.

Protein representations: SPS representations combine protein sequence and structure information in a more compact form than amino-acid sequences.Residue-level sequences could offer higher-resolution precision and interpretability, motivating their evaluation.
Protein representations: Protein sequences were 10-times longer, required 10-times more GRUs, and produced test perplexity over 12 versus 1.001 for SPS.Training used twice the SPS time limit, yet the sequence model was judged inadequate for subsequent semi-supervised learning.
Compound representations: SMILES were selected partly because of advances in sequence modeling, but syntactically invalid SMILES can still represent valid chemical structures.The authors therefore also explored chemical formulae represented as 2D graphs.
Compound representations: The unified RNN/GCNN-CNN model replaces compound-sequence RNN layers with a graph CNN while retaining the rest of the architecture.This graph-based variant uses separate attention and is evaluated for compound-protein interactions.
Compound representations: Graph-based compounds did not outperform SMILES on RMSE and performed much worse on Pearson’s correlation coefficient for pIC50 prediction.The authors interpret this as evidence that graph models need new architectures, not as proof that SMILES are intrinsically superior.

5 Conclusion

The paper concludes that semi-supervised unified RNN-CNN models can predict compound-protein affinity from compound identities and protein sequences while supporting interpretable attention analyses. It also reports transfer-learning gains for protein classes with limited labeled data and identifies representation and interpretability challenges.

Contributions: Unified RNN-CNN models use unlabeled and labeled data to learn context- and task-specific molecular representations and predict compound-protein affinity.The models use compound identities and protein sequences without requiring compound-protein structures.
Predictive performance: Relative IC50 error was within 5-fold on a comprehensive test set and within 10-fold on generalization sets of unseen protein classes.The conclusion also reports that deeper models could further improve results.
Representations and interpretation: Learned SPS and SMILES representations improved predictive power and training efficiency over molecular descriptors and Pfam domains.Attention mechanisms quantify which protein or compound parts receive focus for each prediction.
Representations and interpretation: The models predicted target selectivity in all reported drug case studies, while conventional representations and machine-learning models failed in some cases.Attention analyses also showed promise for predicting binding sites and selectivity origins, while motivating further development.
Remaining challenges: SSE-level protein representations support interpretability, whereas residue-level amino-acid inputs increase computational burden and worsen RNN convergence.The paper also explored SMILES and graph representations, with the graph-based unified model not outperforming the SMILES model.