Despite the disparate types of data and scientific goals in the learning tasks covered above, several challenges are broadly important for deep learning in the biomedical domain. Here we examine these factors that may impede further progress, ask what steps have already been taken to overcome them, and suggest future research directions.
Some of the challenges in applying deep learning are shared with other machine learning methods. In particular, many problem-specific optimizations described in this review reflect a recurring universal tradeoff---controlling the flexibility of a model in order to maximize predictivity. Methods for adjusting the flexibility of deep learning models include dropout, reduced data projections, and transfer learning (described below). One way of understanding such model optimizations is that they incorporate external information to limit model flexibility and thereby improve predictions. This balance is formally described as a tradeoff between "bias and variance" [@url:http://www.deeplearningbook.org/].
Although the bias-variance tradeoff is common to all machine learning applications, recent empirical and theoretical observations suggest that deep learning models may have uniquely advantageous generalization properties [@tag:Zhang2017_generalization; @tag:Lin2017_why_dl_works]. Nevertheless, additional advances will be needed to establish a coherent theoretical foundation that enables practitioners to better reason about their models from first principles.
Making predictions in the presence of high class imbalance and differences between training and generalization data is a common feature of many large biomedical datasets, including deep learning models of genomic features, patient classification, disease detection, and virtual screening. Prediction of transcription factor binding sites exemplifies the difficulties with learning from highly imbalanced data. The human genome has 3 billion base pairs, and only a small fraction of them are implicated in specific biochemical activities. Less than 1% of the genome can be confidently labeled as bound for most transcription factors.
Estimating the false discovery rate (FDR) is a standard method of evaluation in genomics that can also be applied to deep learning model predictions of genomic features. Using deep learning predictions for targeted validation experiments of specific biochemical activities necessitates a more stringent FDR (typically 5--25%). However, when predicted biochemical activities are used as features in other models, such as gene expression models, a low FDR may not be necessary.
What is the correspondence between FDR metrics and commonly used classification metrics such as AUPR and AUROC? AUPR evaluates the average precision, or equivalently, the average FDR across all recall thresholds. This metric provides an overall estimate of performance across all possible use cases, which can be misleading for targeted validation experiments. For example, classification of TF binding sites can exhibit a recall of 0% at 10% FDR and AUPR greater than 0.6. In this case, the AUPR may be competitive, but the predictions are ill-suited for targeted validation that can only examine a few of the highest-confidence predictions. Likewise, AUROC evaluates the average recall across all false positive rate (FPR) thresholds, which is often a highly misleading metric in class-imbalanced domains [@doi:10.1145/1143844.1143874; @doi:10.1038/nmeth.3945]. Consider a classification model with recall of 0% at FDR less than 25% and 100% recall at FDR greater than 25%. In the context of TF binding predictions where only 1% of genomic regions are bound by the TF, this is equivalent to a recall of 100% for FPR greater than 0.33%. In other words, the AUROC would be 0.9967, but the classifier would be useless for targeted validation. It is not unusual to obtain a chromosome-wide AUROC greater than 0.99 for TF binding predictions but a recall of 0% at 10% FDR. Consequently, practitioners must select the metric most tailored to their subsequent use case to use these methods most effectively.
Genome-wide continuous signals are commonly formulated into classification labels through signal peak detection. ChIP-seq peaks are used to identify locations of TF binding and histone modifications. Such procedures rely on thresholding criteria to define what constitutes a peak in the signal. This inevitably results in a set of signal peaks that are close to the threshold, not sufficient to constitute a positive label but too similar to positively labeled examples to constitute a negative label. To avoid an arbitrary label for these examples they may be labeled as "ambiguous". Ambiguously labeled examples can then be ignored during model training and evaluation of recall and FDR. The correlation between model predictions on these examples and their signal values can be used to evaluate if the model correctly ranks these examples between positive and negative examples.
In assessing the upper bound on the predictive performance of a deep learning model, it is necessary to incorporate inherent between-study variation inherent to biomedical research [@tag:Errington2014_reproducibility]. Study-level variability limits classification performance and can lead to underestimating prediction error if the generalization error is estimated by splitting a single dataset. Analyses can incorporate data from multiple labs and experiments to capture between-study variation within the prediction model mitigating some of these issues.
Deep learning based solutions for biomedical applications could substantially benefit from guarantees on the reliability of predictions and a quantification of uncertainty. Due to biological variability and precision limits of equipment, biomedical data do not consist of precise measurements but of estimates with noise. Hence, it is crucial to obtain uncertainty measures that capture how noise in input values propagate through deep neural networks. Such measures can be used for reliability assessment of automated decisions in clinical and public health applications, and for guarding against model vulnerabilities in the face of rare or adversarial cases [@tag:ghahramani_protect]. Moreover, in fundamental biological research, measures of uncertainty help researchers distinguish between true regularities in the data and patterns that are false or merely anecdotal. There are two main uncertainties that one can calculate: epistemic and aleatoric [@tag:uncertainty_types]. Epistemic uncertainty describes uncertainty about the model, its structure, or its parameters. This uncertainty is caused by insufficient training data or by a difference in the training set and testing set distributions, so it vanishes in the limit of infinite data. On the other hand, aleatoric uncertainty describes uncertainty inherent in the observations. This uncertainty is due to noisy or missing data, so it vanishes with the ability to observe all independent variables with infinite precision. A good way to represent aleatoric uncertainty is to design an appropriate loss function with an uncertainty variable. In the case of data-dependent aleatoric uncertainty, one can train the model to increase its uncertainty when it is incorrect due to noisy or missing data, and in the case of task-dependent aleatoric uncertainty, one can optimize for the best uncertainty parameter for each task [@tag:uncertainty_multi_task]. Meanwhile, there are various methods for modeling epistemic uncertainty, outlined below.
In classification tasks, confidence calibration is the problem of using classifier scores to predict class membership probabilities that match the true membership likelihoods. These membership probabilities can be used to assess the uncertainty associated with assigning the example to each of the classes. Guo et al. [@tag:guo_calibration] observed that contemporary neural networks are poorly calibrated and provided a simple recommendation for calibration: temperature scaling, a single parameter special case of Platt scaling [@tag:platt_scaling]. In addition to confidence calibration, there is early work from Chryssolouris et al. [@tag:Chryssolouris1996_confidence] that described a method for obtaining confidence intervals with the assumption of normally distributed error for the neural network. More recently, Hendrycks and Gimpel discovered that incorrect or out-of-distribution examples usually have lower maximum softmax probabilities than correctly classified examples, allowing for effective detection of misclassified examples [@tag:out_dist_baseline]. Liang et al. used temperature scaling and small perturbations to further separate the softmax scores of correctly classified examples and the scores of out-of-distribution examples, allowing for more effective detection [@tag:temp_out_dist]. This approach outperformed the baseline approaches by a large margin, establishing a new state-of-the-art performance.
An alternative approach for obtaining principled uncertainty estimates from deep learning models is to use Bayesian neural networks. Deep learning models are usually trained to obtain the most likely parameters given the data. However, choosing the single most likely set of parameters ignores the uncertainty about which set of parameters (among the possible models that explain the given dataset) should be used. This sometimes leads to uncertainty in predictions when the chosen likely parameters produce high-confidence but incorrect results. On the other hand, the parameters of Bayesian neural networks are modeled as full probability distributions. This Bayesian approach comes with a whole host of benefits, including better calibrated confidence estimates [@tag:ai_safety] and more robustness to adversarial and out-of-distribution examples [@tag:strong_adversary]. Unfortunately, modeling the full posterior distribution for the model’s parameters given the data is usually computationally intractable. One popular method for circumventing this high computational cost is called test-time dropout [@tag:Gal2015_dropout], where an approximate posterior distribution is obtained using variational inference. Gal and Ghahramani showed that a stack of fully connected layers with dropout between the layers is equivalent to approximate inference in a Gaussian process model [@tag:Gal2015_dropout]. The authors interpret dropout as a variational inference method and apply their method to convolutional neural networks. This is simple to implement and preserves the possibility of obtaining cheap samples from the approximate posterior distribution. Operationally, obtaining model uncertainty for a given case becomes as straightforward as leaving dropout turned on and predicting multiple times. The spread of the different predictions is a reasonable proxy for model uncertainty. This technique has been successfully applied in an automated system for detecting diabetic retinopathy [@tag:retinopathy_uncertainty], where uncertainty-informed referrals improved diagnostic performance and allowed the model to meet the National Health Service recommended levels of sensitivity and specificity. The authors also found that entropy performs comparably to the spread obtained via test-time dropout for identifying uncertain cases, and therefore it can be used instead for automated referrals.
Several other techniques have been proposed for effectively estimating predictive uncertainty as uncertainty quantification for neural networks continues to be an active research area. Recently, McClure and Kriegeskorte observed that test-time sampling improved calibration of the probabilistic predictions, sampling weights led to more robust uncertainty estimates than sampling units, and spike-and-slab sampling is superior to Gaussian dropconnect and Bernoulli dropout [@tag:mcclure_bayesian]. Krueger et al. introduced Bayesian hypernetworks [@tag:bayesian_hypernets] as another framework for approximate Bayesian inference in deep learning, where an invertible generative hypernetwork maps isotropic Gaussian noise to parameters of the primary network allowing for computationally cheap sampling and efficient estimation of the posterior. Meanwhile, Lakshminarayanan et al. proposed using deep ensembles, which are traditionally used for boosting predictive performance, on standard (non-Bayesian) neural networks to obtain well-calibrated uncertainty estimates that are comparable to those obtained by Bayesian neural networks [@tag:uncertainty_ensembles]. In cases where model uncertainty is known to be caused by a difference in training and testing distributions, domain adaptation based techniques can help mitigate the problem [@tag:domain_adapt_uncertainty].
Despite the success and popularity of deep learning, some deep learning models can be surprisingly brittle. Researchers are actively working on modifications to deep learning frameworks to enable them to handle probability and embrace uncertainty. Most notably, Bayesian modeling and deep learning are being integrated with renewed enthusiasm. As a result, several opportunities for innovation arise: understanding the causes of model uncertainty can lead to novel optimization and regularization techniques, assessing the utility of uncertainty estimation techniques on various model architectures and structures can be very useful to practitioners, and extending Bayesian deep learning to unsupervised settings can be a significant breakthrough [@tag:gal_thesis]. Unfortunately, uncertainty quantification techniques are underutilized in the computational biology communities and largely ignored in the current deep learning for biomedicine literature. Thus, the practical value of uncertainty quantification in biomedical domains is yet to be appreciated.
As deep learning models achieve state-of-the-art performance in a variety of domains, there is a growing need to make the models more interpretable. Interpretability matters for two main reasons. First, a model that achieves breakthrough performance may have identified patterns in the data that practitioners in the field would like to understand. However, this would not be possible if the model is a black box. Second, interpretability is important for trust. If a model is making medical diagnoses, it is important to ensure the model is making decisions for reliable reasons and is not focusing on an artifact of the data. A motivating example of this can be found in Ba and Caruana [@tag:Caruana2014_need], where a model trained to predict the likelihood of death from pneumonia assigned lower risk to patients with asthma, but only because such patients were treated as higher priority by the hospital. In the context of deep learning, understanding the basis of a model's output is particularly important as deep learning models are unusually susceptible to adversarial examples [@tag:Nguyen2014_adversarial] and can output confidence scores over 99.99% for samples that resemble pure noise.
As the concept of interpretability is quite broad, many methods described as improving the interpretability of deep learning models take disparate and often complementary approaches.
Several approaches ascribe importance on an example-specific basis to the parts of the input that are responsible for a particular output. These can be broadly divided into perturbation-based approaches and backpropagation-based approaches.
Perturbation-based approaches change parts of the input and observe the impact on the output of the network. Alipanahi et al. [@tag:Alipanahi2015_predicting] and Zhou & Troyanskaya [@tag:Zhou2015_deep_sea] scored genomic sequences by introducing virtual mutations at individual positions in the sequence and quantifying the change in the output. Umarov et al. [@doi:10.1371/journal.pone.0171410] used a similar strategy, but with sliding windows where the sequence within each sliding window was substituted with a random sequence. Kelley et al. [@tag:Kelley2016_basset] inserted known protein-binding motifs into the centers of sequences and assessed the change in predicted accessibility. Ribeiro et al. [@tag:Ribeiro2016_lime] introduced LIME, which constructs a linear model to locally approximate the output of the network on perturbed versions of the input and assigns importance scores accordingly. For analyzing images, Zeiler and Fergus [@tag:Zeiler2013_visualizing] applied constant-value masks to different input patches. More recently, marginalizing over the plausible values of an input has been suggested as a way to more accurately estimate contributions [@tag:Zintgraf2017_visualizing].
A common drawback to perturbation-based approaches is computational efficiency: each perturbed version of an input requires a separate forward propagation through the network to compute the output. As noted by Shrikumar et al. [@tag:Shrikumar2017_learning], such methods may also underestimate the impact of features that have saturated their contribution to the output, as can happen when multiple redundant features are present. To reduce the computational overhead of perturbation-based approaches, Fong and Vedaldi [@tag:Fong2017_perturb] solve an optimization problem using gradient descent to discover a minimal subset of inputs to perturb in order to decrease the predicted probability of a selected class. Their method converges in many fewer iterations but requires the perturbation to have a differentiable form.
Backpropagation-based methods, in which the signal from a target output neuron is propagated backwards to the input layer, are another way to interpret deep networks that sidestep inefficiencies of the perturbation-based methods. A classic example of this is calculating the gradients of the output with respect to the input [@tag:Simonyan2013_deep] to compute a "saliency map". Bach et al. [@tag:Bach2015_on] proposed a strategy called Layerwise Relevance Propagation, which was shown to be equivalent to the element-wise product of the gradient and input [@tag:Shrikumar2017_learning; @tag:Kindermans2016_investigating]. Networks with Rectified Linear Units (ReLUs) create nonlinearities that must be addressed. Several variants exist for handling this [@tag:Zeiler2013_visualizing; @tag:Springenberg2014_striving]. Backpropagation-based methods are a highly active area of research. Researchers are still actively identifying weaknesses [@tag:Mahendran2016_salient], and new methods are being developed to address them [@tag:Selvaraju2016_grad;; @tag:Sundararajan2017_axiomatic; @tag:Shrikumar2017_learning]. Lundberg and Lee [@tag:Lundberg2016_an] noted that several importance scoring methods including integrated gradients and LIME could all be considered approximations to Shapely values [@tag:Shapely], which have a long history in game theory for assigning contributions to players in cooperative games.
Matching or exaggerating the hidden representation
Another approach to understanding the network's predictions is to find artificial inputs that produce similar hidden representations to a chosen example. This can elucidate the features that the network uses for prediction and drop the features that the network is insensitive to. In the context of natural images, Mahendran and Vedaldi [@tag:Mahendran2014_understanding] introduced the "inversion" visualization, which uses gradient descent and backpropagation to reconstruct the input from its hidden representation. The method required placing a prior on the input to favor results that resemble natural images. For genomic sequence, Finnegan and Song [@tag:Finnegan2017_maximum] used a Markov chain Monte Carlo algorithm to find the maximum-entropy distribution of inputs that produced a similar hidden representation to the chosen input.
A related idea is "caricaturization", where an initial image is altered to exaggerate patterns that the network searches for [@tag:Mahendran2016_visualizing]. This is done by maximizing the response of neurons that are active in the network, subject to some regularizing constraints. Mordvintsev et al. [@tag:Mordvintsev2015_inceptionism] leveraged caricaturization to generate aesthetically pleasing images using neural networks.
Activation maximization can reveal patterns detected by an individual neuron in the network by generating images which maximally activate that neuron, subject to some regularizing constraints. This technique was first introduced in Ehran et al. [@tag:Ehran2009_visualizing] and applied in subsequent work [@tag:Simonyan2013_deep; @tag:Mordvintsev2015_inceptionism; @tag:Yosinksi2015_understanding; @tag:Mahendran2016_visualizing]. Lanchantin et al. [@tag:Lanchantin2016_motif] applied class-based activation maximization to genomic sequence data. One drawback of this approach is that neural networks often learn highly distributed representations where several neurons cooperatively describe a pattern of interest. Thus, visualizing patterns learned by individual neurons may not always be informative.
Several interpretation methods are specifically tailored to recurrent neural network architectures. The most common form of interpretability provided by RNNs is through attention mechanisms, which have been used in diverse problems such as image captioning and machine translation to select portions of the input to focus on generating a particular output [@tag:Bahdanu2014_neural; @tag:Xu2015_show]. Deming et al. [@tag:Deming2016_genetic] applied the attention mechanism to models trained on genomic sequence. Attention mechanisms provide insight into the model's decision-making process by revealing which portions of the input are used by different outputs. Singh et al. used a hierarchy of attention layers to locate important genome positions and signals for predicting gene expression from histone modifications [@tag:Singh2017_attentivechrome]. In the clinical domain, Choi et al. [@tag:Choi2016_retain] leveraged attention mechanisms to highlight which aspects of a patient's medical history were most relevant for making diagnoses. Choi et al. [@tag:Choi2016_gram] later extended this work to take into account the structure of disease ontologies and found that the concepts represented by the model aligned with medical knowledge. Note that interpretation strategies that rely on an attention mechanism do not provide insight into the logic used by the attention layer.
Visualizing the activation patterns of the hidden state of a recurrent neural network can also be instructive. Early work by Ghosh and Karamcheti [@tag:Ghosh1992_sequence] used cluster analysis to study hidden states of comparatively small networks trained to recognize strings from a finite state machine. More recently, Karpathy et al. [@tag:Karpathy2015_visualizing] showed the existence of individual cells in LSTMs that kept track of quotes and brackets in character-level language models. To facilitate such analyses, LSTMVis [@tag:Strobelt2016_visual] allows interactive exploration of the hidden state of LSTMs on different inputs.
Another strategy, adopted by Lanchatin et al. [@tag:Lanchantin2016_motif] looks at how the output of a recurrent neural network changes as longer and longer subsequences are supplied as input to the network, where the subsequences begin with just the first position and end with the entire sequence. In a binary classification task, this can identify those positions which are responsible for flipping the output of the network from negative to positive. If the RNN is bidirectional, the same process can be repeated on the reverse sequence. As noted by the authors, this approach was less effective at identifying motifs compared to the gradient-based backpropagation approach of Simonyan et al. [@tag:Simonyan2013_deep], illustrating the need for more sophisticated strategies to assign importance scores in recurrent neural networks.
Murdoch and Szlam [@tag:Murdoch2017_automatic] showed that the output of an LSTM can be decomposed into a product of factors, where each factor can be interpreted as the contribution at a particular timestep. The contribution scores were then used to identify key phrases from a model trained for sentiment analysis and obtained superior results compared to scores derived via a gradient-based approach.
Interpretation of embedded or latent space features learned through generative unsupervised models can reveal underlying patterns otherwise masked in the original input. Embedded feature interpretation has been emphasized mostly in image and text based applications [@tag:Radford_dcgan; @tag:Word2Vec], but applications to genomic and biomedical domains are increasing.
For example, Way and Greene trained a variational autoencoder (VAE) on gene expression from The Cancer Genome Atlas (TCGA) [@doi:10.1038/ng.2764] and use latent space arithmetic to rapidly isolate and interpret gene expression features descriptive of high grade serous ovarian cancer subtypes [@tag:WayGreene2017_tybalt]. The most differentiating VAE features were representative of biological processes that are known to distinguish the subtypes. Latent space arithmetic with features derived using other compression algorithms were not as informative in this context [@tag:WayGreene2017_eval]. Embedding discrete chemical structures with autoencoders and interpreting the learned continuous representations with latent space arithmetic has also facilitated predicting drug-like compounds [@tag:Gomezb2016_automatic]. Furthermore, embedding biomedical text into lower dimensional latent spaces have improved name entity recognition in a variety of tasks including annotating clinical abbreviations, genes, cell lines, and drug names [@doi:10.3390/info6040848; @doi:10.1155/2014/240403; @doi:10.18653/v1/w15-3822; @doi:10.18653/v1/w15-3810].
Other approaches have used interpolation through latent space embeddings learned by GANs to interpret unobserved intermediate states. For example, Osokin et al. trained GANs on two-channel fluorescent microscopy images to interpret intermediate states of protein localization in yeast cells [@tag:Osokin2017_biogan]. Goldsborough et al. trained a GAN on fluorescent microscopy images and used latent space interpolation and arithmetic to reveal underlying responses to small molecule perturbations in cell lines [@tag:Goldsborough2017_cytogan].
It can often be informative to understand how the training data affects model learning. Toward this end, Koh and Liang [@tag:Koh2017_understanding] used influence functions, a technique from robust statistics, to trace a model's predictions back through the learning algorithm to identify the datapoints in the training set that had the most impact on a given prediction. A more free-form approach to interpretability is to visualize the activation patterns of the network on individual inputs and on subsets of the data. ActiVis and CNNvis [@tag:Kahng2017_activis; @tag:Liu2016_towards] are two frameworks that enable interactive visualization and exploration of large-scale deep learning models. An orthogonal strategy is to use a knowledge distillation approach to replace a deep learning model with a more interpretable model that achieves comparable performance. Towards this end, Che et al. [@tag:Che2015_distill] used gradient boosted trees to learn interpretable healthcare features from trained deep models.
Finally, it is sometimes possible to train the model to provide justifications for its predictions. Lei et al. [@tag:Lei2016_rationalizing] used a generator to identify "rationales", which are short and coherent pieces of the input text that produce similar results to the whole input when passed through an encoder. The authors applied their approach to a sentiment analysis task and obtained substantially superior results compared to an attention-based method.
While deep learning lags behind most Bayesian models in terms of interpretability, the interpretability of deep learning is comparable to or exceeds that of many other widely-used machine learning methods such as random forests or SVMs. While it is possible to obtain importance scores for different inputs in a random forest, the same is true for deep learning. Similarly, SVMs trained with a nonlinear kernel are not easily interpretable because the use of the kernel means that one does not obtain an explicit weight matrix. Finally, it is worth noting that some simple machine learning methods are less interpretable in practice than one might expect. A linear model trained on heavily engineered features might be difficult to interpret as the input features themselves are difficult to interpret. Similarly, a decision tree with many nodes and branches may also be difficult for a human to make sense of.
There are several directions that might benefit the development of interpretability techniques. The first is the introduction of gold standard benchmarks that different interpretability approaches could be compared against, similar in spirit to how the ImageNet [@doi:10.1007/s11263-015-0816-y] and CIFAR [@url:https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf] datasets spurred the development of deep learning for computer vision. It would also be helpful if the community placed more emphasis on domains outside of computer vision. Computer vision is often used as the example application of interpretability methods, but it is not the domain with the most pressing need. Finally, closer integration of interpretability approaches with popular deep learning frameworks would make it easier for practitioners to apply and experiment with different approaches to understanding their deep learning models.
A lack of large-scale, high-quality, correctly labeled training data has impacted deep learning in nearly all applications we have discussed. The challenges of training complex, high-parameter neural networks from few examples are obvious, but uncertainty in the labels of those examples can be just as problematic. In genomics labeled data may be derived from an experimental assay with known and unknown technical artifacts, biases, and error profiles. It is possible to weight training examples or construct Bayesian models to account for uncertainty or non-independence in the data, as described in the TF binding example above. As another example, Park et al. [@doi:10.1371/journal.pcbi.1002957] estimated shared non-biological signal between datasets to correct for non-independence related to assay platform or other factors in a Bayesian integration of many datasets. However, such techniques are rarely placed front and center in any description of methods and may be easily overlooked.
For some types of data, especially images, it is straightforward to augment training datasets by splitting a single labeled example into multiple examples. For example, an image can easily be rotated, flipped, or translated and retain its label [@doi:10.1101/095794]. 3D MRI and 4D fMRI (with time as a dimension) data can be decomposed into sets of 2D images [@doi:10.1101/070441]. This can greatly expand the number of training examples but artificially treats such derived images as independent instances and sacrifices the structure inherent in the data. CellCnn trains a model to recognize rare cell populations in single-cell data by creating training instances that consist of subsets of cells that are randomly sampled with replacement from the full dataset [@tag:Arvaniti2016_rare_subsets].
Simulated or semi-synthetic training data has been employed in multiple biomedical domains, though many of these ideas are not specific to deep learning. Training and evaluating on simulated data, for instance, generating synthetic TF binding sites with position weight matrices [@tag:Shrikumar2017_reversecomplement] or RNA-seq reads for predicting mRNA transcript boundaries [@doi:10.1101/125229], is a standard practice in bioinformatics. This strategy can help benchmark algorithms when the available gold standard dataset is imperfect, but it should be paired with an evaluation on real data, as in the prior examples [@tag:Shrikumar2017_reversecomplement; @doi:10.1101/125229]. In rare cases, models trained on simulated data have been successfully applied directly to real data [@doi:10.1101/125229].
Data can be simulated to create negative examples when only positive training instances are available. DANN [@doi:10.1093/bioinformatics/btu703] adopts this approach to predict the pathogenicity of genetic variants using semi-synthetic training data from Combined Annotation-Dependent Depletion (CADD) [@doi:10.1038/ng.2892]. Though our emphasis here is on the training strategy, it should be noted that logistic regression outperformed DANN when distinguishing known pathogenic mutations from likely benign variants in real data. Similarly, a somatic mutation caller has been trained by injecting mutations into real sequencing datasets [@tag:Torracinta2016_sim]. This method detected mutations in other semi-synthetic datasets but was not validated on real data.
In settings where the experimental observations are biased toward positive instances, such as MHC protein and peptide ligand binding affinity [@doi:10.1101/054775], or the negative instances vastly outnumber the positives, such as high-throughput chemical screening [@tag:Lusci2015_irv], training datasets have been augmented by adding additional instances and assuming they are negative. There is some evidence that this can improve performance [@tag:Lusci2015_irv], but in other cases it was only beneficial when the real training datasets were extremely small [@doi:10.1101/054775]. Overall, training with simulated and semi-simulated data is a valuable idea for overcoming limited sample sizes but one that requires more rigorous evaluation on real ground-truth datasets before we can recommend it for widespread use. There is a risk that a model will easily discriminate synthetic examples but not generalize to real data.
Multimodal, multi-task, and transfer learning, discussed in detail below, can also combat data limitations to some degree. There are also emerging network architectures, such as Diet Networks for high-dimensional SNP data [@tag:Romero2017_diet]. These use multiple networks to drastically reduce the number of free parameters by first flipping the problem and training a network to predict parameters (weights) for each input (SNP) to learn a feature embedding. This embedding (e.g. from principal component analysis, per class histograms, or a Word2vec [@tag:Word2Vec] generalization) can be learned directly from input data or take advantage of other datasets or domain knowledge. Additionally, in this task the features are the examples, an important advantage when it is typical to have 500 thousand or more SNPs and only a few thousand patients. Finally, this embedding is of a much lower dimension, allowing for a large reduction in the number of free parameters. In the example given, the number of free parameters was reduced from 30 million to 50 thousand, a factor of 600.
Efficiently scaling deep learning is challenging, and there is a high computational cost (e.g. time, memory, and energy) associated with training neural networks and using them to make predictions. This is one of the reasons why neural networks have only recently found widespread use [@tag:Schmidhuber2014_dnn_overview].
Many have sought to curb these costs, with methods ranging from the very applied (e.g. reduced numerical precision [@tag:Gupta2015_prec; @tag:Bengio2015_prec; @tag:Sa2015_buckwild; @tag:Hubara2016_qnn]) to the exotic and theoretic (e.g. training small networks to mimic large networks and ensembles [@tag:Caruana2014_need; @tag:Hinton2015_dark_knowledge]). The largest gains in efficiency have come from computation with GPUs [@tag:Raina2009_gpu; @tag:Vanhoucke2011_cpu; @tag:Seide2014_parallel; @tag:Hadjas2015_cc; @tag:Edwards2015_growing_pains; @tag:Schmidhuber2014_dnn_overview], which excel at the matrix and vector operations so central to deep learning. The massively parallel nature of GPUs allows additional optimizations, such as accelerated mini-batch gradient descent [@tag:Vanhoucke2011_cpu; @tag:Seide2014_parallel; @tag:Su2015_gpu; @tag:Li2014_minibatch]. However, GPUs also have limited memory, making networks of useful size and complexity difficult to implement on a single GPU or machine [@tag:Raina2009_gpu; @tag:Krizhevsky2013_nips_cnn]. This restriction has sometimes forced computational biologists to use workarounds or limit the size of an analysis. Chen et al. [@tag:Chen2016_gene_expr] inferred the expression level of all genes with a single neural network, but due to memory restrictions they randomly partitioned genes into two separately analyzed halves. In other cases, researchers limited the size of their neural network [@tag:Wang2016_protein_contact] or the total number of training instances [@tag:Gomezb2016_automatic]. Some have also chosen to use standard central processing unit (CPU) implementations rather than sacrifice network size or performance [@tag:Yasushi2016_cgbvs_dnn].
While steady improvements in GPU hardware may alleviate this issue, it is unclear whether advances will occur quickly enough to keep pace with the growing biological datasets and increasingly complex neural networks. Much has been done to minimize the memory requirements of neural networks [@tag:CudNN; @tag:Caruana2014_need; @tag:Gupta2015_prec; @tag:Bengio2015_prec; @tag:Sa2015_buckwild; @tag:Chen2015_hashing; @tag:Hubara2016_qnn], but there is also growing interest in specialized hardware, such as field-programmable gate arrays (FPGAs) [@tag:Edwards2015_growing_pains; @tag:Lacey2016_dl_fpga] and application-specific integrated circuits (ASICs) [@arxiv:1704.04760]. Less software is available for such highly specialized hardware [@tag:Lacey2016_dl_fpga]. But specialized hardware promises improvements in deep learning at reduced time, energy, and memory [@tag:Edwards2015_growing_pains]. Specialized hardware may be a difficult investment for those not solely interested in deep learning, but for those with a deep learning focus these solutions may become popular.
Distributed computing is a general solution to intense computational requirements and has enabled many large-scale deep learning efforts. Some types of distributed computation [@tag:Mapreduce; @tag:Graphlab] are not suitable for deep learning [@tag:Dean2012_nips_downpour], but much progress has been made. There now exist a number of algorithms [@tag:Dean2012_nips_downpour; @tag:Dogwild; @tag:Sa2015_buckwild], tools [@tag:Moritz2015_sparknet; @tag:Meng2016_mllib; @tag:TensorFlow], and high-level libraries [@tag:Keras; @tag:Elephas] for deep learning in a distributed environment, and it is possible to train very complex networks with limited infrastructure [@tag:Coates2013_cots_hpc]. Besides handling very large networks, distributed or parallelized approaches offer other advantages, such as improved ensembling [@tag:Sun2016_ensemble] or accelerated hyperparameter optimization [@tag:Bergstra2011_hyper; @tag:Bergstra2012_random].
Cloud computing, which has already seen wide adoption in genomics [@tag:Schatz2010_dna_cloud], could facilitate easier sharing of the large datasets common to biology [@tag:Gerstein2016_scaling; @tag:Stein2010_cloud], and may be key to scaling deep learning. Cloud computing affords researchers flexibility, and enables the use of specialized hardware (e.g. FPGAs, ASICs, GPUs) without major investment. As such, it could be easier to address the different challenges associated with the multitudinous layers and architectures available [@tag:Krizhevsky2014_weird_trick]. Though many are reluctant to store sensitive data (e.g. patient electronic health records) in the cloud, secure, regulation-compliant cloud services do exist [@tag:RAD2010_view_cc].
A robust culture of data, code, and model sharing would speed advances in this domain. The cultural barriers to data sharing in particular are perhaps best captured by the use of the term "research parasite" to describe scientists who use data from other researchers [@doi:10.1056/NEJMe1516564]. A field that honors only discoveries and not the hard work of generating useful data will have difficulty encouraging scientists to share their hard-won data. It's precisely those data that would help to power deep learning in the domain. Efforts are underway to recognize those who promote an ecosystem of rigorous sharing and analysis [@doi:10.1038/ng.3830].
The sharing of high-quality, labeled datasets will be especially valuable. In addition, researchers who invest time to preprocess datasets to be suitable for deep learning can make the preprocessing code (e.g. Basset [@tag:Kelley2016_basset] and variationanalysis [@tag:Torracinta2016_deep_snp]) and cleaned data (e.g. MoleculeNet [@tag:Wu2017_molecule_net]) publicly available to catalyze further research. However, there are complex privacy and legal issues involved in sharing patient data that cannot be ignored. Solving these issues will require increased understanding of privacy risks and standards specifying acceptable levels. In some domains high-quality training data has been generated privately, i.e. high-throughput chemical screening data at pharmaceutical companies. One perspective is that there is little expectation or incentive for this private data to be shared. However, data are not inherently valuable. Instead, the insights that we glean from them are where the value lies. Private companies may establish a competitive advantage by releasing data sufficient for improved methods to be developed. Recently, Ramsundar et al. did this with an open source platform DeepChem, where they released four privately generated datasets [@doi:10.1021/acs.jcim.7b00146].
Code sharing and open source licensing is essential for continued progress in this domain. We strongly advocate following established best practices for sharing source code, archiving code in repositories that generate digital object identifiers, and open licensing [@doi:10.1126/science.aah6168] regardless of the minimal requirements, or lack thereof, set by journals, conferences, or preprint servers. In addition, it is important for authors to share not only code for their core models but also scripts and code used for data cleaning (see above) and hyperparameter optimization. These improve reproducibility and serve as documentation of the detailed decisions that impact model performance but may not be exhaustively captured in a manuscript's methods text.
Because many deep learning models are often built using one of several popular software frameworks, it is also possible to directly share trained predictive models. The availability of pre-trained models can accelerate research, with image classifiers as an apt example. A pre-trained neural network can be quickly fine-tuned on new data and used in transfer learning, as discussed below. Taking this idea to the extreme, genomic data has been artificially encoded as images in order to benefit from pre-trained image classifiers [@tag:Poplin2016_deepvariant]. "Model zoos"---collections of pre-trained models---are not yet common in biomedical domains but have started to appear in genomics applications [@tag:Angermueller2016_single_methyl; @tag:Dragonn]. However, it is important to note that sharing models trained on individual data requires great care because deep learning models can be attacked to identify examples used in training. One possible solution to protect individual samples includes training models under differential privacy [@doi:10.1145/2976749.2978318], which has been used in the biomedical domain [@doi:10.1101/159756]. We discussed this issue as well as recent techniques to mitigate these concerns in the patient categorization section.
DeepChem [@tag:AltaeTran2016_one_shot; @tag:DeepChem; @tag:Wu2017_molecule_net] and DragoNN [@tag:Dragonn] exemplify the benefits of sharing pre-trained models and code under an open source license. DeepChem, which targets drug discovery and quantum chemistry, has actively encouraged and received community contributions of learning algorithms and benchmarking datasets. As a consequence, it now supports a large suite of machine learning approaches, both deep learning and competing strategies, that can be run on diverse test cases. This realistic, continual evaluation will play a critical role in assessing which techniques are most promising for chemical screening and drug discovery. Like formal, organized challenges such as the ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge [@tag:Dream_tf_binding], DeepChem provides a forum for the fair, critical evaluations that are not always conducted in individual methodological papers, which can be biased toward favoring a new proposed algorithm. Likewise DragoNN (Deep RegulAtory GenOmic Neural Networks) offers not only code and a model zoo but also a detailed tutorial and partner package for simulating training data. These resources, especially the ability to simulate datasets that are sufficiently complex to demonstrate the challenges of training neural networks but small enough to train quickly on a CPU, are important for training students and attracting machine learning researchers to problems in genomics and healthcare.
The fact that biomedical datasets often contain a limited number of instances or labels can cause poor performance of deep learning algorithms. These models are particularly prone to overfitting due to their high representational power. However, transfer learning techniques, also known as domain adaptation, enable transfer of extracted patterns between different datasets and even domains. This approach consists of training a model for the base task and subsequently reusing the trained model for the target problem. The first step allows a model to take advantage of a larger amount of data and/or labels to extract better feature representations. Transferring learned features in deep neural networks improves performance compared to randomly initialized features even when pre-training and target sets are dissimilar. However, transferability of features decreases as the distance between the base task and target task increases [@tag:Yosinski2014].
In image analysis, previous examples of deep transfer learning applications proved large-scale natural image sets [@tag:Russakovsky2015_imagenet] to be useful for pre-training models that serve as generic feature extractors for various types of biological images [@tag:Zhang2015_multitask_tl; @tag:Zeng2015; @tag:Angermueller2016_dl_review; @tag:Pawlowski2016]. More recently, deep learning models predicted protein sub-cellular localization for proteins not originally present in a training set [@tag:Parnamaa2017]. Moreover, learned features performed reasonably well even when applied to images obtained using different fluorescent labels, imaging techniques, and different cell types [@tag:Kraus2017_deeploc]. However, there are no established theoretical guarantees for feature transferability between distant domains such as natural images and various modalities of biological imaging. Because learned patterns are represented in deep neural networks in a layer-wise hierarchical fashion, this issue is usually addressed by fixing an empirically chosen number of layers that preserve generic characteristics of both training and target datasets. The model is then fine-tuned by re-training top layers on the specific dataset in order to re-learn domain-specific high level concepts (e.g. fine-tuning for radiology image classification [@tag:Rajkomar2017_radiographs]). Fine-tuning on specific biological datasets enables more focused predictions.
In genomics, the Basset package [@tag:Kelley2016_basset] for predicting chromatin accessibility was shown to rapidly learn and accurately predict on new data by leveraging a model pre-trained on available public data. To simulate this scenario, authors put aside 15 of 164 cell type datasets and trained the Basset model on the remaining 149 datasets. Then, they fine-tuned the model with one training pass of each of the remaining datasets and achieved results close to the model trained on all 164 datasets together. In another example, Min et al. [@tag:Min2016_deepenhancer] demonstrated how training on the experimentally-validated FANTOM5 permissive enhancer dataset followed by fine-tuning on ENCODE enhancer datasets improved cell type-specific predictions, outperforming state-of-the-art results. In drug design, general RNN models trained to generate molecules from the ChEMBL database have been fine-tuned to produce drug-like compounds for specific targets [@tag:Segler2017_drug_design; @tag:Olivecrona2017_drug_design].
Related to transfer learning, multimodal learning assumes simultaneous learning from various types of inputs, such as images and text. It can capture features that describe common concepts across input modalities. Generative graphical models like RBMs, deep Boltzmann machines, and DBNs, demonstrate successful extraction of more informative features for one modality (images or video) when jointly learned with other modalities (audio or text) [@tag:Ngiam2011]. Deep graphical models such as DBNs are well-suited for multimodal learning tasks because they learn a joint probability distribution from inputs. They can be pre-trained in an unsupervised fashion on large unlabeled data and then fine-tuned on a smaller number of labeled examples. When labels are available, convolutional neural networks are ubiquitously used because they can be trained end-to-end with backpropagation and demonstrate state-of-the-art performance in many discriminative tasks [@tag:Angermueller2016_dl_review].
Jha et al. [@tag:Jha2017_integrative_models] showed that integrated training delivered better performance than individual networks. They compared a number of feed-forward architectures trained on RNA-seq data with and without an additional set of CLIP-seq, knockdown, and over-expression based input features. The integrative deep model generalized well for combined data, offering a large performance improvement for alternative splicing event estimation. Chaudhary et al. [@tag:Chaudhary2017_multiom_liver_cancer] trained a deep autoencoder model jointly on RNA-seq, miRNA-seq, and methylation data from TCGA to predict survival subgroups of hepatocellular carcinoma patients. This multimodal approach that treated different omic data types as different modalities outperformed both traditional methods (principal component analysis) and single-omic models. Interestingly, multi-omic model performance did not improve when combined with clinical information, suggesting that the model was able to capture redundant contributions of clinical features through their correlated genomic features. Chen et al. [@tag:Chen2015_trans_species] used deep belief networks to learn phosphorylation states of a common set of signaling proteins in primary cultured bronchial cells collected from rats and humans treated with distinct stimuli. By interpreting species as different modalities representing similar high-level concepts, they showed that DBNs were able to capture cross-species representation of signaling mechanisms in response to a common stimuli. Another application used DBNs for joint unsupervised feature learning from cancer datasets containing gene expression, DNA methylation, and miRNA expression data [@tag:Liang2015_exprs_cancer]. This approach allowed for the capture of intrinsic relationships in different modalities and for better clustering performance over conventional k-means.
Multimodal learning with CNNs is usually implemented as a collection of individual networks in which each learns representations from single data type. These individual representations are further concatenated before or within fully-connected layers. FIDDLE [@tag:Eser2016_fiddle] is an example of a multimodal CNN that represents an ensemble of individual networks that take NET-seq, MNase-seq, ChIP-seq, RNA-seq, and raw DNA sequence as input to predict transcription start sites. The combined model radically improves performance over separately trained datatype-specific networks, suggesting that it learns the synergistic relationship between datasets.
Multi-task learning is an approach related to transfer learning. In a multi-task learning framework, a model learns a number of tasks simultaneously such that features are shared across them. DeepSEA [@tag:Zhou2015_deep_sea] implemented multi-task joint learning of diverse chromatin factors from raw DNA sequence. This allowed a sequence feature that was effective in recognizing binding of a specific TF to be simultaneously used by another predictor for a physically interacting TF. Similarly, TFImpute [@tag:Qin2017_onehot] learned information shared across transcription factors and cell lines to predict cell-specific TF binding for TF-cell line combinations. Yoon et al. [@tag:Yoon2016_cancer_reports] demonstrated that predicting the primary cancer site from cancer pathology reports together with its laterality substantially improved the performance for the latter task, indicating that multi-task learning can effectively leverage the commonality between two tasks using a shared representation. Many studies employed multi-task learning to predict chemical bioactivity [@tag:Dahl2014_multi_qsar; @tag:Ramsundar2015_multitask_drug] and drug toxicity [@tag:Mayr2016_deep_tox; @tag:Hughes2016_macromol_react]. Kearnes et al. [@tag:Kearnes2016_admet] systematically compared single-task and multi-task models for ADMET properties and found that multi-task learning generally improved performance. Smaller datasets tended to benefit more than larger datasets.
Multi-task learning is complementary to multimodal and transfer learning. All three techniques can be used together in the same model. For example, Zhang et al. [@tag:Zhang2015_multitask_tl] combined deep model-based transfer and multi-task learning for cross-domain image annotation. One could imagine extending that approach to multimodal inputs as well. A common characteristic of these methods is better generalization of extracted features at various hierarchical levels of abstraction, which is attained by leveraging relationships between various inputs and task objectives.
Despite demonstrated improvements, transfer learning approaches pose challenges. There are no theoretically sound principles for pre-training and fine-tuning. Best practice recommendations are heuristic and must account for additional hyper-parameters that depend on specific deep architectures, sizes of the pre-training and target datasets, and similarity of domains. However, similarity of datasets and domains in transfer learning and relatedness of tasks in multi-task learning is difficult to access. Most studies address these limitations by empirical evaluation of the model. Unfortunately, negative results are typically not reported. A deep CNN trained on natural images boosts performance in radiographic images [@tag:Rajkomar2017_radiographs]. However, due to differences in imaging domains, the target task required either re-training the initial model from scratch with special pre-processing or fine-tuning of the whole network on radiographs with heavy data augmentation to avoid overfitting. Exclusively fine-tuning top layers led to much lower validation accuracy (81.4 versus 99.5). Fine-tuning the aforementioned Basset model with more than one pass resulted in overfitting [@tag:Kelley2016_basset]. DeepChem successfully improved results for low-data drug discovery with one-shot learning for related tasks. However, it clearly demonstrated the limitations of cross-task generalization across unrelated tasks in one-shot models, specifically nuclear receptor assays and patient adverse reactions [@tag:AltaeTran2016_one_shot].
In the medical domain, multimodal, multi-task and transfer learning strategies not only inherit most methodological issues from natural image, text, and audio domains, but also pose domain-specific challenges. There is a compelling need for the development of privacy-preserving transfer learning algorithms, such as Private Aggregation of Teacher Ensembles [@tag:Papernot2017_pate]. We suggest that these types of models deserve deeper investigation to establish sound theoretical guarantees and determine limits for the transferability of features between various closely related and distant learning tasks.