Opportunities And Obstacles For Deep Learning In Biology And Medicine

Key results

1. SuggestedDeep learning now equals or exceeds earlier state of the art in many tasks but, in the authors' judgement, has not yet transformed any domain; it is most mature for image tasks resembling natural photos and is expected to augment, not replace, experts.

   Clinician-comparable classifiers exist for skin lesions, tuberculosis, diabetic retinopathy and diabetic macular edema; a lymph-node cancer model erred about twice as often as a pathologist (whose error was about 3%), though model and pathologist errors were not strongly correlated; RaptorX-Contact ranked first by F₁ in CASP12; DeepVariant beat GATK. · Conclusions; Imaging applications in healthcare; Protein secondary and tertiary structure; Sequencing and variant calling

2. ShownGains over classical baselines are frequently modest and sometimes absent: non-neural or linear models matched or beat deep networks for transcription factor binding, splicing variant effects, structure-based binding affinity and small-data compound screening.

   FactorNet made the ENCODE-DREAM top three but a non-neural model slightly beat it; a hexamer linear model was three times as accurate as a deep framework at predicting exon skipping (SNPs in three genes); random forests generally beat neural nets on PDBBind; DeepTox topped 9 of 15 Tox21 targets, often narrowly; on 1547 BACE-1 compounds, random forest and deep net were on par. · Splicing; Transcription factors; Ligand-based prediction of bioactivity; Structure-based prediction of bioactivity

3. SuggestedScarce, costly expert labels limit deep learning in nearly all reviewed domains; transfer learning, weak supervision, augmentation and multi-task training mitigate this, but transfer helps less as source and target tasks diverge.

   Expert-validated eMERGE/PheKB sets mostly hold 100 to 300 patients; NegBio negation filtering (~90% F₁) enabled 112,120 weakly labeled chest X-rays from 30,805 patients; Basset pretrained on 149 of 164 cell types, fine-tuned on the other 15, neared the full model; tuning only top layers on radiographs gave 81.4 vs 99.5 validation accuracy for whole-network approaches. · Imaging applications in healthcare; Generating ground-truth labels can be expensive…; Data limitations; Multimodal, multi-task, and transfer learning

4. SuggestedPrivacy rules and attacks on trained models restrict sharing of health data and models; differentially private generative models look promising, while uncertainty estimates are largely ignored and interpretability methods lack gold-standard benchmarks.

   Differentially private generative adversarial networks mimicked SPRINT trial data, and models trained on the synthetic records performed only slightly below real-data models; dropout-based uncertainty referrals let a diabetic retinopathy system meet National Health Service sensitivity and specificity levels; networks can output over 99.99% confidence on noise-like inputs. · Data sharing is hampered by standardization and…; Uncertainty quantification; Interpretation; Future outlook

5. ShownBenchmark metrics are often non-comparable across studies and misleading under class imbalance: areas under the ROC (AUROC) and precision-recall (AUPR) curves can look excellent while a genomic classifier is useless for targeted validation, so metrics must fit the downstream use.

   Basset reported a gkm-SVM mean AUPR of 0.322 over 164 cell types, whereas DeepEnhancer reported 0.899 on nine cell types for that baseline; with 1% of regions bound, a hypothetical classifier with 0% recall below 25% false discovery rate (FDR) still scores AUROC 0.9967, and transcription factor models can exceed chromosome-wide AUROC 0.99 with 0% recall at 10% FDR. · Enhancers; Evaluation metrics for imbalanced classification