Deep learning enhanced deciphering of brain activity maps for discovery of therapeutics for brain disorders

Abstract

This study presents an artificial intelligence enhanced in vivo screening platform, DeepBAM, which enables deep learning of large-scale whole brain activity maps (BAMs) from living, drug-responsive larval zebrafish for neuropharmacological prediction. Automated microfluidics and high-speed microscopy are utilized to achieve high-throughput in vivo phenotypic screening for generating the BAM library. Deep learning is applied to deconvolve the pharmacological information from the BAM library and to predict the therapeutical potential of non-clinical compounds without any prior information about the chemicals. For a validation set composed of blinded clinical neuro-drugs, several potent anti-Parkinson's disease and anti-epileptic drugs are predicted with nearly 45% accuracy. The prediction capability of DeepBAM is further tested with a set of nonclinical compounds, revealing the pharmaceutical potential in 80% of the anti-epileptic and 36% of the anti-Parkinson predictions. These data support the notion of systems-level phenotyping in combination with machine learning to aid therapeutics discovery for brain disorders.

Keywords: Biomedical Engineering; Pharmacology.

Graphical abstract

Figure 1

Overview of proposed DeepBAM platform for high-throughput brain activity mapping. DeepBAM deploys drug screening in awake larval zebrafish, and several potent anti-Parkinson and anti-epileptic drugs are predicted

Figure 2

Development of the DeepBAM platform (A) There are a total of 451 compounds, including the training set (330 drugs with ATC code) and the test set (121 compounds without ATC code). The distribution of the drug library containing clinically used drugs with ATC codes and nonclinical compounds. The abbreviations for different ATC categories are A for the alimentary tract and metabolism, C for the cardiovascular system, M for the musculoskeletal system, N for the nervous system, R for the respiratory system, S for sensory organs, V for various other drugs. Non-ATC indicates non-clinical compounds without any ATC codes. (B) Schematic figure showing the build of deep convolutional autoencoder for comprehensive feature learning from T-score BAMs. (C) Implementation of unsupervised clustering for discovering subgroups and deep neural network for predicting identified subgroups.

Figure 3

Consensus clustering reveals the functional diversity of the drugs (A) Data augmentation by combination-permutation $C_5^3$ strategy to compute the ten T-score BAMs for one drug. (B) The 5-dimensional pheno-prints (featured by the top 5 principal components) for ten DeepBAM clusters. (C) Consensus clustering identified ten phenotypic DeepBAM clusters. The heatmap was the consensus matrix illustrating the frequency of the scenario in which two compounds in a pair were clustered together. The color of the heatmap is proportional to the frequency scores of the consensus matrix, ranging from 0 to 1. (D) The representative cluster patterns were derived by taking the mean of all T-score BAMs for drugs in the ten clusters.

Figure 4

Functional association of clusters with N03 functional drug categories (A) Schematic of the hypergeometric test for overrepresentation of the N03 ATC category by phenotypic BAM cluster 2. (B) BAMs and chemical structures of the N03 drugs in cluster2 with similar structure. (C) Typical representation overlapped drugs between N03 and cluster2.

Figure 5

Functional association of clusters with N04 functional drug categories (A) Schematic of the hypergeometric test for overrepresentation of N04 (anti-Parkinson drugs) by phenotypic BAM cluster 8. (B) Molecular similarity searching based on the MACCS key fingerprints (0-budipine, 1-metixene, 2-piribedil, 3-ropinirole, 4-rasagiline, 5-benzhexol, 6-pergolide, 7-apomorphine, 8-ethopropazine, 9-amantadine). (C) Chemical structures of metixene and ethopropazine. (D) Chemical structures of budipine and amantadine.

Figure 6

Functional prediction of nonclinical compounds as N03 (A) Predicted potent anti-epileptic compounds with the same molecular targets. (B) T-score BAMs of the other predicted compounds.

Figure 7

Prediction of the nonclinical compounds as N04 Five compounds are predicted to be potent N04 anti-Parkinson drugs, with the support of literature.

Figure 8

Validation of predicted anti-epileptic compound by in larval zebrafish (A) Schematic of the setup for the behavioral monitoring and testing in larval zebrafish. (B) The experimental protocol for therapeutic insulation (predicted compounds), seizure induction, and locomotive stimulation in zebrafish larvae. Following a 4-h drug exposure, the zebrafish larvae received a 2-min light stimulation, which typically elicits a rapid increase in their movement for the following 15 min; failure to exhibit this agitated moving pattern suggests the larva was sedated. Afterward, PTZ was applied for seizure induction, and the number of seizures per larva and the average seizure frequency at each concentration were quantified for 15 min post-PTZ exposure. (C) The seizure counts for the tested compounds (SB 216763, Volinanserin, NNC 711, AR A014418, and 7,8-dihydroxyflavone) at different concentrations (10, 25, and 50 μM). (D) Quantification of the sedative index for the tested compounds was calculated by taking the inverse of change in moving distance before and after light stimulation. For (C) and (D), successful prediction was shaded in light green, larva pre-treated with DMSO were used as a control. “n.s.” stands for no significant difference, error bars indicate s.e.m., n = 12, ∗∗∗∗p < 0.0001, ∗∗∗p < 0.001, ∗∗p < 0.01, ∗p < 0.05 by one way ANOVA.