Directed Evolution of a Selective and Sensitive Serotonin Sensor via Machine Learning

Abstract

Serotonin plays a central role in cognition and is the target of most pharmaceuticals for psychiatric disorders. Existing drugs have limited efficacy; creation of improved versions will require better understanding of serotonergic circuitry, which has been hampered by our inability to monitor serotonin release and transport with high spatial and temporal resolution. We developed and applied a binding-pocket redesign strategy, guided by machine learning, to create a high-performance, soluble, fluorescent serotonin sensor (iSeroSnFR), enabling optical detection of millisecond-scale serotonin transients. We demonstrate that iSeroSnFR can be used to detect serotonin release in freely behaving mice during fear conditioning, social interaction, and sleep/wake transitions. We also developed a robust assay of serotonin transporter function and modulation by drugs. We expect that both machine-learning-guided binding-pocket redesign and iSeroSnFR will have broad utility for the development of other sensors and in vitro and in vivo serotonin detection, respectively.

Keywords: OSTA; SERT; fear-learning; fiber photometry; fluorescence protein sensor; iSeroSnFR; machine learning; serotonin; sleep-wake; social behaviors.

Figure 1.. Using Machine Learning to Evolve Binding Proteins

(A) Overview of machine learning method. (B) Schematic showing conversion of an acetylcholine (ACh)-binding protein to a serotonin (5-HT)-binding protein. ACh and 5-HT were docked into the binding pocket of AChSnFR0.4 using Rosetta. Statistical modeling was performed on these models, and promising mutations were synthesized and tested (see Figure S2 and Table S3). iSeroSnFR0.0 was chosen as a starting point for statistical modeling. Positions 66, 143, 170, and 188 were selected for further mutation. (C) The binding pocket of iSeroSnFR0.0 was simulated using Rosetta and 5-HT (magenta) docked. Top-ranked positions are labeled (cyan). (D) Table of DNA libraries created, and number of variants screened from each library. (E) DNA libraries were generated, transformed into bacteria, grown, and lysed. Lysate was then screened with 10 mM 5-HT and compared to the parent sensor (iSeroSnFR0.0). (F) Heatmap of the contribution of each mutation at each position screened, as predicted by the generalized linear model (GLM) (for additional information see Table S3). (G) Combinations of mutations predicted to be better than the parent (iSeroSnFR0.0) were synthesized and tested as purified protein with 10 mM 5-HT. Dashed line represents equivalency. (H) Protein from iSeroSnFR0.0 and the top variant (iSeroSnFR0.1) was purified and tested against multiple concentrations of 5-HT. Shaded area denotes 95% confidence interval. (I) Raincloud plot where iSeroSnFR0.1 was used as the parent for a second round of screening followed by GLM analysis. A small library (32 possible combinations) was generated based on the GLM results and screened (cyan), which led to the discovery of iSeroSnFR0.2. (J) Raincloud plot similar to (I), but using iSeroSnFR0.2 as the parent. This GLM-guided targeted library (96 possible combinations) was created and screened, leading to the discovery of iSeroSnFR. See also Figure S1 and Data S1.

Figure 2.. Affinity and Specificity of the Sensor

(A) Crystal structure of unliganded iSeroSnFR (PDB: 6PER). Mutations in iSeroSnFR relative to iAChSnFR0.6 are mapped onto the crystal structure (red). Positions interrogated by site-saturated mutagenesis (but not mutated in iSeroSnFR) are displayed in blue, mutations interrogated by Rosetta, but not SSM, in purple, and positions that were randomly mutated, in green. (B) Purified iSeroSnFR binding to 5-HT. (C) Purified iSeroSnFR binding to multiple ligands. Due to differential compound solubility, the values displayed match the following concentrations: octopamine, L-phenylalanine, 80 μM; 5-HTP, 85 μM; sertraline, 110 μM; L-DOPA, tyramine, escitalopram, citalopram, amoxapine, 125 μM; all other compounds were tested at either 100 or 105 μM. For the full concentration curve for each compound, see Figure S3. (D-F) Response of membrane-displayed iSeroSnFR in HEK293T cells. Representative images (D), and dose-response curves for higher concentrations (E) and lower concentrations (F). n = 3–4. (G–I) Response of membrane-displayed iSeroSnFR-PDGFR in cultured neurons. (G–I) Representative images (G), and dose-response curve for higher concentrations (H) and lower concentrations (I). n = 3–4. For raw traces, see Figure S5. (B, E, F, H, and I) Shaded area denotes 95% confidence interval. Scale bars represent 50 μm. Insets show magnifications of the points at low concentrations. See also Figure S4.

Figure 3.. Sensor Kinetics

(A–F) iSeroSnFR was purified and tested in a stopped-flow apparatus, with increasing concentrations of 5-HT. (A and B) Average traces showing full time courses for low (A) and high (B) concentrations. (D and E) Magnification of the first 100 ms and 1 s, respectively, of the data in (A) and (B), respectively, with double exponential fits shown. (C and F) The τ (1/rate constant) for each concentration for the slow phase (C) and the fast phase (F) was fit using the Hill equation. (G) Model of iSeroSnFR function showing two rate-limiting steps: isomerization between the inactive and active states, followed by binding of serotonin, for full fluorescence activation. N = 16–18 trials for each concentration. (H–R) Primary cultured neurons (H–Q) and HEK cells (R) were exposed to 200 μM caged-5-HT (PA-N-5-HT) and uncaged using 405 nm laser stimulation. (H) Representative image. (I) 5-HT was uncaged as noted. Traces represent a 9-trial average for each replicate. Biological replicates = 3. There was no image acquisition during 405 nm laser stimulation. (J and K) Data expanded from (I). Faded lines depict raw traces, dark lines represent average traces. (L) Data from (I) was plotted and fit with using the Hill equation. (M–R) 5-HT was uncaged as noted. Red lines represent uncaging epochs. (M and N) Line scans (128 × 1 pixel) at 5 kHz (M) and 250 Hz (N). (O–R) Frame scans (128 × 128 pixel) at 40 Hz. (Q) Red dot represents uncaging spot. Blue and green traces represent data from regions of interest outlined on the image (~2 μm and ~20 μm from the uncaged region). (R) Response of membrane-displayed iSeroSnFR and iAChSnFR to serotonin uncaging on HEK293T cells. For raw traces, see Figure S6. Scale bars represent 50 μm (H) and 10 μm (Q and R). See also Figure S5.

Figure 4.. Detection of Electrically or Behaviorally Triggered Release of 5-HT

(A–E) WT mice were injected with AAV2/9.CAG.iSeroSnFR.Nlgn into the dorsal striatum (DStr). Slices were prepared (300 μm) and imaged using one-photon photometry. (A and B) Schematic (A) and representative (B) image of sensor injection and expression. (C) Slices were stimulated using a monopolar saline-filled glass electrode (0.5 ms; 50 μA) at the frequencies noted. (D) Data from (C) fitted using the Hill equation. Shaded area represents 95% confidence interval. (E) Tetrodotoxin (TTX; 300 nM) was added to the perfusion solution, and the slice was stimulated at 40 Hz for 1 s (n = 11 slices from 3 mice). For more information, see Figure S7F. (F–O) Fiber-photometry recording of 5-HT release in response to fear-conditioning in BLA (F–J) and mPFC (K–O). Mice were injected with either AAV2/9.CAG-iSeroSnFR.Nlgn or AAV2/5.CAG-GFP (as a negative control) followed by optical fiber implantation into BLA (F and G) or mPFC (K and L). Yellow box indicates unconditioned stimulus (tone + house lights); pink box illustrates foot shock. Single-trial traces (H and M) or average trace across all trials (I and N). Shaded area represents SEM. Statistical comparison was made based on the average fluorescence for the 10 s before the cue onset (baseline) and the 10 s of cue presentation (J and O). n = 15 trials/animal, N = 9 BLA^iSeroSnFR, 3 BLA^GFP, 4 mPFC^iSeroSnFR, and 3 mPFC^GFP). See also Figure S5.

Figure 5.. Detection of 5-HT Release during Sleep-Wakefulness Cycles in BLA

(A) Mice were injected with either AAV2/9.CAG-iSeroSnFR.Nlgn or AAV2/5.CAG-GFP (as a negative control), and an optical fiber was implanted into basolateral amygdala (BLA). EEG screw electrodes and EMG wires were implanted to classify sleep-wake states. (B) Representative BLA^iSeroSnFR EEG spectrograms, EMG, and fiber photometry traces over time across sleep-wake cycles (left) and walking episode (right). (C) Temporal dynamics of iSeroSnFR (left) and GFP (right) activity during waking, NREM, and REM episodes (data from B, left) across time, normalized from onset to offset. Statistical comparisons of fluorescence levels were performed on the last 10% of data within each behavioral state (one-way ANOVA with Bonferroni correction). (D) Single-trial or averaged fluorescence change across all trials of iSeroSnFR from NREM to wake, REM to wake, wake to NREM, and NREM to REM transitions. Statistical comparisons of changes in fluorescence using and BLA^iSeroSnFR were made based on the average fluorescence over 15 s before and 15 s after the behavioral state transition. n = 15 trials/animal, N = 3 BLA^GFP, and N = 9 BLA^iSeroSnFR. Data represent mean ± SEM. *p < 0.05, **p < 0.01, ***p < 0.001, paired Student’s t test. See also Figures S6, S7, and S9.

Figure 6.. Application to Monitoring hSERT Transport

(A) Schematic of the oscillating stimulus transporter assay (OSTA). (B) Confirmation of ionic requirements of hSERT for 5-HT uptake: influx only occurs in the presence of 5-HT, Na⁺, and Cl⁻. (C) Ions driving hSERT-mediated 5-HT efflux: rates of efflux increase with K⁺ and are insensitive to Cl⁻. (D–F) Sodium dependence of 5-HT transport through hSERT. (D) Cells were subjected to a decreasing linear gradient of Na⁺ In the Influx buffer, with standardized “fiducial” bouts at saturating sodium concentration Interleaved. Red fluorescence (sulforhodamine 101 at 200 nM in fiducials) was used as a readout for the sodium stimulus (top panel), and green fluorescence (iSeroSnFR) reflected changes in cytosolic [5-HT] due to hSERT function (bottom panel). (E) Fits of single-cell (top) and grouped (bottom) 5-HT transport responses to the stimulus shown in (D). (F) Scatterplot of fitted parameters for individual cells: Hill coefficient versus K_m. (G–L) hSERT-mediated 5-HT transport responses to various pharmacological agents (as indicated) under a standardized stimulus. (M and N) MDMA-mediated 5-HT efflux: MDMA at 20 μM in the efflux buffer significantly increased the rate of 5-HT efflux compared to Na⁺ alone. (N) In the last epoch of the experiment, K⁺ was substituted for Na⁺ in the efflux buffer for comparison to other experiments.