Machine learning designs new GCGR/GLP-1R dual agonists with enhanced biological potency

Potency assaysPeptide potencies for cAMP accumulation had been experimentally decided for the activation of each hGCGR and hGLP-1R expressed in Chinese hamster ovary (CHO) cells for a set of 125 distinctive peptide sequence variants, following strategies described previously16,35. In transient secure CHO cell strains expressing human and mouse GLP-1R had been generated in-house utilizing customary strategies, as beforehand described16. CHO cells expressing both human GLP-1R or human GCGR had been distributed in assay buffer (Hanks balanced salt answer containing 0.1% BSA (Sigma-Aldrich) and 0.5 mM IBMX (Sigma-Aldrich)) in 384-well assay plates containing dilutions of take a look at peptides. After 30 min of incubation, cAMP ranges had been measured utilizing the cAMP dynamic 2 HTRF package (Cisbo) following the manufacturer’s suggestions. Fluorescence emissions at 665 nm and 620 nm following excitation at 320 nm had been detected utilizing an Envision reader (Perkin Elmer), and the information had been remodeled to % Delta F, as described within the manufacturer’s tips, earlier than EC50 dedication. All in vitro cell-based assay knowledge are introduced because the imply of n ≥ 3 impartial experiments, and all particular person EC50 measurements had been inside threefold of the geometric imply. The native peptide reference customary potency was inside threefold of the historic geometric imply for all assays.DataunitsThe GPCR-binding peptides thought of on this work solely comprise naturally occurring amino acids, so the fashions are usually not in a position to seize the impact of any chemical modifications of residues. The preliminary set of sequences was aligned utilizing MAFFT model 736 to disclose regularities in amino-acid occurrences throughout positions. We reasoned that sequence alignment may assist construction the information, thereby rising the predictive energy of the neural-network fashions. The aligned sequences had been truncated to L = 30 amino acids, and redundant sequences had been eliminated. The remaining set of sequences used on this research comprised N = 125 distinctive peptide sequences examined in opposition to human GPCR and GLP-1R receptors. Within this dataset, 122 information had been C-terminally amidated. The sequences had been subsequently encoded utilizing a one-hot illustration and used to coach numerous regression fashions.Data encodingTo encode the amino acid at every sequence place we used a one-hot (binary) illustration. Here, we thought of 21 classes: 20 amino acids and the hole image ‘-’, launched by alignment. Because practically all peptides utilized in these research (122/125) had been C-terminally amidated, we didn’t introduce an extra parameter to encode this characteristic. In this strategy, every peptide sequence of size L is transformed to a binary matrix S of measurement 21 × L, the entries of which point out the presence of an amino acid Ai on the given sequence web site, such that Sab = 1 if a = i and 0 elsewhere, ∀b ∈ {1, …, L}. The binary matrix is then re-shaped right into a vector: S21 × L → v1 × 21L. The alignment course of ensures that L = 30 for all peptides, such that every sequence is represented by the binary vector ({vin {{mathbb{R}}}^{1times 630}}).Evaluation metricsWe employed the next generally used regression metrics to guage the prediction accuracy of the fashions developed on this work.

1.

Root-mean-square error (r.m.s.e.): ({rm{r.m.s.e.}}={sqrt{frac{1}{N}mathop{sum }nolimits_{i = 1}^{N}{(;{y}_{i}-hat{y})}^{2}}})

2.

Mean absolute error (m.a.e.): ({rm{m.a.e.}}={frac{1}{N}mathop{sum }nolimits_{i = 1}^{N}| ;{y}_{i}-hat{y}|})

3.

Coefficient of dedication (R2): ({R}^{2}={1-frac{mathop{sum }nolimits_{i = 1}^{N}{(;{y}_{i}-hat{y})}^{2}}{mathop{sum }nolimits_{i = 1}^{N}{(;{y}_{i}-bar{y})}^{2}}})

Notation: yi, true worth of the goal for the ith pattern; ({hat{y}}), predicted worth of the goal for the ith pattern; ({bar{y}}), common worth of the goal; N, variety of examples within the batch.Neural-network mannequinWe used the Keras/Tensorflow practical API to construct the deep community model37. The first Conv1D layer in our mannequin has 256 filters, a kernel window of three amino acids, with out padding, and an L2 regularization penalty on the kernel, with weight = 0.01. The layer makes use of ReLU activation. We subsequent added batch normalization and a MaxPool1D operation with stride 2 and used Dropout = 0.5. The second Conv1D layer comprises 512 filters and the identical configuration of parameters as the primary layer, with an extra L2 regularization penalty on the bias time period, with weight = 0.01. The layer is activated with ReLU, adopted by batch normalization, MaxPool1D operation with stride 2, and Dropout = 0.5. The third convolutional layer has 128 filters; right here the padding preserves the form of the enter, and the kernel in addition to bias are regularized with L2. This layer is adopted by MaxPool1D operation with stride 2. Next, the output from convolutional layers is flattened and two dense layers terminate the community. The first dense layer contains 256 models, and the second layer has 64 models. Both layers are ReLU-activated. The remaining two dense layers with a single unit convert the mannequin output to the prediction. These layers are usually not activated.Network ensembleWe constructed a neural-network ensemble mannequin the place the ultimate prediction is given by the typical of the person predictions made by M = 12 separate copies of the mannequin. Different copies of the mannequin, skilled on the identical knowledge, differ of their predictions because of components such because the random initialization of the community parameters. Ensembling predictions over a number of copies of the mannequin has the impact of mitigating a few of this randomness and decreasing mannequin variance. The ensuing ensemble prediction is given by the typical of the ensemble aspect predictions.Model coaching and hyperparameter tuningTo alter capability and choose non-trainable mannequin parameters, the accessible knowledge had been used for efficiency validation. Initially, the dataset of 125 examples was divided into three subsets: 105 coaching sequences, ten sequences for validation and ten held-out sequences for remaining mannequin efficiency analysis (unseen throughout coaching). We carried out ten sixfold cross-validations splits with completely different seeds to separate the information, acquiring 60 (take a look at set measurement) × 10 = 600 knowledge factors (errors) in complete for every mannequin. Retraining on completely different knowledge splits allowed us to take note of the variance ensuing from coaching on completely different knowledge, along with the variance that arises because of the random mannequin initialization.For every baseline mannequin, we used the sklearn grid search (GridSearchCV) to search out the set of hyperparameters that present the most effective cross-validation efficiency (listed in Supplementary Table 2). Parameters for which the optimum worth differs between duties are marked with a double worth v1/v2 within the respective column of Supplementary Table 2, the place v1 is the optimum parameter worth for the GCGR job, and v2 is the optimum parameter worth for the GLP-1R job. For the neural networks, numerous configurations of layers, unit numbers and regularization had been tried, and we chosen the mannequin that gave the most effective efficiency on the validation set.In addition, to stop overfitting of the neural networks, we monitored efficiency utilizing early stopping. Training was terminated when the optimization loss reported on the validation set goes up after a specific variety of parameter updates. Here, we use the Early Stopping monitor applied within the Keras call-back module38. Deep fashions with 120 coaching examples (remaining fashions) had been skilled for as much as 1,500 epochs, monitoring the validation loss, with the endurance of 100 epochs. Each batch for the gradient step contained 25 samples. The deep fashions with 105 coaching examples used for validation had been skilled for as much as 1,500 epochs, monitoring the validation loss with the endurance of 75 epochs, and 20 examples per batch (every epoch had 5 parameter updates). Model coaching is illustrated in Supplementary Fig. 2.Baseline fashionsAll baseline regressors in Table 1 had been applied utilizing the sklearn Python module39. To affirm that the ML fashions don’t merely be taught the underlying potency distributions or amino-acid sequence compositions, we skilled management ensembles of multi-task neural networks utilizing the method described above, the place we (1) shuffled every peptide sequence used to coach the fashions and (2) shuffled the measured potencies between coaching examples. The ensuing management fashions make a lot bigger prediction errors; the outcomes are proven in Supplementary Fig. 4 and summarized in Supplementary Table 4. Finally, we applied a easy nearest-neighbours strategy through which the expected potency for a held-out take a look at sequence is predicted by the measured potency of the closest neighbour within the coaching knowledge. For every take a look at sequence we used the pairwise2 BioPython module with the BLOSUM62 matrix to attain alignments with each coaching sequence; within the case of a number of equidistant coaching sequences, the typical potency was reported. Results throughout sixfold cross-validation are summarized in Supplementary Table 4 and present that this strategy is outperformed by the ML fashions described above.We used a t-test (two-sided) to check whether or not the variations in mannequin efficiency had been important between the ensemble of multi-task neural networks and the opposite fashions. Distributions of 600 prediction errors (squared distinction between the true and predicted potency for every take a look at sequence) obtained for every mannequin for the GCGR and GLP-1R duties are proven in Supplementary Table 3. For every pair of fashions we take a look at the null speculation that the 2 impartial populations of error samples have the identical common values (we don’t assume equal variances). Supplementary Table 3 exhibits that at a confidence degree of 0.05, the multi-task neural-network ensemble performs considerably higher in all circumstances for the GLP-1R job, whereas the efficiency variations are insignificant in all besides one case for the GCGR job.Multi-task coachingMulti-task learning goals to enhance generalization and improve prediction accuracy by learning aims for a number of goal variables from shared representations28. The primary concept is that by coaching all duties utilizing shared hidden layers, every job advantages from the presence of the others, which act as regularizers, making the mannequin much less delicate to the specificity of a single target28,40. This is as a result of the shared layers are shared representations—the mannequin makes use of the identical weights on every job. The efficient variety of coaching examples is subsequently elevated, and overfitting on every job individually is diminished by optimizing the mannequin with respect to the typical knowledge noise40.We used the Kereas deep-learning framework to construct our mannequin (https://keras.io) utilizing the TensorFlow back-end37. The mannequin consists of eight totally related layers, comprising the enter layer, adopted by three 1D convolutional layers, three pooling layers and two dense layers on the backside of the mannequin, related to 2 remaining models that convert the output to real-valued predictions. The general goal is the weighted common of the loss for every of the 2 particular person duties:$${L}_{rm{complete}}={mathop{sum }limits_{i=1}^{okay=2}{alpha }_{i}{L}_{i}}$$
(1)
the place Li is the loss operate of the ith job, αi is the corresponding weight and okay denotes the variety of duties. We set α1 = α2 = 0.5 so that every loss contributes with equal weight to the general loss. We use the mean-squared-error (m.s.e.) because the loss for every job, ({rm{m.s.e.}}={frac{1}{n}mathop{sum }nolimits_{j = 1}^{n}{(;{y}_{j}-hat{y})}^{2}}), the place n is the variety of coaching examples per batch.Our multi-task neural-network mannequin shares all inside hidden layers between the duties. Two output models return the expected potencies, ({hat{y}}_{1}) and ({hat{y}}_{2}). The convolutional layers on the high of the mannequin are designed to encode the peptide representations. We use a kernel with a window measurement of three amino acids and stride equal to 1. Each convolutional layer is adopted by a max pooling layer, with stride equal to 2. We use batch normalization41 and Dropout42 for regularization. Each convolutional and dense layer is activated with ReLU43 activation. We skilled the mannequin with an optimization goal as given in equation (1) utilizing the Adam optimizer44. The remaining community was skilled on an equal variety of coaching examples for each duties, N = 120.Model-guided ligand designOur aim was to design peptide sequences with the next properties:

Highly energetic in opposition to each receptors:$$start{array}{r}{rm{Activity}}=left{start{array}{l}{log }_{10}{{{rm{EC}}}}_{50}^{rm{GCGR}}[M,] < {-11.5}quad {log }_{10}{{{rm{EC}}}}_{50}^{rm{GLP-1R}}[M,] < {-11.5}quad {{{rm{EC}}}}_{50}^{rm{GCGR}}/{{{rm{EC}}}}_{50}^{rm{GLP-1R}}approx {1}quad finish{array}proper.finish{array}$$ (2) Selectively energetic in the direction of GCGR:$$start{array}{r}{rm{Activity}}=left{start{array}{l}{log }_{10}{{{rm{EC}}}}_{50}^{rm{GCGR}}[M,] < {-11}quad {log }_{10}{{{rm{EC}}}}_{50}^{rm{GLP-1R}}[M,] > {-9}quad {{{rm{EC}}}}_{50}^{rm{GCGR}}/{{{rm{EC}}}}_{50}^{rm{GLP-1R}}approx {100}quad finish{array}proper.finish{array}$$
(3)

Selectively energetic in the direction of GLP-1R:$$start{array}{r}{rm{Activity}}=left{start{array}{l}{log }_{10}{{{rm{EC}}}}_{50}^{rm{GCGR}}[M,] > {-9}quad {log }_{10}{{{rm{EC}}}}_{50}^{rm{GLP-1R}}[M,] < {-11.5}quad {{{rm{EC}}}}_{50}^{rm{GLP-1R}}/{{{rm{EC}}}}_{50}^{rm{GCGR}}approx {100}quad finish{array}proper.finish{array}$$ (4) We use model-guided directed evolution, an optimization technique that makes an attempt to unravel the optimization drawback by imitating the pure evolutionary course of. In every successive era (iteration), a change within the sequence is proposed, adopted by the analysis of a health operate (right here, potency predicted by the ensemble of multi-task neural networks) and the most effective options are progressed to the subsequent era. This course of repeats till a passable answer is reached. In this work we assume that the ensemble of multi-task convolutional neural networks makes dependable predictions as much as three mutation steps from the closest training-set analogue sequence.We first generated all single-step mutations from every training-set sequence within the three teams of curiosity, eradicating any duplicates throughout the generated set, and any overlaps with the coaching set. Because every sequence within the preliminary alignment has a size of 30 amino acids and every place may be mutated to considered one of 19 amino acids (20 if the place is gapped), this provides 570 single-step mutants within the first era for every sequence within the coaching set, that's, 71,304 sequences, decreasing to 69,639 sequences after eradicating duplicates. We then used every mannequin to pick out the 50 greatest sequences for every of the three goal designs outlined above, and chosen the ten most numerous sequences as beginning factors for a second spherical of optimization. Note that within the first era for the multi-task CNN solely 5 candidate dual agonists had been discovered and used as mother and father for the second era. For the second era we repeated the method described for the primary era, and from the 50 greatest sequences for every group, we chosen 5 numerous sequences as mother and father for the third era. The whole course of was then repeated for a remaining era, taking the 50 sequences with the most effective predicted potencies inside every of the three teams, contemplating GCGR for group one.We recognized six biophysical properties that may be predicted from a sequence utilizing the ProtParam module (https://biopython.org/wiki/ProtParam) from the biopython Python package45: (1) the isoelectric level in impartial pH, (2) GRAVY (grand common of hydropathy)46, (3) the instability index47, (4) aromaticity, (5) the molar extinction coefficient and (6) molecular weight. We in contrast the expected worth for every designed peptide with the expected properties of peptides within the coaching set throughout the identical potency group. We ranked the 50 greatest sequences in every group by computing the variety of options whose values are inside one customary deviation of the imply calculated for the corresponding group of training-set sequences. As the final step of filtering, we predicted the secondary construction for every remaining candidate utilizing PSIPRED48 (http://bioinf.cs.ucl.ac.uk/psipred/) to verify that the chosen sequences are helical peptides. Using this rating, we chosen 5 remaining samples in every potency category—4 from the third era, and one from the primary era of mutants. We prioritized designed sequences with the smallest (first era) and largest (third era) distance from the coaching set. Sequences chosen with the ensemble of multi-task neural-network fashions experimentally examined on this research and mentioned in the primary textual content are listed in Table 2 (Supplementary Tables 7–11 present extra particulars).To study the similarity of peptides predicted by completely different fashions inside every potency profile, we used PCA, contemplating the five hundred one-hot-encoded sequences generated throughout all 5 in contrast fashions, the wild-type peptides—hGCG and hGLP-1—and their single-step mutants (551 hGCG and 570 hGLP-1), such that the projection was computed for an array [1,621 × 21L]. The projected knowledge are proven in Supplementary Fig. 7. The chosen remaining sequences listed in Supplementary Table 6 had been analysed when it comes to the full variety of mutations from wild-type and predicted potencies. The predictions made with completely different fashions are constant, as evidenced by the low values of ordinary deviation (<0.5) from the typical prediction computed throughout the fashions.To consider the data content material generated by the sequence design course of, we calculated the entropy throughout every set of designed sequences, and the relative entropy (Kullback–Leibler divergence, KL) between the distribution of amino acids at every sequence place estimated for the model-designed samples, and the coaching knowledge. KL divergence is the same as zero if and provided that the distribution of amino acids throughout the designed samples matches precisely the respective distribution of amino acids estimated from the training-set sequences. The relative entropy between two discrete distributions s(x) and t(x) is given by$$ t)={-mathop{sum }limits_{i=1}^{21}s({x}_{i}){log }_{21}frac{t({x}_{i})}{s({x}_{i})}}$$ (5) the place xi is among the 21 symbols at a specific place j. We additionally measured the dependence between model-generated samples and the coaching knowledge utilizing mutual info (MI). Given two alignment columns A and B, every with discrete distributions of amino acids, their MI may be calculated as$${I(A;,B)}={mathop{sum }limits_{i}^{Okay=21}mathop{sum }limits_{j}^{L=21}p({x}_{i},,{y}_{j}){log }_{21}frac{p({x}_{i},,{y}_{j})}{p({x}_{i})p(;{y}_{j})}= p(a)p(b))}$$ (6) the place xi is among the 21 symbols at place A, and yj is among the 21 symbols at place B. The MI describes the discount of uncertainty concerning the amino acid at place i in our generated samples after we are informed what the amino acid at place i within the coaching knowledge is. The larger the worth, the extra dependent the variables.Predicted properties of pure homologuesAs described in the primary textual content, we used our multi-task neural-network ensemble mannequin to make predictions for pure GCG and GLP-1 peptide orthologues which can be present in numerous organisms, recognized utilizing BLASTp to look the NCBI RefSeq49 database to determine non-redundant proglucagon sequences from numerous organisms throughout numerous phylogenetic teams. In vertebrates, the pre-proglucagon polypeptide is a product of the GCG gene, which encodes 4 peptide hormones: glucagon, glucagon-like peptide-1, glucagon-like peptide-2 and oxyntomodulin50. In people, pre-proglucagon has 180 amino acids and is cleaved to yield proglucagon (158 amino acids), which lacks the N-terminal signalling sequence. Proglucagon is subsequently processed by the prohormone convertases 1, 2 and 350 to provide, amongst different merchandise, the 29-amino-acid-long GCG (in human, positions 53–81, PSCK2) and the 30-amino-acid-long GLP-17−36 (positions 98–127, PSCK1), that are the main target of this work.We recognized 450 preliminary information, which we aligned utilizing MAFFT model 736 with default parameters to assemble a a number of sequence alignment (MSA). We additionally eliminated duplicated sequence isoforms, leaving a single consultant for every species. Columns with low occupancy (f < 30% amino acids) had been additionally eliminated, leaving 294 distinctive samples, such that the ultimate MSA contained 294 rows (species) and 179 columns (positions). MSA areas similar to the human GCG sequence (positions 53–81) and the GLP-1 human sequence (positions 98–127) had been extracted, yielding two units of corresponding homologues. Species that lacked both a GCG or GLP-1 sequence within the alignment had been additional eliminated to yield two remaining peptide sequence units, every comprising 288 orthologous sequences. The record of species and NCBI accession numbers, in addition to the corresponding peptide sequences, are supplied in Supplementary Tables 15 and 16.Reporting summaryFurther info on analysis design is offered within the Nature Portfolio Reporting Summary linked to this Article.
https://www.nature.com/articles/s41557-024-01532-x