Utilizing compound-protein machine learning representations to generate bioactivity predictions
The present disclosure relates to systems, non-transitory computer-readable media, and methods that utilizing compound-protein machine learning representations to generate target results. For example, the disclosed systems can utilize a compound-protein interaction machine learning model to generate a compound-protein machine learning representation for compound protein pairs. The disclosed systems can utilize the compound-protein machine learning representation to train and utilize other target machine learning models in generating predicted bioactivity results. For example, the disclosed systems train a target machine learning model from compound-protein machine learning representations to generate ADMET predictions and/or biological perturbation program predictions. Furthermore, the disclosed systems can utilize one or more explainability models in conjunction with target machine learning models trained based on compound-protein machine learning representations to identify proteins that contribute to predicted bioactivity results.
1 . A computer-implemented method comprising:
identifying a plurality of compound-protein pairs comprising a training compound matched to a plurality of proteins;
generating, utilizing a compound-protein interaction machine learning model from the plurality of compound-protein pairs, a plurality of binding scores between the training compound and the plurality of proteins;
generating a compound-protein machine learning representation comprising the plurality of binding scores between the training compound and the plurality of proteins;
generating, utilizing the compound-protein interaction machine learning model, an additional compound-protein machine learning representation comprising an additional plurality of binding scores between an additional training compound and the plurality of proteins; and
iteratively training a target machine learning model to improve accuracy of the target machine learning model by,
for a first training iteration:
inputting the compound-protein machine learning representation comprising the plurality of binding scores between the training compound and the plurality of proteins into the target machine learning model to generate a first predicted bioactivity result for the training compound; and
modifying parameters of the target machine learning model by comparing the first predicted bioactivity result to a ground truth bioactivity result corresponding to the training compound; and
for a second training iteration:
inputting the additional compound-protein machine learning representation comprising the additional plurality of binding scores between the additional training compound and the plurality of proteins into the target machine learning model to generate an additional predicted bioactivity result for the additional training compound; and
modifying the parameters of the target machine learning model by comparing the additional predicted bioactivity result to an additional ground truth bioactivity result corresponding to the additional training compound.
2 . The computer-implemented method of claim 1 , further comprising:
determining, utilizing a loss function, a measure of loss based on comparing the first predicted bioactivity result to the ground truth bioactivity result;
determining, utilizing the loss function, an additional measure of loss based on comparing the additional predicted bioactivity result to the additional ground truth bioactivity result; and
modifying the parameters of the target machine learning model to reduce the measure of loss and the additional measure of loss.
3 . The computer-implemented method of claim 1 , wherein the compound-protein interaction machine learning model comprises a compound-protein interaction neural network having parameters trained to generate binding scores indicating probabilities that compounds will bind to protein pockets of proteins.
4 . The computer-implemented method of claim 1 , further comprising:
determining that the compound-protein interaction machine learning model will perform below a threshold confidence accuracy based on applying a protein confidence filter to the compound-protein machine learning representation; and
in response to determining that the compound-protein interaction machine learning model will perform below a threshold confidence accuracy, generating a refined compound-protein machine learning representation by removing one or more binding scores from the compound-protein machine learning representation.
5 . The computer-implemented method of claim 4 , wherein the first predicted bioactivity result comprises a biological perturbation program prediction and training the target machine learning model comprises training the target machine learning model to generate biological perturbation program predictions utilizing a training dataset by, for a biological perturbation program corresponding to identifying compounds demonstrating a target biological activity:
generating, from the compound-protein machine learning representation utilizing the target machine learning model, the biological perturbation program prediction for the training compound; and
modifying the parameters of the target machine learning model by comparing the biological perturbation program prediction with a ground truth perturbation program result.
6 . The computer-implemented method of claim 5 , further comprising generating the training dataset by:
generating measures of similarity between a target gene of the biological perturbation program and datapoints of the training dataset, wherein the measures of similarity are based on at least one of phenomic data, transcriptomic data, metabolomic data, or proteomic data; and
generating the training dataset by filtering datapoints based on the measures of similarity between the datapoints and the target gene.
7 . The computer-implemented method of claim 6 , wherein the phenomic data comprises phenomic image embeddings and further comprising generating the training dataset by:
identifying phenomic digital images of cell perturbations;
generating, utilizing a machine learning model, phenomic image embeddings from the phenomic digital images; and
generating the training dataset by filtering datapoints based on pheno-similarity measures of the phenomic image embeddings relative to a target gene.
8 . The computer-implemented method of claim 5 , further comprising generating the training dataset by:
generating, utilizing a clustering model, clusters from a dataset utilizing chemical fingerprints of compounds; and
splitting the dataset into the training dataset and a testing data set based on the clusters.
9 . The computer-implemented method of claim 1 , wherein:
the first predicted bioactivity result comprises an absorption, distribution, metabolism, excretion, or toxicity (ADMET) prediction; and
training the target machine learning model comprises training the target machine learning model to generate ADMET predictions for the first training iteration by:
generating, from the compound-protein machine learning representation utilizing the target machine learning model, the ADMET prediction for the training compound; and
modifying the parameters of the target machine learning model by comparing the ADMET prediction to the ground truth bioactivity result, the ground truth bioactivity result comprising a measured ADMET result for the training compound.
10 . A system comprising:
at least one processor; and
at least one non-transitory computer-readable storage medium storing instructions that, when executed by the at least one processor, cause the system to:
identifying a plurality of compound-protein pairs comprising a training compound matched to a plurality of proteins;
generate, utilizing a compound-protein interaction machine learning model from the plurality of compound-protein pairs, a plurality of binding scores between the training compound and the plurality of proteins;
generate a compound-protein machine learning representation comprising the plurality of binding scores between the training compound and the plurality of proteins;
generate, utilizing the compound-protein interaction machine learning model, an additional compound-protein machine learning representation comprising an additional plurality of binding scores between an additional training compound and the plurality of proteins; and
iteratively train a target machine learning model to improve accuracy of the target machine learning model by,
for a first training iteration:
inputting the compound-protein machine learning representation comprising the plurality of binding scores between the training compound and the plurality of proteins into the target machine learning model to generate a first predicted bioactivity result for the training compound; and
modifying parameters of the target machine learning model by comparing the first predicted bioactivity result to a ground truth bioactivity result corresponding to the training compound; and
for a second training iteration:
inputting the additional compound-protein machine learning representation comprising the additional plurality of binding scores between the additional training compound and the plurality of proteins into the target machine learning model to generate an additional predicted bioactivity result for the additional training compound; and
modifying the parameters of the target machine learning model by comparing the additional predicted bioactivity result to an additional ground truth bioactivity result corresponding to the additional training compound.
11 . The system of claim 10 , further comprising instructions that, when executed by the at least one processor, cause the system to:
determine, utilizing a loss function, a measure of loss based on comparing the first predicted bioactivity result to the ground truth bioactivity result;
determine, utilizing the loss function, an additional measure of loss based on comparing the additional predicted bioactivity result to the additional ground truth bioactivity result; and
modify the parameters of the target machine learning model to reduce the measure of loss and the additional measure of loss.
12 . The system of claim 10 , further comprising instructions that, when executed by the at least one processor, cause the system to generate the compound-protein machine learning representation by:
determining machine learning protein confidence scores indicating a measure of confidence of the compound-protein interaction machine learning model in generating binding predictions for the plurality of proteins; and
filtering one or more features based on the machine learning protein confidence scores to generate the compound-protein machine learning representation.
13 . The system of claim 10 , further comprising instructions that, when executed by the at least one processor, cause the system to:
determine that the compound-protein interaction machine learning model will perform below a threshold confidence accuracy based on applying a protein confidence filter to the compound-protein machine learning representation; and
in response to determining that the compound-protein interaction machine learning model will perform below a threshold confidence accuracy, generate a refined compound-protein machine learning representation by removing one or more binding scores from the compound-protein machine learning representation.
14 . The system of claim 13 , wherein the first predicted bioactivity result comprises a biological perturbation program prediction and further comprising instructions that, when executed by the at least one processor, cause the system to train the target machine learning model to generate biological perturbation program predictions utilizing a training dataset by, for a biological perturbation program corresponding to identifying compounds demonstrating a target biological activity:
generating, from the compound-protein machine learning representation utilizing the target machine learning model, the biological perturbation program prediction for the training compound; and
modifying the parameters of the target machine learning model by comparing the biological perturbation program prediction with a ground truth biological perturbation program result.
15 . The system of claim 14 , further comprising instructions that, when executed by the at least one processor, cause the system to generate the training dataset by:
identifying phenomic digital images of cell perturbations;
generating, utilizing a machine learning model, phenomic image embeddings from the phenomic digital images; and
generating the training dataset by filtering datapoints based on a measure of similarity of the phenomic image embeddings relative to the target biological activity.
16 . The system of claim 10 :
wherein the first predicted bioactivity result comprises an absorption, distribution, metabolism, excretion, or toxicity (ADMET) prediction; and
further comprising instructions that, when executed by the at least one processor, cause the system to train the target machine learning model to generate ADMET predictions for the first training iteration by:
generating, from the compound-protein machine learning representation utilizing the target machine learning model, the ADMET prediction for the training compound; and
modifying the parameters of the target machine learning model by comparing the ADMET prediction to the ground truth bioactivity result, the ground truth bioactivity result comprising a measured ADMET result for the training compound.
17 . A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause a computing device to:
identify a plurality of compound-protein pairs comprising a training compound matched to a plurality of proteins;
generate, utilizing a compound-protein interaction machine learning model from the plurality of compound-protein pairs, a plurality of binding scores between the training compound and the plurality of proteins;
generate a compound-protein machine learning representation comprising the plurality of binding scores between the training compound and the plurality of proteins;
generate, utilizing the compound-protein interaction machine learning model, an additional compound-protein machine learning representation comprising an additional plurality of binding scores between an additional training compound and the plurality of proteins; and
iteratively train a target machine learning model to improve accuracy of the target machine learning model by,
for a first training iteration:
inputting the compound-protein machine learning representation comprising the plurality of binding scores between the training compound and the plurality of proteins into the target machine learning model to generate a first predicted bioactivity result for the training compound; and
modifying parameters of the target machine learning model by comparing the first predicted bioactivity result to a ground truth bioactivity result corresponding to the training compound; and
for a second training iteration:
inputting the additional compound-protein machine learning representation comprising the additional plurality of binding scores between the additional training compound and the plurality of proteins into the target machine learning model to generate an additional predicted bioactivity result for the additional training compound; and
modifying the parameters of the target machine learning model by comparing the additional predicted bioactivity result to an additional ground truth bioactivity result corresponding to the additional training compound.
18 . The non-transitory computer-readable medium of claim 17 , further comprising instructions that, when executed by the at least one processor, cause the computing device to generate the compound-protein machine learning representation by:
determining machine learning protein confidence scores indicating a measure of confidence of the compound-protein interaction machine learning model in generating binding predictions for the plurality of proteins; and
filtering one or more features based on the machine learning protein confidence scores to generate the compound-protein machine learning representation.
19 . The non-transitory computer-readable medium of claim 17 :
wherein the first predicted bioactivity result comprises an absorption, distribution, metabolism, excretion, or toxicity (ADMET) prediction; and
further comprising instructions that, when executed by the at least one processor, cause the computing device to train the target machine learning model to generate ADMET predictions for the first training iteration by:
generating, from the compound-protein machine learning representation utilizing the target machine learning model, the ADMET prediction for the training compound; and
modifying the parameters of the target machine learning model by comparing the ADMET prediction to the ground truth bioactivity result, the ground truth bioactivity result comprising a measured ADMET result for the training compound.
20 . The non-transitory computer-readable medium of claim 17 , wherein the first predicted bioactivity result comprises a biological perturbation program prediction and further comprising instructions that, when executed by the at least one processor, cause the computing device to train the target machine learning model to generate biological perturbation program predictions utilizing a training dataset by, for a biological perturbation program corresponding to identifying compounds demonstrating a target biological activity:
generating, from the compound-protein machine learning representation utilizing the target machine learning model, the biological perturbation program prediction for a training compound; and
modifying the parameters of the target machine learning model by comparing the biological perturbation program prediction with a ground truth biological perturbation program result.