Utilizing machine learning models to synthesize perturbation data to generate perturbation heatmap graphical user interfaces
The present disclosure relates to systems, non-transitory computer-readable media, and methods for embedding perturbation data via a machine learning model and filtering, aligning, and aggregating the embeddings to generate a genome-wide perturbation database for real-time generation of perturbation heatmaps. In particular, in one or more embodiments, the disclosed systems can receive a plurality of perturbation images portraying cells from a plurality of wells corresponding to a plurality of cell perturbations. Further, the systems can generate, utilizing a machine learning model, a plurality of well-level image embeddings from the plurality of perturbation images. Moreover, the systems can align, utilizing an alignment model, the plurality of well-level image embeddings to generate aligned well-level image embeddings. Additionally, the systems can aggregate, according to perturbations of one or more perturbation experiments, the well-level image embeddings to generate perturbation-level image embeddings. Furthermore, the systems can generate perturbation comparisons utilizing the perturbation-level image embeddings.
1 . A computer-implemented method comprising:
receiving a plurality of perturbation images portraying cells from a plurality of wells corresponding to a plurality of cell perturbations;
generating, utilizing a machine learning model, a plurality of well-level image embeddings from the plurality of perturbation images;
aligning, utilizing an alignment model, the plurality of well-level image embeddings to generate aligned well-level image embeddings;
aggregating, according to perturbations of one or more perturbation experiments, the aligned well-level image embeddings to generate perturbation-level image embeddings comprising:
a first perturbation-level image embedding aggregated from a first set of well-level image embeddings having a first shared cell perturbation, and
a second perturbation-level image embedding aggregated from a second set of well-level image embeddings having a second shared cell perturbation different from the first shared cell perturbation; and
generating perturbation comparisons utilizing the perturbation-level image embeddings.
2 . The computer-implemented method of claim 1 , wherein receiving the plurality of perturbation images comprises:
receiving a first perturbation image portraying a first cell resulting from at least one of: a first gene knockout perturbation of a first gene or a first compound perturbation; and
receiving a second perturbation image portraying a second cell resulting from at least one of: a second gene knockout perturbation of a second gene or a second compound perturbation.
3 . The computer-implemented method of claim 2 , wherein generating, utilizing the machine learning model, the plurality of well-level image embeddings comprises:
generating, utilizing a convolutional neural network, a first well-level embedding for the first perturbation image; and
generating, utilizing the convolutional neural network, a second well-level embedding for the second perturbation image.
4 . The computer-implemented method of claim 1 , wherein the plurality of perturbation images comprise a plurality of patch images portraying portions of wells from the plurality of wells and wherein generating the plurality of well-level image embeddings comprises:
generating, utilizing the machine learning model, a plurality of patch-level embeddings from the plurality of perturbation images; and
aggregating the plurality of patch-level embeddings according to the plurality of wells to generate the plurality of well-level image embeddings.
5 . The computer-implemented method of claim 1 , further comprising:
receiving an additional plurality of perturbation images portraying additional cells;
generating, utilizing the machine learning model, an additional plurality of well-level image embeddings; and
generating modified perturbation-level image embeddings from the additional plurality of well-level image embeddings and the plurality of well-level image embeddings.
6 . The computer-implemented method of claim 1 , further comprising filtering the plurality of well-level image embeddings according to one or more quality criterion.
7 . The computer-implemented method of claim 1 , wherein aligning, utilizing the alignment model, the plurality of well-level image embeddings to generate the aligned well-level image embeddings comprises aligning a set of well-level image embeddings from a plurality of different perturbation experiments having a shared perturbation according to a statistical aligning model.
8 . The computer-implemented method of claim 1 , further comprising:
generating, utilizing a proximity bias model, proximity bias corrected perturbation-level image embeddings; and
generating the perturbation comparisons from the proximity bias corrected perturbation-level image embeddings.
9 . The computer-implemented method of claim 1 , wherein generating the perturbation comparisons utilizing the perturbation-level image embeddings comprises:
determining similarity measures between the perturbation-level image embeddings;
generating a perturbation heatmap comprising the similarity measures for the plurality of cell perturbations; and
providing the perturbation heatmap for display via a client device.
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:
receive a plurality of perturbation images portraying cells from a plurality of wells corresponding to a plurality of cell perturbations;
generate, utilizing a machine learning model, a plurality of well-level image embeddings from the plurality of perturbation images;
align, utilizing an alignment model, the plurality of well-level image embeddings to generate aligned well-level image embeddings;
aggregate, according to perturbations of one or more perturbation experiments, the aligned well-level image embeddings to generate perturbation-level image embeddings comprising:
a first perturbation-level image embedding aggregated from a first set of well-level image embeddings having a first shared cell perturbation, and
a second perturbation-level image embedding aggregated from a second set of well-level image embeddings having a second shared cell perturbation different from the first shared cell perturbation; and
generate perturbation comparisons utilizing the perturbation-level image embeddings.
11 . The system of claim 10 , further comprising instructions that, when executed by the at least one processor, cause the system to:
determine statistical significance values for the perturbation-level image embeddings; and
filter one or more of the perturbation-level image embeddings having a statistical significance value below a statistical significance threshold.
12 . The system of claim 11 , wherein the instructions cause the system to generate, utilizing the machine learning model, the plurality of well-level image embeddings by:
generating, utilizing a convolutional neural network, a first well-level embedding for a first perturbation image; and
generating, utilizing the convolutional neural network, a second well-level embedding for a second perturbation image.
13 . The system of claim 10 , wherein the plurality of perturbation images comprise a plurality of patch images portraying portions of wells from the plurality of wells and wherein the instructions cause the system to generate the plurality of well-level image embeddings by:
generating, utilizing the machine learning model, a plurality of patch-level embeddings from the plurality of perturbation images; and
aggregating the plurality of patch-level embeddings according to the plurality of wells to generate the plurality of well-level image embeddings.
14 . The system of claim 10 , further comprising instructions that, when executed by the at least one processor, cause the system to:
receive an additional plurality of perturbation images portraying additional cells;
generate, utilizing the machine learning model, an additional plurality of well-level image embeddings; and
generate modified perturbation-level image embeddings from the additional plurality of well-level image embeddings and the plurality of well-level image embeddings.
15 . The system of claim 10 , wherein the instructions cause the system to align, utilizing the alignment model, the plurality of well-level image embeddings to generate the aligned well-level image embeddings comprises aligning a set of well-level image embeddings from a plurality of different perturbation experiments having a shared perturbation according to a statistical aligning model.
16 . A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause a computing device to:
receive a plurality of perturbation images portraying cells from a plurality of wells corresponding to a plurality of cell perturbations;
generate, utilizing a machine learning model, a plurality of well-level image embeddings from the plurality of perturbation images;
align, utilizing an alignment model, the plurality of well-level image embeddings to generate aligned well-level image embeddings;
aggregate, according to perturbations of one or more perturbation experiments, the aligned well-level image embeddings to generate perturbation-level image embeddings comprising:
a first perturbation-level image embedding aggregated from a first set of well-level image embeddings having a first shared cell perturbation, and
a second perturbation-level image embedding aggregated from a second set of well-level image embeddings having a second shared cell perturbation different from the first shared cell perturbation; and
generate perturbation comparisons utilizing the perturbation-level image embeddings.
17 . The non-transitory computer-readable medium of claim 16 , wherein the instructions cause the computing device to receive the plurality of perturbation images by:
receiving a first perturbation image portraying a first cell resulting from at least one of: a first gene knockout perturbation of a first gene or a first compound perturbation; and
receiving a second perturbation image portraying a second cell resulting from at least one of: a second gene knockout perturbation of a second gene or a second compound perturbation.
18 . The non-transitory computer-readable medium of claim 17 , wherein the instructions cause the computing device to generate, utilizing the machine learning model, the plurality of well-level image embeddings by:
generating, utilizing a convolutional neural network, a first well-level embedding for the first perturbation image; and
generating, utilizing the convolutional neural network, a second well-level embedding for the second perturbation image.
19 . The non-transitory computer-readable medium of claim 16 , wherein the plurality of perturbation images comprise a plurality of patch images portraying portions of wells from the plurality of wells and wherein the instructions cause the computing device to generate the plurality of well-level image embeddings by:
generating, utilizing the machine learning model, a plurality of patch-level embeddings from the plurality of perturbation images; and
aggregating the plurality of patch-level embeddings according to the plurality of wells to generate the plurality of well-level image embeddings.
20 . The non-transitory computer-readable medium of claim 16 , further comprising instructions that, when executed by the at least one processor, cause the computing device to:
receive an additional plurality of perturbation images portraying additional cells;
generate, utilizing the machine learning model, an additional plurality of well-level image embeddings; and
generate modified perturbation-level image embeddings from the additional plurality of well-level image embeddings and the plurality of well-level image embeddings.