IP Library Granted Patent US 12665048
Granted Patent B2
US 12665048 · App. 19/000,940 · Granted Jun 23, 2026

Synthon embeddings for modeling DNA-encoded libraries

Inventors: Benson Chen (Palo Alto, CA); Mohammad Muneeb Sultan (San Francisco, CA); Theofanis Karaletsos (San Francisco, CA)
Assignee: Insitro, Inc.
G16B15/30G06N20/00G16C20/10
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12665048
App. No.
19/000,940
Granted
Jun 23, 2026
Kind
B2
Abstract

Embodiments of the disclosure involve modeling DEL data using factorized molecular representations (e.g., hierarchical mono-synthon and di-synthon building blocks), which capitalizes on the inherent hierarchical structure of these molecules. Using the factorized molecular representations, machine learning models are trained to learn latent binding affinity of compounds for targets and one or more covariates (e.g., load/replicate noise). This leads to improved predictions by the machine learning models in the form of higher enrichment scores, which are well-correlated with compound-target binding affinity.

Claims (36)

1 . A method for performing molecular screening of a compound from a virtual library of compounds for binding to a protein target, the method comprising:

obtaining a plurality of synthons forming the compound;

transforming the plurality of synthons into a plurality of synthon representations comprising one or more monosynthon representations, one or more disynthon representations, and one or more trisynthon representations, wherein the transforming comprises:

transforming the plurality of synthons to generate one or more monosynthon representations;

transforming the one or more monosynthon representations to generate one or more disynthon representations; and

transforming the one or more disynthon representations to generate one or more trisynthon representations;

combining the plurality of synthon representations into a molecular embedding by implementing a multi-head attention mechanism across the plurality of synthon representations, wherein the molecular embedding comprises learned attention weights for the plurality of synthon representations; and

using a multilayer perceptron (MLP) neural network, analyzing the molecular embedding to generate at least a target enrichment prediction representing a measure of binding between the compound and the target.

2 . The method of claim 1 , further comprising determining a binding affinity value between the compound and the target using the target enrichment prediction.

3 . The method of claim 1 , wherein implementing the multi-head attention mechanism comprises using one or more learned attention weights to rank the plurality of synthons for their ability to bind to the target.

4 . The method of claim 1 , further comprising performing probabilistic modeling using at least the target enrichment prediction by applying a probability density function that models experimental target counts.

5 . The method of claim 4 , wherein the probability density function is represented by any one of a Poisson distribution, Binomial distribution, Gamma distribution, Binomial-Poisson distribution, or negative binomial distribution.

6 . A non-transitory computer readable medium for developing a machine learning model for use in ML enabled molecular screening of a compound from a virtual library of compounds for binding to a protein target, the non-transitory computer readable medium comprising instructions that, when executed by a processor, cause the processor to perform steps comprising:

obtain a plurality of synthons forming the compound;

transform the plurality of synthons into a plurality of synthon representations comprising one or more monosynthon representations, one or more disynthon representations, and one or more trisynthon representations, wherein the transform step comprises:

transforming the plurality of synthons to generate one or more monosynthon representations;

transforming the one or more monosynthon representations to generate one or more disynthon representations; and

transforming the one or more disynthon representations to generate one or more trisynthon representations;

combine the plurality of synthon representations into a molecular embedding by implementing a multi-head attention mechanism across the plurality of synthon representations, wherein the molecular embedding comprises learned attention weights for the plurality of synthon representations; and

use a multilayer perceptron (MLP) neural network, analyzing the molecular embedding to generate at least a target enrichment prediction representing a measure of binding between the compound and the target.

7 . The non-transitory computer readable medium of claim 6 , wherein instructions for training the MLP further comprises instructions that, when executed by the processor, cause the processor to perform steps comprising: perform probabilistic modeling using at least the target enrichment prediction by applying a probability density function that models experimental target counts.

8 . The non-transitory computer readable medium of claim 6 , further comprising instructions that, when executed by the processor, cause the processor to determine a binding affinity value between the compound and the target using the target enrichment prediction.

9 . The non-transitory computer readable medium of claim 6 , wherein the instructions that cause the processor to implement the multi-head attention mechanism further comprises instructions that, when executed by the processor, cause the processor to use one or more learned attention weights to rank the plurality of synthons for their ability to bind to the target.

10 . The non-transitory computer readable medium of claim 6 , further comprising instructions that, when executed by a processor, cause the processor to: perform probabilistic modeling using at least the target enrichment prediction by applying a probability density function that models experimental target counts.

11 . The non-transitory computer readable medium of claim 10 , wherein the probability density function is represented by any one of a Poisson distribution, Binomial distribution, Gamma distribution, Binomial-Poisson distribution, or negative binomial distribution.

12 . A method for performing molecular screening of a compound from a virtual library of compounds for binding to a protein target, the method comprising:

obtaining a plurality of synthons forming the compound;

transforming the plurality of synthons into a plurality of synthon representations comprising one or more monosynthon representations, one or more disynthon representations, and one or more trisynthon representations, wherein the transforming comprises:

transforming the plurality of synthons using a first learned representation model to generate one or more monosynthon representations;

transforming the one or more monosynthon representations using a second learned representation model to generate one or more disynthon representations; and

transforming the one or more disynthon representations using a third learned representation model to generate one or more trisynthon representations;

combining the plurality of synthon representations into a molecular embedding; and

using a multilayer perceptron (MLP) neural network, analyzing the molecular embedding to generate at least a target enrichment prediction representing a measure of binding between the compound and the target.

13 . The method of claim 12 , further comprising determining a binding affinity value between the compound and the target using the target enrichment prediction.

14 . The method of claim 12 , further comprising performing probabilistic modeling using at least the target enrichment prediction by applying a probability density function that models experimental target counts.

15 . The method of claim 14 , wherein the probability density function that models the experimental control counts is represented by any one of a Poisson distribution, Binomial distribution, Gamma distribution, Binomial-Poisson distribution, or negative binomial distribution.