AI-accelerated data sequence engineering system

1 . A system, comprising:

a data store configured to store a plurality of protein sequences that originates from one or more non-target data sources that correspond to one or more non-human species;

a computing device comprising memory and one or more processors, the memory storing instructions, wherein the instructions, when executed, cause the one or more processors to:

receive a protein sequence that has a majority of regions originating from the one or more non-target data sources that correspond to the one or more non-human species;

replace windows of the protein sequence with humanized regions to generate a humanized intermediate candidate sequence;

identify one or more contact residue regions in the humanized intermediate candidate sequence, wherein the one or more contact residue regions are verifiable from laboratory tests to be regions that interact with a drug target molecule;

input the humanized intermediate candidate sequence and the one or more contact residue regions into a deep neural network to generate a set of one or more improved humanized candidate sequences that have higher affinity than the protein sequence to the drug target molecule, wherein the one or more contact residue regions are inputted as preservation constraints; and

output the set of one or more improved target compatible humanized candidate sequences to be stored in the data store.

2 . The system of claim 1 , wherein the instruction to replace the windows of the protein sequence with the humanized regions to generate the humanized intermediate candidate sequence comprises using a lookup-based approach, the lookup-based approach comprises:

retrieving a database of known human protein sequences;

identifying sequence windows in the protein sequence that are non-human;

mapping a non-human sequence window to a matching human sequence window based on a similarity threshold; and

substituting the non-human sequence window with a corresponding human sequence window in response to at least meeting the similarity threshold.

3 . The system of claim 1 , wherein the instruction to replace the windows of protein sequence with the humanized intermediate candidate sequence comprises instructions to:

input the protein sequence into a machine learning model that is trained on a dataset of human and non-human protein sequences;

generate, using the machine learning model, one or more candidate humanized sequence windows that are candidates to replace the windows of the protein sequence;

score the one or more candidate humanized sequence windows based on sequence similarity; and

substitute one or more windows of the protein sequence with one or more candidate humanized sequence windows based on the scoring.

4 . The system of claim 1 , wherein the instruction to replace the windows of protein sequence with the humanized regions comprises instructions to:

retrieve a reference set of human protein sequences;

compute pairwise similarity scores between non-human sequence windows and the reference set;

select a sequence similarity threshold for viable adaptation humanization candidates; and

select humanized sequence windows that meet the sequence similarity threshold.

5 . The system of claim 1 , wherein the instruction to replace the windows of protein sequence with the humanized regions comprises instructions to:

retrieve humanized protein sequences from a curated protein entity database;

identify sequence motifs pattern to a target protein function;

match a non-human sequence region with the sequence motifs based on a similarity in the target protein function; and

substitute the non-human sequence region with one of the sequence motifs based on the similarity in the target protein function.

6 . The system of claim 1 , wherein the instruction to identify one or more contact residue regions in the humanized intermediate candidate sequence comprises instructions to:

determine a protein structure of the humanized intermediate candidate sequence;

compute intermolecular distances between amino acid residues of the humanized intermediate candidate sequence and a drug target molecule;

determine a distance threshold for contact residue site determination; and

label amino acid residues that fall within the distance threshold as the contact residue regions.

7 . The system of claim 6 , wherein the instruction to determine the distance threshold comprises instructions to:

analyze structural interaction data from known antibody-antigen complexes;

establish an empirical distance range that correlates with binding interactions in the structural interaction data; and

determine the threshold based on the empirical distance range.

8 . The system of claim 1 , wherein the instruction to input the humanized intermediate candidate sequence and the one or more contact residue regions into the deep neural network to generate the set of one or more improved humanized intermediate candidate sequences comprises instructions to:

define contact residue regions that remain unchanged during affinity improvement;

restrict sequence modifications to non-interaction residues; and

generate the set of one or more improved humanized candidate sequences.

9 . The system of claim 1 , wherein the instruction to input the humanized intermediate candidate sequence and the one or more contact residue regions into a deep neural network to generate a set of one or more improved humanized candidate sequences comprises instructions to:

establish the preservation constraints as including a feature of the contact residue regions;

allow limited modification of the contact residue regions within a predefined extent;

iteratively adjust candidate sequences based on both affinity and humanization scores.

10 . The system of claim 1 , wherein the instruction to replace the windows of protein sequence with the humanized regions comprises instructions to:

retrieve structural data from crystallography, cryo-electron microscopy, or computational structure prediction;

identify conserved structural features that are maintained during humanization;

validate humanized candidate sequences through structural stability simulations; and

select candidates that maintain folding and binding interactions of the protein sequence.

11 . The system of claim 1 , wherein the instruction to input the humanized intermediate candidate sequence and the one or more contact residue regions into the deep neural network to generate the set of one or more improved humanized candidate sequences comprises instructions to:

use a transformer-based model trained on antibody sequence databases to predict affinity-enhancing mutations;

provide the one or more contact residue regions to the transformer-based model as the preservation constraints; and

generate sequence modifications to the humanized intermediate candidate sequence using the transformer-based model.

12 . The system of claim 1 , wherein the instruction to input the humanized intermediate candidate sequence and the one or more contact residue regions into the deep neural network to generate the set of one or more improved humanized candidate sequences comprises instructions to:

use a diffusion model to generate incremental sequence modifications; and

generate variants by iteratively refining the humanized intermediate candidate sequence.

13 . The system of claim 1 , wherein the instruction to input the humanized intermediate candidate sequence and the one or more contact residue regions into the deep neural network to generate the set of one or more improved humanized candidate sequences comprises instructions to:

employ a reinforcement learning technique to reward affinity while maintaining structural and functional constraints;

implement a reward-based learning that prioritizes rewarding the affinity; and

generate an improved humanized candidate sequence.

14 . The system of claim 1 , wherein the instructions, when executed, further cause the one or more processors to:

collect experimental validation data on the set of one or more improved humanized candidate sequences;

feed back the experimental validation data as model training data; and

retrain the deep neural network using the model training data.

15 . The system of claim 1 , wherein the majority of regions of the protein sequence originates from a rodent-derived antibody.

16 . The system of claim 1 , wherein the majority of regions of the protein sequence originates from a camelid nanobody.

17 . The system of claim 1 , wherein the protein sequence is an approved therapeutic entity.

18 . The system of claim 1 , wherein the instructions, when executed, further cause the one or more processors to distill the deep neural network to generate a second deep neural network using outputs of the deep neural network as training data of the second deep neural network.

19 . A non-transitory computer-readable medium configured to store code comprising instructions, wherein the instructions, when executed by one or more processors, cause the one or more processors to:

receive a protein sequence that has a majority of regions originating from one or more non-target data sources that correspond to one or more non-human species;

replace windows of the protein sequence with humanized regions to generate a humanized intermediate candidate sequence;

identify one or more contact residue regions in the humanized intermediate candidate sequence, wherein the one or more contact residue regions are verifiable from laboratory tests to be regions that interact with a drug target molecule;

input the humanized intermediate candidate sequence and the one or more contact residue regions into a deep neural network to generate a set of one or more improved humanized candidate sequences that have higher affinity than the protein sequence to the drug target molecule, wherein the one or more contact residue regions are inputted as preservation constraints; and

output the set of one or more improved humanized candidate sequences to be stored in the data store.

20 . A computer-implemented method, comprising:

receiving a protein sequence that has a majority of regions originating from one or more non-target data sources that correspond to one or more non-human species;

replacing windows of the protein sequence with humanized regions to generate a humanized intermediate candidate sequence;

identifying one or more contact residue regions in the humanized intermediate candidate sequence, wherein the one or more contact residue regions are verifiable from laboratory tests to be regions that interact with a drug target molecule;

inputting the humanized intermediate candidate sequence and the one or more contact residue regions into a deep neural network to generate a set of one or more improved humanized candidate sequences that have higher affinity than the protein sequence to the drug target molecule, wherein the one or more contact residue regions are inputted as preservation constraints; and

outputting the set of one or more improved humanized candidate sequences to be stored in the data store.