DETAILED ACTION
Notice of Pre-AIA or AIA Status
1. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim status
2. Claims 9 and 21 are cancelled.
Claims 1-8, 10-20 and 22 are currently pending and under exam herein.
Claims 1-8, 10-20 and 22 are rejected.
Priority
3. Domestic priority to a national stage application under 35 U.S.C. § 371 of International Application No. PCT/CN2021/142904, filed December 30, 2021 is acknowledged. In this action, all claims are examined as though they had an effective filing date of December 30, 2021. In future actions, the effective filing date of one or more claims may change, due to amendments to the claims, or further analysis of the disclosure(s) of the priority application(s).
Information Disclosure Statement
4. No information disclosure statement has been filed herein.
Drawings
5. The drawings submitted on 27 September, 2022 have been accepted by the examiner and are under consideration herein.
Claim Interpretation
6. Claims 14-18 merely show how similarities and distances are expressed but they are not interpreted to be part of the method because there is no step to calculate similarities.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
7. Claims 7 and 10 are rejected under 3 U.S.C. 112(b) or pre-AIA 35 U.S.C. 112, second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or joint inventor regards as the invention.
Claim 7 recites “the method of claim 5, wherein sampling L number of negative samples from the respective intermediate step is performed based on probability” lacks antecedent basis. The claim does not previously introduce a sampling step, rendering the scope of the claim uncertain. It is unclear if the method includes a sampling step or the clause merely describes a property of the method without requiring the step. For the purpose of examination and under broadest reasonable interpretation, the sampling step will not be considered a step of the invention.
Claim 10 recites “the use of the positive sample set and the negative sample set generated by the method of claim 1”, but no steps in the process of using the samples are disclosed. Claiming a process without setting forth any steps involved in the process is generally considered indefinite (see MPEP 2173.05(q).
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
8. Claims 10 and 22 are rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter.
Step 1
In accordance with MPEP § 2106, claims are analyzed to determine if they recite statutory subject matter. In the instant application, the following claims do not recite statutory subject matter:
Claim 10 recites: a method of predicting macromolecule-macromolecule interaction using the positive sample set and the negative sample set generated by the method of claim 1
Claim 22 recites: a neural network model for predicting macromolecule-macromolecule interaction, trained by the method of claim 11.
The claimed subject matter of claims 10 and 22 do not fall within at least one of the four categories of patent eligible subject matter because they are not directed to a method, a manufacture, a composition of matter or a process.
Claim 10 is directed to using the samples for predicting interactions between macromolecules, but there is no method listed for predicting interactions. Claim 22 is directed to a neural network model, which under the broadest reasonable interpretation, encompasses data per se.
Therefore claims 10 and 22 are rejected for claiming non-statutory subject matter (STEP 1: NO).
9. Claims 1-8 and 11-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 2A, Prong 1
In accordance with MPEP § 2106, claims found to recite statutory subject matter (Step 1: YES) are then analyzed to determine if the claims recite any concepts that equate to an abstract idea, law of nature or natural phenomenon (Step 2A, Prong 1). In the instant application, the claims recite the following limitations that equate to an abstract idea:
Claim 1 recites: generating a first similarity map of the macromolecules of the first type
Claim 1 recites: generating a second similarity map of the macromolecules of the second type
Claim 1 recites: generating vectorized representations of nodes in the first similarity map and vectorized representations of nodes in the second similarity map
Claim 1 recites: generating the negative sample set using the vectorized representations of nodes in the first similarity map and the vectorized representations of nodes in the second similarity map
Claim 2 recites: the method of claim 1, wherein the first similarity map or the second similarity map comprises nodes and edges connecting adjacent nodes, wherein a respective node represents a respective macromolecule, a respective edge represents a respective distance between a respective pair of the macromolecules, and a respective weight of the respective edge represents a respective similarity between the respective pair of the macromolecules
Claim 3 recite: the method of claim 1, further comprising generating a plurality of intermediate sets
Claim 3 recites: wherein the positive sample set is represented by {(m1i, m2i), i=1,...,K}, wherein m1i stands for an i-th macromolecule of the first type and m2i stands for an i-th macromolecule of the second type
Claim 3 recite: wherein generating the respective intermediate set of the plurality of intermediate sets comprises: sorting m1j (j=1,...,K, and j ≠ i) based on similarities between m1i and m1j to obtain a subset of m1j(j=K, and j ≠ i)
Claim 3 recites: determining a probability of interaction between m2i and each sample in the subset; and generating the respective intermediate set based on the probability of interaction
Claim 4 recites: the method of claim 3, further comprising calculating similarities between m1i and m1j (j=1,...,K, and j ≠ i) by:
s
d
r
j
,
d
r
i
=
<
d
r
j
,
d
r
i
>
<
d
r
j
,
d
r
j
>
X
<
d
r
i
,
d
r
i
>
wherein dr stands for vectorized representations of nodes in the subset; and dri stands for a vectorized representation of mli.
Claim 5 recites: the method of claim 3, wherein the probability of interaction between m2i and each sample in the subset is determined by:
P
1
d
r
j
,
d
p
i
=
1
1
+
exp
-
ϴ
d
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
, wherein drj stands for vectorized representations of nodes in the subset; dpl stands for vectorized representations of m2i; P(1|drj, dpi) stands for a probability of interaction between m2i and each sample in the subset; [,] stands for stitching between elements;
⦻
stands for product of two vectors; and
ϴ
stands for a parameter that is tunable.
Claim 6 recites: the method of claim 4, further comprising placing (m1j, 1-P(1|drj,dpi,)) into the respective intermediate set when P(1|drj, dpi) is less than a threshold value
Claim 7 recites: the method of claim 5, wherein sampling L number of negative samples from the respective intermediate set is performed based on probability pk,k=1,…,|T|}; wherein
p
k
=
1
-
P
1
d
r
j
,
d
p
i
∑
I
=
1
T
1
-
P
1
d
r
j
,
d
p
i
, |T| stands for a number of elements in the respective intermediate set; and 1- P(1 | drj, dpi)) stands for a k-th element in the intermediate set
Claim 8 recites: the method of claim 3, wherein generating the negative sample set comprises: sampling L number of negative samples from a respective intermediate set of the plurality of intermediate sets; wherein the negative sample set comprises negative samples sampled from the plurality of intermediate sets; and L is an integer equal to or greater than 1
Claim 11 recite: generating a first similarity map of macromolecules of a first type
Claim 11 recites: generating a second similarity map of macromolecules of a second type
Claim 11 recites: generating vectorized representations of nodes in the first similarity map and vectorized representations of nodes in the second similarity map
Claim 11 recites: determining a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map
Claim 11 recites: training the model at least partially based on the probability of interaction
Claim 12 recites: the method of claim 11, wherein the probability of interaction is determined by:
P
1
d
m
1
i
,
d
m
2
j
=
1
1
+
exp
-
ϴ
d
m
1
i
,
d
m
2
j
,
d
m
1
i
-
d
m
2
j
,
d
m
1
i
⦻
d
m
2
j
, wherein dm1 stands for vectorized representations of nodes in the first similarity map; dm2 stands for vectorized representations of nodes in the second similarity map; p (1 |dm1i, dm2i) stands for a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map; [,] stands for stitching between elements; and
⦻
stands for product of two vectors; and
ϴ
stands for a parameter that is tunable
Claim 13 recites: the method of claim 11, wherein the first similarity map or the second similarity map comprises nodes and edges connecting adjacent nodes, wherein a respective node represents a respective macromolecule, a respective edge represents a respective distance between a respective pair of the macromolecules, and a respective weight of the respective edge represents a respective similarity between the respective pair of the macromolecules
Claim 14 recites: the method of claim 11, wherein a respective similarity between a respective pair of the macromolecules of the first type is expressed as: sim1 (m1-1,m1-2)= 1 - d1(m1-1,m1-2); wherein (m1-1 , m1-2) stands for the respective pair of the macromolecules of the first type, sim1 stands for the respective similarity between the respective pair of the macromolecules of the first type, and dl stands for a distance between the respective pair of the macromolecules of the first type
Claim 15 recites: the method of claim 14, wherein dl is expressed as:
d
1
m
1
-
1
,
m
1
-
2
=
l
e
v
m
1
-
1
,
m
1
-
2
max
l
e
n
m
1
-
1
,
m
1
-
2
,
wherein lev(m1-1 , m1-2) stands for an edit distance between the respective pair of the macromolecules of the first type, len(m1-1) stands for a length of a first macromolecule of the first type in the respective pair, and len(m1-2) stands for a length of a second macromolecule of the first type in the respective pair
Claim 16 recites: the method of claim 11, wherein a respective similarity between a respective pair of the macromolecules of the second type is expressed as: sim2 (m2-1,m2-2)= 1 - d2(m2-1,m2-2); wherein (m2-1,m2-2)) stands for the respective pair of the macromolecules of the second type, sim2 stands for the respective similarity between the respective pair of the macromolecules of the second type, and d2 stands for a distance between the respective pair of the macromolecules of the second type
Claim 17 recites: the method of claim 16, wherein d2 is expressed as:
d
2
m
2
-
1
,
m
2
-
2
=
l
e
v
m
2
-
1
,
m
2
-
2
max
l
e
n
m
2
-
1
,
m
2
-
2
; wherein, lev(m2-1,m2-2) stands for an edit distance between the respective pair of the macromolecules of the second type, len(m2-1) stands for a length of a first macromolecule of the second type in the respective pair, and len(m2-2) stands for a length of a second macromolecule of the second type in the respective pair
Claim 20 recites: wherein training the model comprises minimizing a loss function:
L
=
-
∑
i
=
1
K
l
o
g
p
1
|
d
m
1
i
,
d
m
2
i
;
wherein dm1i stands for vectorized representations of nodes in the first similarity map; dm2i stands for vectorized representations of nodes in the second similarity map; p (1 |dm1i, dm2i) stands for a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map.
The limitations regarding generating vectorized representations of nodes in claims 1 and 11, calculating similarities between macromolecules using vectorized representations in claim 4, determining a probability based on vectorized representations of nodes in claim 5, determining a probability of interaction between vectorized representations of nodes in claim 11, training a model based on probability of interactions in claim 11, determining a probability of interaction using an equation with vectorized representations of nodes in claim 12 and minimizing a loss function in claim 20 are equations and/or verbal equivalents that describe mathematical calculations performed on vectorized representations of data. Therefore these limitations fall under the “Mathematical concepts” groupings of abstract ideas.
The limitation directed to determining a probability of interaction between two macromolecules in claim 3 describes a mathematical calculation that is so simple that it could be performed in the human mind or with pen and paper. Therefore, these limitations fall under the "Mathematical concepts" and "Mental processes" groupings of abstract ideas.
The remaining limitations for placing nodes/macromolecules into the respective intermediate set in claim 1, sampling a number of negative samples in claims 1 and 11, generating a similarity map in claims 1 and 11, sorting samples/macromolecules based on similarity to other proteins in claim 3, generating an intermediate set based on probability scores in claim 3, placing a sample in an intermediate set based on a threshold in claim 6, and sampling negative samples in claim 8 are generically recited data analysis steps that can be practically performed in the human mind because the human mind is capable of identifying relevant information, comparing values, and determining information from other values.
Regarding the limitation of claim 7 that further limits ‘sampling L number of negative samples’, although it recites a mathematical equation, this limitation is not considered to be a step of the invention because there is no sampling step recited in the preceding claim family. Claims 2 and 13 further limit the edges of the similarity maps generated in claims 1 and 11 but do not change the position of these limitations as a mental process. The limitations of claims 14-17 further limit how the similarities used in the ‘generating similarity maps’ in claim 11 are expressed, but how the similarities are expressed does not change the position of the limitation as a mental process.
As such, claims 1-8 and 11-20 recite an abstract idea (Step 2A, Prong 1: YES).
Step 2A, Prong 2
Claims found to recite a judicial exception under Step 2A, Prong 1 are then further analyzed to determine if the claims as a whole integrate the recited judicial exception into a practical application or not (Step 2A, Prong 2). This judicial exception is not integrated into a practical application because the claims do not recite an additional element that reflects an improvement to technology or applies or uses the recited judicial exception in some other meaningful way. Rather, the instant claims recite additional elements that amount to mere instructions to implement the abstract idea in a generic computing environment or insignificant extra-solution activity. Specifically, the claims recite the following additional elements:
Claim 1 recites: receiving a positive sample set comprising pairs of macromolecules of a first type and macromolecules of a second type having macromolecule-macromolecule interaction
Claim 11 recites: a method of training a model for generating a negative sample set for predicting macromolecule-macromolecule interaction, comprising: receiving a positive sample set comprising pairs of macromolecules of a first type and macromolecules of a second type having macromolecule-macromolecule interaction
Claim 18 recites: the method of claim 11, wherein the first similarity map includes N1 number of nodes, {ei, i=1,...,N1}; and M1 number of edges, {rj,j=1,...,M}; a respective vectorized representation of a respective node in the first similarity map is expressed as:
h
t
1
+
1
e
i
=
σ
(
W
p
X
h
t
1
e
i
+
∑
e
k
⋲
N
e
i
W
p
h
X
h
t
1
(
e
k
)
)
;
wherein ei stands for a respective node in the first similarity map; ht1(ei) stands for a respective vectorized representation of the respective node ei prior to a t1-th step reiteration; ht1+1(ei) stands for an updated respective vectorized representation of the respective node ei subsequent to the t1-th step reiteration;
σ
stands for a leaky relu activation function; N(ei) stands for a set of nodes neighboring the respective node ei; and Wp, Wph stand for parameters of a graph neural network for generating the vectorized representation
Claim 19 recites: the method of claim 11, wherein the second similarity map includes N2 number of nodes, {e'i, i=1,...,N2}; and M2 number of edges, {r'j,j=1,...,M2}; a respective vectorized representation of a respective node in the second similarity map is expressed as:
h
t
2
+
1
e
'
i
=
σ
∑
e
'
k
⋲
N
e
'
i
U
e
i
α
i
,
k
t
2
+
1
W
p
t
+
1
X
h
t
2
e
'
k
;
α
i
,
k
t
2
+
1
=
s
o
f
t
m
a
x
h
t
2
e
i
'
,
h
t
2
e
'
k
wherein e'i stands for a respective node in the second similarity map; (e'i) stands for a respective vectorized representation of the respective node e'i prior to a t2-th step reiteration; htz+1 (e'i) stands for an updated respective vectorized representation of the respective node e'i subsequent to the t2-th step reiteration; a stands for a leaky relu activation function; ; N(e'i) stands for a set of nodes neighboring the respective node e'i;
h
t
2
e
i
'
,
h
t
2
e
'
k
stands for an inner product of ht2(e’j) and ht2 (e’k); Wpt2+1 stands for a parameter of a graph neural network for generating the vectorized representation; and
α
i
,
k
t
2
+
1
stands for attention weights representing a link strength between node e’i and node e’k.
Claim 20 recites: the method of claim 11, wherein the positive sample set is represented by {(m1i, m2i), i=1,...,K}, wherein m1i stands for an i-th macromolecule of the first type and m2i stands for an i-th macromolecule of the second type
The limitations for receiving a positive sample in claims 1 and 11 merely serve to gather data that is used an input for the judicial exception. Therefore, these limitations are mere data gathering activities. As set forth in MPEP 2106.05(g), mere data gathering activity has been identified by the courts as insignificant extra-solution activity that does not provide a practical application.
The limitations in claim 20 further limit the attributes of the positive sample received in claims 1 and 11, but do not integrate the judicial exception into a practical application because they just further limit data gathering activities but don’t change their position as data gathering activities.
The limitations of claims 18 and 19 directed to network embedding and using a graph neural network (GNN) to predict macromolecular interactions describe the GNN at a high level of generality. While detail is given to generating node vectors from similarity scores, no structure is recited for the operation of the GNN itself. The focus remains on the mathematical concepts (vectorization, similarity calculations and probability calculations), while the GNN is described functionally as a tool for processing the mathematical relationships and generating predictions, which does not integrate the abstract idea into a practical application. Instructions to apply a generic neural network without reciting a particular structure or configuration that improves technology is akin to performing the abstract idea in a generic computing environment, which do not render the abstract idea eligible (see MPEP 2106.05(f)).
The above recited additional elements do not provide a practical application of the recited judicial exception. As such, claims 1-8 and 11-20 are directed to an abstract idea (Step 2A, Prong 2: NO).
Step 2B
Claims found to be directed to a judicial exception are then further evaluated to determine if the claims recite an inventive concept that provides significantly more than the judicial exception itself (Step 2B).
The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception because the claims recite additional elements that equate to mere instructions to apply the recited exception in a generic computing environment or well-understood, routine and conventional activity.
Regarding the data gathering steps, as set forth in MPEP section 2106.05(g), the courts have decided that limitations that merely add an insignificant extra-solution activity, do not amount to an inventive concept, particularly when the activities are well-understood and conventional. Parker v. Flook, 437 U.S. 584, 588-89, 198 USPQ 193, 196 (1978). As set forth in MPEP section 2106.05(d), the courts have recognized that limitations directed to data gathering and output that are claimed as insignificant extra-solution activity are routine, well understood and conventional (Mayo Collaborative servs. V. Prometheus Labs., Inc., 566 U.S. at 79, 101 USPQ2d at 1968).
Similarly, the limitations of claims 18 and 19 directed to using a graph neural network (GNN) to make predictions, are describing an algorithm at a high level of generality. Instructions to apply a generic neural network without reciting a particular structure or configuration that improves technology is akin to performing the abstract idea in a generic computing environment. As set forth in MPEP 2106.05(f), claims that amount to nothing more than an instruction to apply the abstract idea using a generic computer do not render an abstract idea eligible (see Alice Corp., 573 U.S. at 223, 110 USPQ2d at 1983. See also 573 U.S. at 224, 110 USPQ2d at 1984). Furthermore, as set forth in MPEP 2106.05(d), performing repetitive calculations is a computer function that is recognized as well-understood, routine and conventional functions when they are claimed in merely a generic manner (Bancorp Services v. Sun Life, 687 F.3d 1266, 1278, 103 USPQ2d 1425, 1433 (Fed. Cir. 2012).
The additional elements do not comprise an inventive concept when considered individually or as an ordered combination that transforms the claimed judicial exception into a patent-eligible application of the judicial exception. Therefore, the claims do not amount to significantly more than the judicial exception itself (Step 2B: No). As such, claims 1-8 and 11-20 are not patent eligible.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
10. Claims 1-2 and 10 are rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11). The italicized text corresponds to the instant claim limitations.
Pertaining to claim 1, Thafar et al. and Cheng et al. teach a method for generating a negative sample set for predicting macromolecule-macromolecule interactions as indicated below. Note that Thafar et al. teaches characterizing protein-drug interactions (i.e. macromolecule-drug) interactions using similarity networks for both types of molecules, whereas Cheng et al. teaches identifying negative samples for macromolecule-macromolecule interactions (protein-RNA) interactions by using similarity networks for the proteins only. At the end of this section, an obviousness type rationales are provided for combining the teachings of these prior arts to teach the claimed invention.
Pertaining to claim 1, Thafar et al. discloses obtaining four gold standard datasets comprised of drugs, protein, and known drug-target interactions. Thafar et al. discloses the targets include enzymes, ion channels, G-protein-coupled receptors and nuclear receptors (p. 3, col. 2, para. 2; Table 1; receiving a positive sample set comprising pairs of macromolecules of a first type and macromolecules of a second type having macromolecule-macromolecule interaction).
Regarding claim 1, Thafar et al. discloses computing or retrieving 10 representations or characteristics that were used to determine drug similarity. Thafar et al. further discloses representing each similarity measure by a square matrix and applying a similarity network fusion algorithm (SNF) to construct a sample similarity network for each matrix, wherein molecules are represented by nodes and similarity is represented by edges. Thafar et al. further discloses that the SNF iteratively integrates the networks with the information from the other networks using K-nearest neighbor. (p. 4, col. 1, para. 1; p. 4, col. 2, para. 3-p. 5, col. 1, para. 1; Fig. 2; generating a first similarity map of the macromolecules of the first type).
Regarding claim 1, Thafar et al. discloses computing or retrieving 10 drug target/protein similarity matrices that were combined into a similarity. Thafar et al. further discloses representing each similarity measure by a square matrix and applying a similarity network fusion algorithm (SNF) to construct a sample similarity network for each matrix, wherein molecules are represented by nodes and similarity is represented by edges. Thafar et al. further discloses that the SNF iteratively integrates the networks with the information from the other networks using K-nearest neighbor ( p. 4, col. 1, para. 2; Fig. 2; p. 4, col. 2, para. 3-p. 5, col. 1, para. 1; generating a second similarity map of the macromolecules of the second type).
Regarding Claim 1, Thafar et al. further representing each similarity measure by a square matrix and for each macromolecule type (i.e. drugs or targets) applying a similarity network fusion algorithm (SNF) to construct a sample similarity network for each matrix, wherein molecules are represented by nodes and similarity is represented by edges. Thafar et al. further discloses that the SNF iteratively integrates each of the networks (i.e. all drug networks or all target networks) with the information from the other networks of the same type using K-nearest neighbor (p. 4, col. 2, para. 3-p. 5, col. 1, para. 1; Fig. 2; generating vectorized representations of nodes in the first similarity map and vectorized representations of nodes in the second similarity map).
Regarding claims 1, Thafar et al. is silent to the method wherein both molecules of the first type and the second type are macromolecules; and generating the negative sample set using the vectorized representations of nodes in the first similarity map and the vectorized representations of nodes in the second similarity map. However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Cheng et al.
Regarding claim 1, Cheng et al. teaches selecting high-quality negative samples for effectively predicting protein-RNA interactions, which is done by a method called FIRE, wherein reliable negative samples are selected based on first computing the protein-protein similarity matrix for each protein involved in positively-identified protein-RNA interaction with other proteins, and then using these data to select negative protein-RNA combinations to minimize potential interactions. To do this, Cheng et al. discloses first calculating protein-protein similarity scores based on 1) sequence similarity using Smith-Waterman score, 2) functional annotation semantic similarity and 3) protein domain similarity, wherein each protein is represented by a domain fingerprint (binary vector) whose elements encode the presence or absence of each retained conserved protein region (Pfam domain), and then combining the scores into an aggregated similarity score (AS) for each protein pair. Note that this score is derived using a binary vectorized representation of the protein-protein interactions.
Cheng et al. discloses using these data to select a negative sample set in three steps: 1) first constructing a positive set of protein-RNA interactors (Pk-Rj) and remove them from the protein-RNA interaction dataset; 2) generating aggregate protein-protein similarity scores for all proteins with positive interactions (Pk) and other proteins (Pi) and for all pairs Pi-Rj that do not have positive interactions, computing a score of interaction between Pi-Rj based on the aggregate similarity score between Pi and Pk, which is the confidence of Pi-Rj being a positive interaction; and 3) sorting the protein/RNA pairs (Pi-Rj ) by the aggregate score in increasing order and select those as negative samples as the top-m protein-RNA pairs (p. 4, col. 1, para. 2-col. 2, para. 3; Fig. 2; Fig. 3; generating the negative sample set using the vectorized representations of nodes in the first similarity map).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Cheng et al. taught that across 18 independent datasets and using three different classifiers, all tested models achieved better performance on the negative set selected by the disclosed FIRE method than on a randomly selected negative set (p. 9, col 2, para. 4). Therefore, one of ordinary skill in the art would have been motivated to utilize the method of selecting negative samples for improving protein-RNA interaction predictions taught Cheng et al. in the method to predict drug target interactions taught by Thafar et al. in order to improve the performance of their neural network-based models to predict molecular interactions. Furthermore, one of ordinary skill in the art would predict that the negative sample selection methods taught by Cheng et al. could be readily added to the method of Thafar et al. with a reasonable expectation of success because they both pertain to training machine learning models to predict macromolecular interactions involving proteins with other molecules (i.e. drugs and RNA, respectively) and because the operations taught by Cheng et al., which were applied to similarity values could also be applied to vectorized representations of similarity scores taught by Thafar et al. (e.g. one could sort by the first element of a vector, or sort by vector magnitude). Furthermore, the method of Thafar et al. already includes a step of calculating similarity scores for both proteins and drugs and Thafar et al. discloses extending their method to create a reliable set of negative drug-target interactions (p. 14, col. 2, para. 3). The invention is therefore prima facie obvious.
Furthermore, while Thafar et al. taught characterizing protein-drug interactions (i.e. macromolecule-drug) interactions using similarity networks for both types of molecules and Cheng et al. taught identifying negative samples for predicting macromolecule-macromolecule interactions (protein-RNA) interactions by using similarity networks for the proteins only, neither taught calculating a similarity score for two types of macromolecules for use in the prediction of macromolecule-macromolecule interactions. However, this would be obvious to try because there was a market need to solve the problem at the time of the effective filing date of the instant application. Cheng et al. disclosed that the quality of negative sets selected randomly cannot be guaranteed and that this will unavoidably impact prediction performance of classifiers trained on datasets with random negative samples (p. 2, col. 1, para. 5). Cheng et a. disclosed a solution to this problem by using protein similarities to improve selection of negative samples to improve detection of true protein-RNA interactions. However, another solution to the problem would be to use both protein similarities and RNA similarities in the selection of negative samples. A person having ordinary skill in the art could have pursued the known potential solution with a reasonable expectation of success because the same algorithms applied to quantify protein-protein similarities are also used for quantifying RNA-RNA similarities and these techniques have been widely used in the field.
Pertaining to Claim 2, Thafar et al. discloses using a similarity network fusion algorithm (SNF) to first construct a sample similarity network for each of the similarity matrices (i.e., drugs represent network nodes, and the similarity represents the networks’ weighted edges but without self-loop edges, and the same thing is done for the target proteins separately). Thafar et al. further discloses that the second step is to combine each set of networks (of either drugs or targets) into an aggregate network by iteratively updating the first network with information from other networks using K-nearest neighbor (p. 4, col. 2, para. 3. – p. 5, col. 1, para. 1; the method of claim 1, wherein the first similarity map or the second similarity map comprises nodes and edges connecting adjacent nodes, wherein a respective node represents a respective macromolecule, a respective edge represents a respective distance between a respective pair of the macromolecules, and a respective weight of the respective edge represents a respective similarity between the respective pair of the macromolecules).
Regarding claim 10, Thafar et al discloses generating a model and testing its ability to predict the novel drug-target interactions (DTIs) in each of the benchmark datasets separately using the following procedure: for each dataset, the model was first trained using all known interactions (positive labels) and split the unknown interactions (negative labels) into training and testing sets for each fold in the tenfold CV. In this manner, it was determined if any of the unknown DTI are predicted to be positive DTIs, and then DTIs predicted to be positive were ranked, based on their prediction scores. Only novel DTIs that were not part of the training set were reported and validated (p. 11, col. 1, para. 2; a method of predicting macromolecule-macromolecule interaction using the positive sample set and the negative sample set generated by the method of claim 1).
11. Claims 3, 4, 6 and 8 are rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11), as applied to claims 1-2 and 10 above, as evidenced by Lahitani et al. (Cosine similarity to determine similarity measure: Study case in online essay assessment, 2016, 2016 4 international conference, IEEE p. 1-6), and further in view of Gao et al. (2018, Proceedings of 27th international joint conference on artificial intelligence, JJCAI-18, p. 3371 - 3378). The italicized text corresponds to the instant claim limitations.
The limitations of claims 1-2 and 10 have been taught by Thafar et al. and Cheng et al. above.
Regarding claim 3, in an example case that macromolecule type 1 is protein and molecule type 2 is RNA, claim 3 is interpreted to mean that a positive set of protein-RNA interactions is used to generate intermediate sets of proteins, and for each intermediate set, the proteins are sorted based on their pairwise similarity to a protein in the positive set. Furthermore, the selected protein from the positive set is excluded from the intermediate set of similarity partners.
Regarding claim 3, Cheng et al. discloses constructing a positive set (PS) of protein-RNA interactions, and for protein pi and RNA rj that do not form a positive protein-RNA interactions in PS, compute a score between pi and rj as follows: a) if protein pk (k≠i) and rj forms a protein-RNA interaction in PS, then the score indicating the confidence of (pi,rj) being a positive protein-RNA interaction is evaluated based on the similarity between pi and pk. b) some scores are aggregated to account for multiple positive interactions involving rj. Cheng et al. further discloses that negative protein-RNA pairs are selected from this intermediate set (described above) by first sorting them by their scores in increasing order and selecting the top-m protein-RNA pairs in the sorted list as negative protein-RNA interactions if m negative PRIs are to be generated (p. 4, col. 2, para. 1-3; Fig. 3; claim 1, further comprising generating a plurality of intermediate sets; wherein the positive sample set is represented by {(m1i, m2i), i=1,...,K}, wherein m1i stands for an i-th macromolecule of the first type and m2i stands for an i-th macromolecule of the second type; wherein generating the respective intermediate set of the plurality of intermediate sets comprises: sorting m1j (j=1,...,K, and j ≠ i) based on similarities between m1i and m1j to obtain a subset of m1j (j=K, and j ≠ i); determining a probability of interaction between m2i and each sample in the subset; and generating the respective intermediate set based on the probability of interaction.
Regarding claim 3, Thafar et al. and Change et al. are silent to specifically determining a probability of interaction between m2i and each sample in the subset (rather than a score) (claim 3) and the method of claim 4, further comprising placing (m1j, 1-P(1|drj,dpi,)) into the respective intermediate set when P(1|drj, dpi) is less than a threshold value) (claim 6). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Gao et al.
Regarding claim 3, Gao et al. teaches a method to determine a probability from vectorized nodes using a sigmoid function to predict the probability that an interaction exists between a pair of protein and drug, from learned drug and protein representation using a neural-network classifier that outputs an interaction probability. Gao et al. discloses using the equation
P
y
=
1
p
,
d
=
1
1
+
e
-
v
p
*
v
d
to calculate the probability, wherein vp is a vector-based transformation of the Siamese network for protein and fd is the vector-based transformation of the Siamese network for drugs, respectively. (p. 3371, col. 2, para. 3 – p. 3372, col. 1, para. 1; p. 3374, col. 1, para. 4-6 (section 3.5); Equation 5; determining a probability of interaction between m2i and each sample in the subset).
Regarding claim 4, as evidenced by Lahitani et al., the claimed equation is the cosine similarity equation (p. 3, col. 2, para. 6, equation 3). Thafar et al. discloses first using node2fec graph embedding on drug-drug similarity graph and target-target similarity graph to represent each node in each graph with a feature vector in the graph. Thafar et al. further discloses computing two cosine similarity matrices based on node2vec feature representations for each drug pair and target pair because it gives unique similarity between nodes that carry meaningful topological, relational, and structural information (p. 5, col. 2, para. 2 – p. 6, col. 1, para. 2; p. 13, col. 2, para. 1- p. 14, col. 1, para. 1, p. 14, col. 2, para. 2; the method of claim 3, further comprising calculating similarities between m1i and m1j (j=1,...,K, and j ≠ i) by:
s
d
r
j
,
d
r
i
=
<
d
r
j
,
d
r
i
>
<
d
r
j
,
d
r
j
>
X
<
d
r
i
,
d
r
i
>
wherein dr stands for vectorized representations of nodes in the subset; and dri stands for a vectorized representation of m1i).
Claim 6 is a contingent claim and is executed only when the probability is less than a threshold; therefore, it is not considered to be a necessary step of the invention. Pertaining to claim 6, Gao et al. teaches that in a classification scenario, a hyper-parameter threshold d is selected as a classification boundary where given class categories y of either 0 or 1, if P(y = 1 | p,d) > d, category is one, otherwise, category is 0 (Equation 6; p. 3374, col. 1, para. 4; the method of claim 4, further comprising placing (m1j, 1-P(1|drj,dpi,)) into the respective intermediate set when P(1|drj, dpi) is less than a threshold value).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Gao et al. taught that by applying the probability equation to vectors and learning directly from molecular structures and protein sequences, their approach saves the effort of designing biochemical descriptors, which can be an expensive processes of feature engineering (p. 3372, col. 2, para. 1). Therefore, one of ordinary skill in the art would have been motivated to utilize the probability calculation taught by Gao et al. in the method to predict drug target interactions taught by Thafar et al. and Cheng et al. in order to reduce computational expense. Furthermore, one of ordinary skill in the art would predict that the method of calculating probabilities of macromolecular interactions taught by Gao et al. could be readily added to the method of Thafar et al. and Cheng et al. with a reasonable expectation of success because both methods employ learned vector representations of paired macromolecules to predict interactions. The invention is therefore prima facie obvious.
Pertaining to claim 8, Cheng et al. discloses sorting all generated potential protein-RNA interactions (PRIs) (pi, rj) in increasing order via their SPRij scores, scores which represent protein-protein similarity with proteins in positive examples. Cheng et al. further discloses that after sorting, the top-m protein-RNA pairs in the sorted list are taken as negative PRIs if m negative PRIs are to be generated. (p. 4, col. 2, para. 3; the method of claim 3, wherein generating the negative sample set comprises: sampling L number of negative samples from a respective intermediate set of the plurality of intermediate sets; wherein the negative sample set comprises negative samples sampled from the plurality of intermediate sets; and L is an integer equal to or greater than 1).
12. Claims 5 and 7 are rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11) as evidenced by Lahitani et al. (Cosine similarity to determine similarity measure: Study case in online essay assessment, 2016, 2016 4 international conference, IEEE p. 1-6), in view of Gao et al. (2018, Proceedings of 27th international joint conference on artificial intelligence, JJCAI-18, p. 3371 - 3378), as applied to claims 3, 4, 6 and 8 above, and further in view of Mou et al.( 2016, proceedings of the 54th Annual Meeting of the Association Of Computational Linguistics, Berlin, Germany, August 2026, p. 130-136). The italicized text corresponds to the instant claim limitations.
The limitations of claims 1-2 and 10 have been taught by Thafar et al. and Cheng et al. above.
The limitations of claims 3, 4, 6 and 8 have been taught by Thafar et al., Cheng et al. and Gao et al. above.
Pertaining to claim 5, the claimed equation
P
1
d
r
j
,
d
p
i
=
1
1
+
exp
-
ϴ
d
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
is a combination of two concepts: 1) a simply logistic regression (or sigmoid) of the structure:
y
=
1
1
+
e
-
x
, wherein x represents either features or feature vector(s) and 2) a feature vector in three parts including: vector concatenation (drj, dpi), vector difference (drj-dpi) and Hadamard product of vectors
d
r
j
⦻
d
p
i
. The three parts together are: (
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
)
. As described below, these two concepts are taught separately.
Regarding claim 5, Gao et al. teaches using a sigmoid function to predict the probability that an interaction exists between a pair of protein and drug, from learned drug and protein representation using a neural-network classifier that outputs an interaction probability. Gao et al. discloses using the equation
P
y
=
1
p
,
d
=
1
1
+
e
-
v
p
*
v
d
to calculate the probability, wherein vp is a transformation of the Siamese network for protein and fd is the transformation of the Siamese network for drugs, respectively. (p. 3371, col. 2, para. 3 – p. 3372, col. 1, para. 1; p. 3374, col. 1, para. 4-6 (section 3.5); Equation 5; the method of claim 3, wherein the probability of interaction between m2i and each sample in the subset is determined by:
P
1
d
r
j
,
d
p
i
=
1
1
+
exp
-
ϴ
d
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
, wherein drj stands for vectorized representations of nodes in the subset; dpl stands for vectorized representations of m2i; P(1|drj, dpi) stands for a probability of interaction between m2i and each sample in the subset; [,] stands for stitching between elements;
⦻
stands for product of two vectors; and
ϴ
stands for a parameter that is tunable.
Pertaining to claim 5, Thafar et al., Chang et al. and Gao et al. are silent to the three-part feature vector described as (
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
)
. However, this limitation was known in the art at the time of the effective filing date of the invention as taught by Mou et al.
Regarding claim 5, Mou et al. teaches a composite feature vector architecture to classify their relationship or sentence matching between two vectors using vector concatenation, vector difference and Hadamard product of vectors. Mou et al. teaches a method to combine vector representations of 2 individual sentences in 3 ways to compare similarity of the vectors including: concatenation of the two sentence vectors, Element-wise product (Hadamard product), and Element-wise difference (m=[h1;h2;h1 −h2;h1 ◦h2]), wherein “◦” denotes element-wise product; semi colons refer to column vector concatenation (p. 131, col. 2, para. 3 – p. 132, col. 1, para. 1; the method of claim 3, wherein the probability of interaction between m2i and each sample in the subset is determined by:
P
1
d
r
j
,
d
p
i
=
1
1
+
exp
-
ϴ
d
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
, wherein drj stands for vectorized representations of nodes in the subset; dpl stands for vectorized representations of m2i; P(1|drj, dpi) stands for a probability of interaction between m2i and each sample in the subset; [,] stands for stitching between elements;
⦻
stands for product of two vectors; and
ϴ
stands for a parameter that is tunable.
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Mou et al. taught that their approach to aggregate information from vectors using three matching heuristics significantly improves the performance in similarity analysis over other heuristic combinations (p. 133, col. 1, para. 2). Therefore, one of ordinary skill in the art would have been motivated to utilize the probability calculation taught by Mou et al. in the method to predict drug target interactions taught by Thafar et al., Cheng et al. and Gao et al. in order to improve the performance of their neural network-based models to predict drug targets. Furthermore, one of ordinary skill in the art would predict that the method of calculating probabilities of macromolecular interactions taught by Mou et al. could be readily added to the method of Thafar et al., Cheng et al. and Gao et al. with a reasonable expectation of success because both methods employ learned vector representations of paired entities to predict interactions. The invention is therefore prima facie obvious.
Claim 7 is interpreted to mean that selecting the number of negative samples from a set is determined based on a normalized probability value pk, which represents a softmax-style normalization wherein an individual probability component is divided by the sum of all components in a set to ensure the resulting probabilities pk sum to 1.
Regarding claim 7, Gao et al. teaches the claimed softmax equation for an attention mechanism to produce normalized importance weights for proteins and drugs. Gao et al. further teaches these weights are derived using a softmax operation in the form
α
i
e
x
p
(
a
i
)
∑
j
e
x
p
(
a
j
)
(
e
q
u
a
t
i
o
n
3
o
f
G
a
o
e
t
a
l
.
)
, which is the same softmax style of normalization claimed in the instant application with a different use (i.e. applied to interaction weights instead of interaction probabilities) (p. 3373, col. 2, para. 2 - p. 3374, col. 1, para. 3 (Section 3.4); Equation 3; the method of claim 5, wherein sampling L number of negative samples from the respective intermediate set is performed based on probability pk,k=1,…,|T|}; wherein
p
k
=
1
-
P
1
d
r
j
,
d
p
i
∑
I
=
1
T
1
-
P
1
d
r
j
,
d
p
i
, |T| stands for a number of elements in the respective intermediate set; and 1- P(1 | drj, dpi)) stands for a k-th element in the intermediate set).
Gao et al. taught a softmax normalization that was the same as the normalization taught in the claimed method in except for the simple substitution of the data being normalized. Whereas Gao et al. applied the equation to normalize importance weights, the claimed invention applied the same approach to normalize interaction probabilities. Thafar et al., Cheng et al., Gao et al. and Mou et al. taught the generation of interaction probabilities claimed in claim 5, as discussed above. Additionally, the softmax normalization taught by Gao et al. is a simple division of an individual quantity by the sum of corresponding quantities over the set. One of ordinary skill in the art could have substituted the vectors representing importance weights for the interaction probabilities, and the results of the substitution would have been predictable because the equation simply normalizes the vectors to produce a probability distribution whose components sum to one.
13. Claims 11, 13, 20 and 22 are rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Gao et al. (2018, Proceedings of 27th international joint conference on artificial intelligence, JJCAI-18, p. 3371 - 3378) and Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11). The italicized text corresponds to the instant claim limitations.
Pertaining to claim 1, Thafar et al., Gao et al. and Cheng et al. teach a method for generating a negative sample set for predicting macromolecule-macromolecule interactions as indicated below. Note that Thafar et al. and Gao et al. teach characterizing protein-drug interactions (i.e. macromolecule-drug) interactions using similarity networks for both types of molecules, whereas Cheng et al. teaches identifying negative samples for macromolecule-macromolecule interactions (protein-RNA) interactions by using similarity networks for the proteins only. At the end of this section, an obviousness type rationales are provided for combining the teachings of these prior arts to teach the claimed invention.
With respect to claim 11, Thafar et al. discloses obtaining four gold standard datasets comprised of drugs, targets (proteins), and known drug-target interactions. Thafar et al. discloses the targets include enzymes, ion channels, G-protein-coupled receptors and nuclear receptors (p. 3, col. 2, para. 2; Table 1; receiving a positive sample set comprising pairs of macromolecules of a first type and macromolecules of a second type having macromolecule-macromolecule interaction
Regarding claim 11 Thafar et al. further teaches representing each similarity measure by a square matrix and for each macromolecule type (i.e. drugs or targets) applying a similarity network fusion algorithm (SNF) to construct a sample similarity network for each matrix, wherein molecules are represented by nodes and similarity is represented by edges. Thafar et al. further discloses that the SNF iteratively integrates each of the networks (i.e. all drug networks or all target networks) with the information from the other networks of the same type using K-nearest neighbor (p. 4, col. 2, para. 3-p. 5, col. 1, para. 1; Fig. 2; recite: generating a first similarity map of macromolecules of a first type and generating a second similarity map of macromolecules of a second type).
Regarding claim 11, Thafar et al. further representing each similarity measure by a square matrix and for each macromolecule type (i.e. drugs or targets) applying a similarity network fusion algorithm (SNF) to construct a sample similarity network for each matrix, wherein molecules are represented by nodes and similarity is represented by edges. Thafar et al. further discloses that the SNF iteratively integrates each of the networks (i.e. all drug networks or all target networks) with the information from the other networks of the same type using K-nearest neighbor (p. 4, col. 2, para. 3-p. 5, col. 1, para. 1; Fig. 2; generating vectorized representations of nodes in the first similarity map and vectorized representations of nodes in the second similarity map).
With respect to claim 11, Thafar et al. further discloses using the vectorized representations of the nodes in the first and second similarity maps (i.e. the drug-drug and target-target similarity maps) to calculate drug-drug cosine similarities and target-target cosine similarities and using these data to calculate path scores of drug target interactions (Fig. 4; p. 8, col. 1, para. 4 – col. 2, para. 1; determining a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map).
Pertaining to claim 11, Thafar et al. discloses using a supervised machine learning model to predict drug target interactions based on three different classifiers for each dataset: 1) an artificial neural network, 2) random forest and 3) adaptive boosting classifiers. Thafar et al. further discloses using the path scores representing the likelihood of drug-target interactions to generate the class labels for the samples used to train and test the machine learning classifiers (p. 8, col. 1, para. 3 – col. 2, para. 1; Fig. 4; recites: training the model at least partially based on the probability of interaction).
Regarding claim 11, Thafer et al. is silent to specifically determining a probability of interaction between a first and second respective vectorized representation of a node (rather than a score of interaction) (claim 11) and the method of claim 11, wherein the positive sample set is represented by {(m1i, m2i), i=1,...,K}, wherein m1i stands for an i-th macromolecule of the first type and m2i stands for an i-th macromolecule of the second type; wherein training the model comprises minimizing a loss function:
L
=
-
∑
i
=
1
K
l
o
g
p
1
|
d
m
1
i
,
d
m
2
i
;
wherein dm1i stands for vectorized representations of nodes in the first similarity map; dm2i stands for vectorized representations of nodes in the second similarity map; p (1 |dm1i, dm2i) stands for a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map (claim 20). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Gao et al.
Regarding claim 11, Gao et al. teaches a method to determine a probability from vectorized nodes using a sigmoid function to predict the probability that an interaction exists between a pair of protein and drug, from learned drug and protein representation using a neural-network classifier that outputs an interaction probability. Gao et al. discloses using the equation
P
y
=
1
p
,
d
=
1
1
+
e
-
v
p
*
v
d
to calculate the probability, wherein vp is a vector-based transformation of the Siamese network for protein and fd is the vector-based transformation of the Siamese network for drugs, respectively. (p. 3371, col. 2, para. 3 – p. 3372, col. 1, para. 1; p. 3374, col. 1, para. 4-6 (section 3.5); Equation 5; determining a probability of interaction between m2i and each sample in the subset).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Gao et al. taught that by applying the probability equation to vectors and learning directly from molecular structures and protein sequences, their approach saves the effort of designing biochemical descriptors, which can be an expensive processes of feature engineering (p. 3372, col. 2, para. 1). Therefore, one of ordinary skill in the art would have been motivated to utilize the probability calculation taught by Gao et al. in the method to predict drug target interactions taught by Thafar et al. in order to reduce computational expense. Furthermore, one of ordinary skill in the art would predict that the method of calculating probabilities of macromolecular interactions taught by Gao et al. could be readily added to the method of Thafar et al. with a reasonable expectation of success because both methods employ learned vector representations of paired macromolecules to predict interactions. The invention is therefore prima facie obvious.
Pertaining to claim 11, Thafar et al. further discloses that output of the classifier is the class label, which is either a positive or negative label (p. 8, col. 2, para. 1). However, Thafar et al. and Gao et al. are silent to training a model for the specific purpose of generating a negative sample set for predicting macromolecule-macromolecule interaction. Thafar et al. and Gao et al. are also silent to the method wherein the method wherein molecules of the first type and the second type are both macromolecules. However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Cheng et al.
Regarding claim 11, Cheng et al. taught selecting high-quality negative samples for effectively predicting protein-RNA interactions. Cheng et al. taught doing this based on the idea that for an experimentally-validated protein RNA interaction of protein p and RNA r, r is highly possible to interact with any protein p’ similar to p. On the contrary, if protein p’ is dissimilar to p, there is low possibility that p’ interacts r. Cheng et al. further disclosed applying this idea by first computing similarity between each pair of proteins and then selecting negative protein-RNA interaction samples based on lack of similarity between the proteins in the negative interacting pair and the protein in a positive interacting pair. Cheng et al. further taught training a model to identify protein-RNA interacting pairs using the identified negative samples as the negative class and positive samples as the positive class (abstract, p. 4, col. 1, para. 2 – col. 2, para. 3; p. 2, col. 2, para. 2; a method of training a model for generating a negative sample set for predicting macromolecule-macromolecule interaction).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Cheng et al. taught that when using the computationally determined negative sample set for ML-based classification of protein-RNA interactions, all tested models achieved better performance on the negative set selected by the disclosed FIRE method than on a randomly selected negative set. (p. 9, col 2, para. 4). Therefore, one of ordinary skill in the art would have been motivated to use computationally determined negative samples taught by Thafar et al. and Gao et al. (i.e. the negative samples identified based on drug-drug similarity and protein-protein similarity followed by neural network based classification) as gold standard negative samples in future classification studies as taught by Cheng et al. in order to improve the performance of their neural network-based models to predict drug targets. Furthermore, one of ordinary skill in the art would predict that the reliable negative sample selection method taught by Cheng et al. could be readily added to the method of Thafar et al. and Gao et al. with a reasonable expectation of success because they both pertain to training machine learning models to predict macromolecular interactions between proteins with other macromolecules. Furthermore, all aspects of the claimed invention are taught by Thafar et al. and Gao et al. except for the utilization of predicted negative samples from the machine learning classification as gold standard negative samples in future machine learning classification studies. Furthermore, Thafar et al. discloses the importance of having a reliable set of negative drug target interactions for modeling (p. 14, col. 2, para. 3). The invention is therefore prima facie obvious.
Furthermore, while Thafar et al. and Gao et al. taught characterizing protein-drug interactions (i.e. macromolecule-drug) interactions using similarity networks for both types of molecules and Cheng et al. taught identifying negative samples for predicting macromolecule-macromolecule interactions (protein-RNA) interactions by using similarity networks for the proteins only, neither taught calculating a similarity score for two types of macromolecules for use in the prediction of macromolecule-macromolecule interactions. However, this would be obvious to try because there was a market need to solve the problem at the time of the effective filing date of the instant application. Cheng et al. disclosed that the quality of negative sets selected randomly cannot be guaranteed and that this will unavoidably impact prediction performance of classifiers trained on datasets with random negative samples (p. 2, col. 1, para. 5). Cheng et a. disclosed a solution to this problem by using protein similarities to improve selection of negative samples to improve detection of true protein-RNA interactions. However, another solution to the problem would be to use both protein similarities and RNA similarities in the selection of negative samples. A person having ordinary skill in the art could have pursued the known potential solution with a reasonable expectation of success because the same algorithms applied to quantify protein-protein similarities are also used for quantifying RNA-RNA similarities and these techniques have been widely used in the field.
Pertaining to claim 20, Gao et al. teaches a loss function equivalent to the loss function claimed, wherein given a dataset D = {(pi, di, yi)}, i = 1…n, the model can be trained by minimizing the likelihood of observing the training data, which is equivalent to minimizing the cross entropy loss function:
-
∑
i
=
1
n
y
i
l
o
g
σ
p
i
,
d
i
+
1
-
y
i
l
o
g
1
-
σ
p
i
,
d
i
.
Wherein the goal is to find the set of parameters that minimizes the total loss. yi is the ground truth label for the ith sample (either 0 or 1) and s(pi,di) is the predicted probability that sample i belongs to class 1. (1 – yi) selects the second term of the equation when the true label is 0. pi and di are the are the vectorized representations of protein and drug nodes, respective. When this equation taught by Gao et al. is applied to samples that are only positive (as the application as claimed in the instant application), the second half of the equation
1
-
y
i
l
o
g
1
-
σ
p
i
,
d
i
is not used and thus the equation disclosed by Gao et al. is identical to the claimed equation (p. 3374, col. 1, para. 7; the method of claim 11, wherein the positive sample set is represented by {(m1i, m2i), i=1,...,K}, wherein m1i stands for an i-th macromolecule of the first type and m2i stands for an i-th macromolecule of the second type; wherein training the model comprises minimizing a loss function:
L
=
-
∑
i
=
1
K
l
o
g
p
1
|
d
m
1
i
,
d
m
2
i
;
wherein dm1i stands for vectorized representations of nodes in the first similarity map; dm2i stands for vectorized representations of nodes in the second similarity map; p (1 |dm1i, dm2i) stands for a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map.
Regarding claim 13, Thafar et al. discloses using a similarity network fusion algorithm (SNF) to first construct a sample similarity network for each of the similarity matrices (i.e., drugs represent network nodes, and the similarity represents the networks’ weighted edges but without self-loop edges, and the same thing is done for the target proteins separately). Thafar et al. further discloses that the second step is to combine each set of networks (of either drugs or targets) into an aggregate network by iteratively updating the first network with information from other networks using K-nearest neighbor (p. 4, col. 2, para. 3. – p. 5, col. 1, para. 1; the method of claim 11, wherein the first similarity map or the second similarity map comprises nodes and edges connecting adjacent nodes, wherein a respective node represents a respective macromolecule, a respective edge represents a respective distance between a respective pair of the macromolecules, and a respective weight of the respective edge represents a respective similarity between the respective pair of the macromolecules).
Regarding claim 22, Thafar et al. discloses developing supervised machine learning models to predict drug-target interactions based on three different classifiers for each dataset including: 1) artificial neural network, 2) random forest and 3) adaptive boosting (p. 8, col. 1, para. 3; a neural network model for predicting macromolecule-macromolecule interaction, trained by the method of claim 11).
14. Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Gao et al. (2018, Proceedings of 27th international joint conference on artificial intelligence, JJCAI-18, p. 3371 - 3378) and ) and Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11), as applied to claims 11, 13, 20 and 22 above, and further in view of Mou et al.( 2016, proceedings of the 54th Annual Meeting of the Association Of Computational Linguistics, Berlin, Germany, August 2026, p. 130-136). The italicized text corresponds to the instant claim limitations.
The limitations of claims 11, 13, 20 and 22 have been taught by Thafar et al., Gao et al. and Cheng et al. above.
Pertaining to claim 12, the claimed equation
P
1
d
r
j
,
d
p
i
=
1
1
+
exp
-
ϴ
d
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
is a combination of two concepts: 1) a simply logistic regression (or sigmoid) of the structure:
y
=
1
1
+
e
-
x
, wherein x represents either features or feature vector(s) and 2) a feature vector in three parts including: vector concatenation (drj, dpi), vector difference (drj-dpi) and Hadamard product of vectors
d
r
j
⦻
d
p
i
. The three parts together are: (
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
)
. As described below, these two concepts are taught separately.
Pertaining to claim 12, Gao et al. teaches a method to determine a probability from vectorized nodes using a sigmoid function to predict the probability that an interaction exists between a pair of protein and drug, from learned drug and protein representation using a neural-network classifier that outputs an interaction probability. Gao et al. discloses using the equation
P
y
=
1
p
,
d
=
1
1
+
e
-
v
p
*
v
d
to calculate the probability, wherein vp is a vector-based transformation of the Siamese network for protein and fd is the vector-based transformation of the Siamese network for drugs, respectively. (p. 3371, col. 2, para. 3 – p. 3372, col. 1, para. 1; p. 3374, col. 1, para. 4-6 (section 3.5); Equation 5; the method of claim 11, wherein the probability of interaction is determined by:
P
1
d
m
1
i
,
d
m
2
j
=
1
1
+
exp
-
ϴ
d
m
1
i
,
d
m
2
j
,
d
m
1
i
-
d
m
2
j
,
d
m
1
i
⦻
d
m
2
j
, wherein dm1 stands for vectorized representations of nodes in the first similarity map; dm2 stands for vectorized representations of nodes in the second similarity map; p (1 |dm1i, dm2i) stands for a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map; [,] stands for stitching between elements; and
⦻
stands for product of two vectors; and
ϴ
stands for a parameter that is tunable).
Regarding claim 12, Thafer et al. and Gao et al. are silent to the three-part feature vector described as (
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
)
in the denominator of the claimed equation used for determining the probability of interactions: (
P
1
d
r
j
,
d
p
i
=
1
1
+
exp
-
ϴ
d
r
j
,
d
p
i
,
d
r
j
-
d
p
i
,
d
r
j
⦻
d
p
i
). However, this limitation was known in the art at the time of the effective filing date of the invention as taught by Mou et al.
Regarding claim 12, Mou et al. teaches a composite feature vector architecture to classify their relationship or sentence matching between two vectors using vector concatenation, vector difference and Hadamard product of vectors. Mou et al. teaches a method to combine vector representations of 2 individual sentences in 3 ways to compare similarity of the vectors including: concatenation of the two sentence vectors, Element-wise product (Hadamard product), and Element-wise difference (m=[h1;h2;h1 −h2;h1 ◦h2]), wherein “◦” denotes element-wise product; semi colons refer to column vector concatenation (p. 131, col. 2, para. 3 – p. 132, col. 1, para. 1; the method of claim 11, wherein the probability of interaction is determined by:
P
1
d
m
1
i
,
d
m
2
j
=
1
1
+
exp
-
ϴ
d
m
1
i
,
d
m
2
j
,
d
m
1
i
-
d
m
2
j
,
d
m
1
i
⦻
d
m
2
j
, wherein dm1 stands for vectorized representations of nodes in the first similarity map; dm2 stands for vectorized representations of nodes in the second similarity map; p (1 |dm1i, dm2i) stands for a probability of interaction between a first respective vectorized representation of a node in the first similarity map and a second respective vectorized representation of a node in the second similarity map; [,] stands for stitching between elements; and
⦻
stands for product of two vectors; and
ϴ
stands for a parameter that is tunable).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Mou et al. taught that their approach to aggregate information from vectors using three matching heuristics significantly improves the performance in similarity analysis over other heuristic combinations (p. 133, col. 1, para. 2). Therefore, one of ordinary skill in the art would have been motivated to utilize the probability calculation taught by Mou et al. in the method to predict drug target interactions taught by Thafar et al. and Gao et al. in order to improve the performance of their neural network-based models to predict drug targets. Furthermore, one of ordinary skill in the art would predict that the method of calculating probabilities of macromolecular interactions taught by Mou et al. could be readily added to the method of Thafar et al. and Gao et al. with a reasonable expectation of success because both methods employ learned vector representations of paired entities to predict interactions. The invention is therefore prima facie obvious.
15. Claims 14-18 are rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Gao et al. (2018, Proceedings of 27th international joint conference on artificial intelligence, JJCAI-18, p. 3371 - 3378) and ) and Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11), as applied to claims 11, 13, 20 and 22 above, in view of Zhang et al. (SIGMOD’10 conference proceedings, 2010, Indianapolis, Indiana, June 6-11, 2010), as evidenced by NovoPro (NovoPro Biosciences Inc., news highlights, 2014, p. 1-2, What is the difference between homology, similarity and identity?; https://www.novoprolabs.com/support/articles/what-is-the-difference-between-homology-similarity-and-identity-201803011302.html). The italicized text corresponds to the instant claim limitations.
The limitations of claims 11, 13, 20 and 22 have been taught by Thafar et al., Gao et al. and Cheng et al. above.
Regarding claims 14-18, these claims merely show how similarities and distances are expressed but they are not interpreted to be part of the method because there is no step to calculate similarities.
Regarding to claims 14-17, Thafar et al. and Gao et al. are silent to: the method of claim 11, wherein a respective similarity between a respective pair of the macromolecules of the first type is expressed as: sim1 (m1-1,m1-2)= 1 - d1(m1-1,m1-2); wherein (m1-1 , m1-2) stands for the respective pair of the macromolecules of the first type, sim1 stands for the respective similarity between the respective pair of the macromolecules of the first type, and d1 stands for a distance between the respective pair of the macromolecules of the first type (claim 14); and the method of claim 11, wherein a respective similarity between a respective pair of the macromolecules of the second type is expressed as: sim2 (m2-1,m2-2)= 1 - d2(m2-1,m2-2); wherein (m2-1,m2-2)) stands for the respective pair of the macromolecules of the second type, sim2 stands for the respective similarity between the respective pair of the macromolecules of the second type, and d2 stands for a distance between the respective pair of the macromolecules of the second type (claim 16). However, this limitation was known in the art at the time of the effective filing date of the invention as taught by Zhang et al.
Pertaining to claims 14 and 16, Zhang et al. teaches a widely used measure of string similarity called edit distance, which is defined as the minimum number of primitive operations (insertions, deletions and substitutions needed to transfer one string into another. As evidenced by NovoPro, the similarity between a pair of strings, for example the similarity between two DNA sequences, is calculated using this equation: similarity = 1 – edit distance/unaligned length of the shorter sequence (Zhang et al. p. 915, col. 2, para. 1-3; NovoPro et al. p. 2, para. 4: the method of claim 11, wherein a respective similarity between a respective pair of the macromolecules of the first type is expressed as: sim1 (m1-1,m1-2)= 1 - d1(m1-1,m1-2); wherein (m1-1 , m1-2) stands for the respective pair of the macromolecules of the first type, sim1 stands for the respective similarity between the respective pair of the macromolecules of the first type, and d1 stands for a distance between the respective pair of the macromolecules of the first type (claim 14); and the method of claim 11, wherein a respective similarity between a respective pair of the macromolecules of the second type is expressed as: sim2 (m2-1,m2-2)= 1 - d2(m2-1,m2-2); wherein (m2-1,m2-2)) stands for the respective pair of the macromolecules of the second type, sim2 stands for the respective similarity between the respective pair of the macromolecules of the second type, and d2 stands for a distance between the respective pair of the macromolecules of the second type (claim 16).
Pertaining to claims 15 and 17, Zhang et al. teaches a generic and widely-used equation for calculating normalized edit distance between sequences that could be readily applied to proteins. Specifically, Zhang et al. teaches that edit distance between two strings si and sj, is the minimum number of primitive potations needed to transform si to sj, denoted by d(si,sj). Zhang et al. further teaches that normalized edit distance is defined as
d
'
s
i
,
s
j
=
d
(
s
i
,
s
j
)
m
a
x
{
s
i
,
s
j
}
, wherein |si| and |sj| are string lengths (p. 917, col. 2, para. 1-2: the method of claim 14, wherein d1 is expressed as:
d
1
m
1
-
1
,
m
1
-
2
=
l
e
v
m
1
-
1
,
m
1
-
2
max
l
e
n
m
1
-
1
,
m
1
-
2
, wherein lev(m1-1 , m1-2) stands for an edit distance between the respective pair of the macromolecules of the first type, len(m1-1) stands for a length of a first macromolecule of the first type in the respective pair, and len(m1-2) stands for a length of a second macromolecule of the first type in the respective pair (claim 15); : the method of claim 16, wherein d2 is expressed as:
d
2
m
2
-
1
,
m
2
-
2
=
l
e
v
m
2
-
1
,
m
2
-
2
max
l
e
n
m
2
-
1
,
m
2
-
2
; wherein, lev(m2-1,m2-2) stands for an edit distance between the respective pair of the macromolecules of the second type, len(m2-1) stands for a length of a first macromolecule of the second type in the respective pair, and len(m2-2) stands for a length of a second macromolecule of the second type in the respective pair (claim 17)).
As set forth in the MPEP 2143(A), an invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if the prior art included each element claimed although not necessarily in a single reference, with the only difference between the claimed invention and the prior art being the lack of actual combination of the elements in a single reference. Zhang et al. teaches that there is a long stream of research on defining string similarity measures including ‘edit distance’ and that several variations of edit distance have been proposed (p. 915, col. 2, para. 3). One of ordinary sill in the art could have combined the edit distance calculations taught by Zhang et al. and the molecular interaction prediction method taught by xxx et al. by the known methods of applying the generic edit distance calculations to strings such as DNA and protein sequences taught by Zhang et al. Furthermore, in combination, the similarity and distance calculations taught by Zhang et al. and the interaction prediction method taught by xxx merely performs the same functions as the do separately. Furthermore, one of ordinary skill in the art would predict that the generic similarity and distance metrics taught by Zhang et al. could be readily added to determine protein-protein similarities in the method of xxx et al. with a reasonable expectation of success because Zhang et al. provides a method of determining length-independent measures of similarities between strings suitable for comparing similarities between proteins of different sizes. The invention is therefore prima facie obvious.
16. Claims 18-19 are rejected under 35 U.S.C. 103 as being unpatentable over of Thafar et al. (J Cheminform (2020) 12, p. 1-17), in view of Gao et al. (2018, Proceedings of 27th international joint conference on artificial intelligence, JJCAI-18, p. 3371 - 3378) and Cheng et al. (BMC Systems Biology 2017, vol 11 (Suppl 2), p. 1-11) as applied to claims 11, 13, 20 and 22 above, in view of Zhi et al. (Biomolecules, Vol. 11, 2021, p. 1-37). The italicized text corresponds to the instant claim limitations.
The limitations of claims 11, 13, 20 and 22 have been taught by Thafar et al., Gao et al. and Cheng et al. above.
Regarding claims 18-20 and 22, Thafar et al. and Gao et al. are silent to: the method of claim 11, wherein the first similarity map includes N1 number of nodes, {ei, i=1,...,N1}; and M1 number of edges, {rj,j=1,...,M}; a respective vectorized representation of a respective node in the first similarity map is expressed as:
h
t
1
+
1
e
i
=
σ
(
W
p
X
h
t
1
e
i
+
∑
e
k
⋲
N
e
i
W
p
h
X
h
t
1
(
e
k
)
)
;
(claim 18); and the method of claim 11, wherein the second similarity map includes N2 number of nodes, {e'i, i=1,...,N2}; and M2 number of edges, {r'j,j=1,...,M2}; a respective vectorized representation of a respective node in the second similarity map is expressed as:
h
t
2
+
1
e
'
i
=
σ
∑
e
'
k
⋲
N
e
'
i
U
e
i
α
i
,
k
t
2
+
1
W
p
t
+
1
X
h
t
2
e
'
k
;
α
i
,
k
t
2
+
1
=
s
o
f
t
m
a
x
h
t
2
e
i
'
,
h
t
2
e
'
k
(claim 19). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Zhi et al.
Regarding claim 18, the equation claimed describes the update rule for the hidden state of a node in a Graph Neural Network at time step t1+1 which is equal to a non-linear activation function (Sigmoid or ReLU), a transformation of the entity’s own previous state using a weight matrix, plus the aggregation states from the neighboring nodes, which sums the transformed features of all adjacent entities using the weight matrix.
Pertaining to claim 18, Zhi et al. . the method of claim 11, wherein the first similarity map includes N1 number of nodes, {ei, i=1,...,N1}; and M1 number of edges, {rj,j=1,...,M}; a respective vectorized representation of a respective node in the first similarity map is expressed as:
h
t
1
+
1
e
i
=
σ
(
W
p
X
h
t
1
e
i
+
∑
e
k
⋲
N
e
i
W
p
h
X
h
t
1
(
e
k
)
)
;
wherein ei stands for a respective node in the first similarity map; ht1(ei) stands for a respective vectorized representation of the respective node ei prior to a t1-th step reiteration; ht1+1(ei) stands for an updated respective vectorized representation of the respective node ei subsequent to the t1-th step reiteration;
σ
stands for a leaky relu activation function; N(ei) stands for a set of nodes neighboring the respective node ei; and Wp, Wph stand for parameters of a graph neural network for generating the vectorized representation
Pertaining to claim 19, the claim is interpreted to include 1) a node update approach, wherein a nodes representation is updated based on its neighbors. An updated feature vector for a node i at step t2+1 equals an activation function (e.g. ReLU or sigmoid activation) multiplied by the sum of (a weight matrix, attention coefficients and a feature vector) and 2) An attention mechanism that computes importance of neighbor k to node i using a dot product similarity (indicated by <.,.>), followed by a softmax normalization over all neighbors. It calculates attention weights based on feature similarity and uses the weights to perform a weighted sum of neighboring node features to compute the next state of the central node. this mechanism helps the model identify which parts of a molecule (e.g., specific atoms or reactive sites) are most influential for a given property or binding affinity, effectively "mapping" the structural importance of the molecular graph.
PNG
media_image1.png
92
186
media_image1.png
Greyscale
Regarding claim 19, Zhi et al. discloses applying a graph neural network (GNN) methodology to investigate protein drug interactions. Zhi et al. further discloses developing a GIAT model that combines two spatial-based GNN approaches including GIN (which distinguishes between each of the different graph structures based on the entire molecular structure) and GAT (graph attention network), which uses an attention mechanism to achieve better neighbor aggregation to update the target node’s representation by learning from its neighbors and local environment. Zhi et al. further discloses an equation used in GAT wherein the updated vectorized representation of a node (h) is calculated as the product of an activation function (s) and the sum of the aggregate influence of a vectorized representation of neighboring nodes that is equivalent to the claimed equation (see Fig. 1 of this office action).
Figure 1, Equation 8 from Zhi et al. showing calculation of output features (h’i) of the GAT model for a node i where the output features is equal to a non-linear activation function s, multiplied by a weighted sum (where i is the target node, hi is the input features vector of node i, as is hj to the neighbor node j). aij is the weighted and normalized attention coefficients and W is a trainable weight matrix.
[AltContent: textbox (Figure 2. Equation 7of Zhi et al. calculating cross-attention between drug and protein features. eij is the attention coefficient leakyReLu(alpha[Whi·Whi]))]
PNG
media_image2.png
67
276
media_image2.png
Greyscale
Regarding the second equation claimed, Zhi et al. teaches an equation, which is computation of cross-attention between drug features and protein features. Attention scores are generated from transformed embeddings and normalized with softmax, which is very close conceptually to the claimed equation as it is simply a normalized version of the claimed equation (see Figure 2 of this office action)
(p. 10, para. 2, section 2.6.3; p. 6, para. 1 – p. 7. para. 1; Fig. 1, Fig. 3, abstract, equations 6-8; the method of claim 11, wherein the second similarity map includes N2 number of nodes, {e'i, i=1,...,N2}; and M2 number of edges, {r'j,j=1,...,M2}; a respective vectorized representation of a respective node in the second similarity map is expressed as:
h
t
2
+
1
e
'
i
=
σ
∑
e
'
k
⋲
N
e
'
i
U
e
i
α
i
,
k
t
2
+
1
W
p
t
+
1
X
h
t
2
e
'
k
;
α
i
,
k
t
2
+
1
=
s
o
f
t
m
a
x
h
t
2
e
i
'
,
h
t
2
e
'
k
; wherein e'i stands for a respective node in the second similarity map; (e'i) stands for a respective vectorized representation of the respective node e'i prior to a t2-th step reiteration; htz+1 (e'i) stands for an updated respective vectorized representation of the respective node e'i subsequent to the t2-th step reiteration; a stands for a leaky relu activation function; ; N(e'i) stands for a set of nodes neighboring the respective node e'i;
h
t
2
e
i
'
,
h
t
2
e
'
k
stands for an inner product of ht2(e’j) and ht2 (e’k); Wpt2+1 stands for a parameter of a graph neural network for generating the vectorized representation; and
α
i
,
k
t
2
+
1
stands for attention weights representing a link strength between node e’i and node e’k).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Zhi et al. discloses that the GIAT model that combines the GIN and GAT spatial models combines a global approach and a local approach, which improves performance of the prediction model (p. 10, para. 2). Zhi et al. also discloses that the application of GNNs to the discovery and development of multitarget drugs is feasible and reliable and provides a good basis for further biological experiments to identify and validate three candidate inhibitors of dihydroorotate dehydrogenase, a drug target for lung cancer (p. 35, para. 1). Therefore, one of ordinary skill in the art would have been motivated to utilize the node update and attention mechanism taught by Zhi et al. in the macromolecular interaction prediction method taught by Thafar et al. and Gao et al. in order to make predictions of good drug inhibitors for targets in treatment of lung cancer. Furthermore, one of ordinary skill in the art would predict that the node update approach and attention mechanism taught by Zhi et al. could be readily added to the interaction prediction method of Thafar et al. and Gao et al. with a reasonable expectation of success because they both predicts protein-ligand complexes based on GNNs. By the method of Zhi et al., the algorithms are applied to protein-drug networks, whereas in the teaching of Thafar et al. and Gao et al. the networks are either drug-drug or protein-protein similarity networks; however, the algorithms are not specific to one particular type of network. As disclosed by Zhi et al., the GAT method has been widely used in various types of deep learning tasks and it is one of the most noteworthy core technologies in deep learning (p. 6, para. 1). The invention is therefore prima facie obvious.
E-mail Communications Authorization
17. Per updated USPTO Internet usage policies, Applicant and/or applicant's representative is encouraged to authorize the USPTO examiner to discuss any subject matter concerning the above application via Internet e-mail communications. See MPEP 502.03. To approve such communications, Applicant must provide written authorization for e-mail communication by submitting the following statement via EFS-Web (using PTO/SB/439) or Central Fax (571-273-8300): "Recognizing that Internet communications are not secure, / hereby authorize the USPTO to communicate with the undersigned and practitioners in accordance with 37 CFR 1.33 and 37 CFR 1.34 concerning any subject matter of this application by video conferencing, instant messaging, or electronic mail. / understand that a copy of these communications will be made of record in the application file."
Written authorizations submitted to the Examiner via e-mail are NOT proper. Written authorizations must be submitted via EFS-Web (using PTO/SB/439) or Central Fax (571-273- 8300). A paper copy of e-mail correspondence will be placed in the patent application when appropriate. E-mails from the USPTO are for the sole use of the intended recipient, and may contain information subject to the confidentiality requirement set forth in 35 USC § 122. See also MPEP 502.03.
Inquiries
18. Any inquiry concerning this communication or earlier communications from the examiner should be directed to JENNIFER J SMITH whose telephone number is (571)272-7801. The examiner can normally be reached Monday-Friday 7:00 AM - 3:00 PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Olivia Wise can be reached at (571) 272-2249. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/J.J.S./Examiner, Art Unit 1685
/OLIVIA M. WISE/Supervisory Patent Examiner, Art Unit 1685
Prosecution Insights