KR20220004945A - Scaffold-based molecular design with a graph generative model - Google Patents

Abstract

The present invention relates to a scaffold-based molecular design method capable of simultaneously controlling multiple desired properties. When a method for generating an extended molecular structure having a predetermined scaffold structure of the present invention is used, a desired molecular structure with optimized performance can be efficiently derived by using only a small amount of labeled data.

Description

Scaffold-based molecular design with a graph generative model

The present invention relates to a scaffold-based molecular design method capable of simultaneously controlling multiple desired properties.

The goal of molecular design is to find molecules with the desired functionality. The extensive chemical space offers limitless possibilities for discovering new molecules in a variety of applications. But at the same time, the very wide range of characteristics presents difficulties when trying to find the most promising candidates in an acceptable time period. For example, in drug discovery, the number of potential drug-like molecules ^{is estimated to be about 10 23} to 10 ⁶⁰ [1], of which ^{only 10 8} have been synthesized [2]. Therefore, it is virtually impossible to discover new drugs by indiscriminate exploration of the chemical space, and the demand for more rational approaches has inevitably surged.

A common strategy in rational molecular design is to narrow the search space, starting with molecules known to be probable. The design of a new molecule proceeds by finding the right combination of functional groups to optimize the necessary properties, while the core structure or scaffold of the original molecule is intentionally maintained to retain its basic properties. This scaffold-based design has been one of the standard approaches in small molecule drug discovery, where we first identify the scaffolds based on information from the target protein and use derivative compounds to find those that exhibit optimal potency and selectivity. Search the library [3-5]. Molecular design in other fields, such as materials chemistry, employs a similar strategy. A representative example is organic electronics in which typical electron-donor and acceptor moieties are known. Proper selection or combination of scaffold moieties followed by side-chain optimization enables the design of novel components in organic light-emitting diodes [6], field-effect transistors [7], photocatalysts [8] and solar cells [9]. can do. However, while modifying the core structure significantly reduces the search space, in most practical applications the rest is still likely to be beyond the reach of human intuition.

As more thorough and intelligent exploration of the chemical space is required, the recent development of deep generation models has increasingly stimulated the use of models in in silico molecular design [10, 11]. All related studies have a common goal of reducing the effort of actually discovering new molecules [12-31]. Despite the probabilities described above, not much attention has been paid to scaffold-based molecular design to this day. One exception is the work of Li et al., [25], where the authors use a fingerprint representation of the scaffold embedded in the molecule. The fingerprint is then used to control the production process so that the resulting molecule tends to contain the desired scaffold. Another possible way to maintain scaffolds is to learn sequential representations of molecules [15,17,24,32]. Searching the vicinity of a query molecule in the generated latent space can generate molecules with similar shapes. Therefore, when the distance in the learned space is closely related to the core structure similarity, a derivative of the scaffold can be generated.

The described scheme of scaffold-based molecular generation has the following disadvantages. First, a predefined vocabulary of scaffolds is needed to categorically represent molecules of the scaffold type. This places constraints on the querying of any kind of scaffold when generating molecules. Second, due to its stochastic nature, conditional generation and latent spatial search allow molecules that do not contain the desired scaffold. Finally, to the best of our knowledge, there have been no studies that simultaneously control the scaffold and properties of the resulting molecule, which are essential to achieve the goal of scaffold-based molecular design.

The related studies are as follows.

To define a generative model for a molecule, it is necessary to select a molecular representation suitable for feature learning and molecular generation [37]. One of the most widely used specifications is SMILES, which uses strings of letters to represent 2D structures and stereoisomers of molecules. Therefore, the SMILES-based generative model is a language model [12-14], VAE [15-17, 32], adversarial autoencoders [18] and generative adversarial networks [19], including It was developed under various frameworks. Some models have used reinforcement learning strategies to increase goal-directed generation [19-23].

Despite the well-proven success of SMILES-based models, SMILES has fundamental limitations in delivering molecular similarities consistently [24]. In contrast to SMILES, molecular graphs as representations can more naturally represent the structural properties of molecules. Since learning a distribution over a graph introduces more tricky problems such as low scalability and non-unique node ordering [25,38], it is not until recently that generative models of graphs, In particular, reports on molecular graphs began to appear. Most studies employ a sequential method in generating graphs. See, for example, Li et al., [25] You et al. [26] and Li et al. The model in [27] creates a graph by predicting a series of graphing actions consisting of node addition and edge addition. Assouel et al. [28] and Liu et al. [29] similarly adopted a sequential scheme, but their graph generation involves two (fully [28] or partially [29]) separate steps: node initialization and concatenation. proceeds through Standard node-by-node generation can also be coarse-grained in a structure-by-structure manner. Jin et al. The junction tree VAE (junction tree VAE) of VAE creates a tree of molecular fragments and then recovers the entire molecular graph through fine-grained assembly of fragments [24]. A sharp contrast to the sequential generation scheme is the single-step generation of the entire graph. The GraphVAE model by Simonovsky and Komodakis [30] and MolGAN by De Cao and Kipf [31] generate graphs at a time by predicting both adjacency matrix and graph attributes.

The present inventors made intensive research efforts to develop a method for generating new molecules that can simultaneously control two or more desired properties and can efficiently generate new molecules only by learning from limited data. As a result, the present invention was completed by accepting the scaffold as an input and finding out that it is possible to synthesize a molecule exhibiting the desired performance by sequentially adding atoms and bonds.

Accordingly, it is an object of the present invention to provide a method for generating an extended molecular structure having a desired scaffold structure.

According to one aspect of the present invention, the present invention provides a method for generating an extended molecular structure having a predetermined scaffold structure comprising the steps of:

(a) learning a scaffold-based molecular generation model comprising: and

를 통해 전체-분자 그래프 G 를 입력 받아, G 를 잠재 벡터 z로 인코딩하도록 인코더

를 학습시키는 단계; (a-1) Encoder

Encoder to take full-molecular graph G as input and encode G as a latent vector z

learning;

를 통해 스캐폴드 그래프 S 를 추가 입력 받고, S 에 대해 순차적으로 노드(node) 및 에지(edge)를 추가함으로써, z로부터 G 를 회복시키도록 디코더

를 학습시키는 단계;(a-2) decoder

A decoder to recover G from z by receiving an additional input of the scaffold graph S and sequentially adding nodes and edges to S

learning;

(b) A target molecule generation step, in which the expanded molecular structure is obtained by inputting the scaffold graph into the learned scaffold-based molecular generation model.

The present inventors made intensive research efforts to develop a method for generating new molecules that can simultaneously control two or more desired properties and can efficiently generate new molecules only by learning from limited data. As a result, it was confirmed that the synthesis of molecules exhibiting the desired performance is possible by accepting the scaffold as an input and sequentially adding atoms and bonds.

In one embodiment of the present invention, the encoder of the present invention comprises an algorithm consisting of:

As the propagation step, H' _V(G) = propagate (H _V(G) , H _E(G) ); and

As the read phase, h _G = readout (H' _V(G) ).

In one embodiment of the present invention, the propagation step of the present invention consists of the following steps:

(a) a first propagation step of calculating an aggregated message between each node and its adjacent node by Equation 1 below; and

[Equation 1]

[Equation 2]

In one embodiment of the present invention, the propagation step of the present invention calculates an aggregated message including an additional vector c to adjust graph propagation.

는 다음 단계를 포함하는 디코딩 프로세스에 의해 z로부터 G 를 회복시킨다: In one embodiment of the present invention, the decoder of the present invention

recovers G from z by a decoding process comprising the following steps:

(a) (i) selecting the atom type to be added or (ii) terminating the building process, node addition;

(b) (i) selecting a binding type for the newly added node in step (a) or (ii) returning to step (a), an edge addition step; and

(c) select node v from existing nodes excluding w, then add the join type selected in step (b) and additional edges (v,w) and return to step (b), node selection (node selection) step.

In an embodiment of the present invention, the probability vector for determining the next action in steps (a) to (c) of the present invention is calculated by any one of the following equations 3 to 5:

[Equation 3]

[Equation 4]

[Equation 5]

The present invention shows the development of generative models targeted for use in scaffold-based molecular design. Our model is a variational autoencode (VAE) that takes a molecular scaffold as input and extends it to generate derivative molecules [33]. Scaffold expansion is accomplished by sequentially adding new atoms and bonds to a portion of the scaffold. The present invention uses a graph representing the molecular structure instead of a widely used text format such as the simplified molecular-input line-entry system (SMILES) [34]. This allows the graph to more naturally represent how a new atom or bond, i.e., a node or edge, is added each time, while the string representation can change substantially throughout the sequential expansion process of the present invention. Because there is [24]. By making every decision to add a new element according to the structure of the graph being processed, graph neural networks are used to extract structural dependencies between nodes and edges [25,35,36]. The properties of the resulting molecule can be controlled by adjusting the production process to the desired properties. Through learning, the model of the present invention can accommodate any arbitrary scaffold and generate new molecules while controlling its properties and maintaining its core structure.

Ideally, a generative model without any starting structures, ie a de novo generative model for the molecule, would be more desirable. However, the performance of optimizing properties in de novo generation can be significantly degraded when the related structure-property relationships are difficult to learn [18]. This tendency can be exacerbated when multiple properties have to be controlled, since obtaining data for a sufficient amount of individual properties is often impractical. By utilizing the properties of the scaffold, the scaffold-based generative model of the present invention can more easily optimize target properties and preserve desirable properties. As an example of the latter, it is possible to maintain the synthetic accessibility of the generated molecules by using a scaffold with well-established synthetic pathways, which is a crucial aspect in in silico design.

The study of the present invention contributes to molecular design in two aspects. One of the more practically important is to propose a model as a tool for scaffold-based molecular design. Another of more analytically important is the supergraph space search of the present invention. To specify the latter, we first noted that once the scaffold is maintained, the set of possible generations includes only molecules whose graph is a supergraph of the scaffold graph. In other words, since fixing the scaffold imposes strong constraints on the search space, it is meaningful to investigate how such constraints affect generation performance and property controllability. The first part of our experiment addresses this point. Further divided into three parts, the experiments of the present invention address (i) overall generation performance, (ii) scaffold dependence and (iii) property controllability of the model of the present invention. In part (i), it is shown that the model of the present invention achieves a high rate of validity, originality and novelty in molecular generation. In part (ii), our model shows consistently good performance in extending observed as well as unobserved scaffolds, and our model memorizes molecular-scaffold scaffold matching from training data. Instead, it shows that new molecules can be made from smaller molecules by learning valid chemical rules. In part (iii), the present invention shows that, despite the limited search space, the model of the present invention can control single or multiple properties of the generated molecule to an acceptable deviation from the target value.

Finally, returning to a more practical aspect, the present invention applied a model to design inhibitors of human epidermal growth factor receptor (EGFR). We performed semi-supervised learning, including a property predictor, to also learn from unlabeled molecules, which helps to deal with the frequently occurring problem of data shortage. can give As a result, this model was able to generate new inhibitors with improved potency, where 20% of the ingenious production showed a 100-fold or greater reduction in the predicted half-maximal inhibitory concentration. This shows that our model can learn to design molecules in more realistic problems that are difficult to learn due to the lack of available data.

(a) The present invention provides a method for generating an extended molecular structure having a desired scaffold structure.

(b) When the method of the present invention is used, it is possible to efficiently derive a molecular structure with an optimized desired performance using only a small amount of labeled data.

1 shows the model architecture of the present invention in the learning phase.
Figure 2 shows the distribution of properties of the resulting molecules. The values in the legend represent the target attribute values of the creation operation. The red line in each plot shows the distribution of each characteristic of the molecule in the training dataset.
3 shows exemplary molecules generated from three scaffolds. The indicated values represent the target conditions of generation and characteristic values of the scaffold.
4 shows the predicted co-distribution of characteristic values of the resulting molecules. The legend indicates the target values used for creation. In all distributions, the innermost contour surrounds 10%, the outermost contour surrounds 90%, and each nth in the middle surrounds n × 10% of the population. At the top and right ends of each plot are marginal distributions of features on the horizontal and vertical axes, respectively.
5 shows a scatter plot of characteristic values of the generated molecules. The legend lists the eight sets of property values used for creation.
6 shows the distribution of predicted human-EGFR pIC50 values for test scaffolds and generated molecules in a semi-supervised learning experiment. We extended our model into semi-supervised VAEs by adding feature predictors.

Hereinafter, the present invention will be described in more detail through examples. These examples are only for illustrating the present invention in more detail, and it will be apparent to those of ordinary skill in the art that the scope of the present invention is not limited by these examples according to the gist of the present invention. .

Example

Experimental method

Example 1: Overall process and model structure

It is an object of the present invention to generate molecules with target properties (property properties) while maintaining a given scaffold as a substructure. To this end, the generative model of the present invention was set up to accept the graph representation S of the molecular scaffold and generate a graph G, which is a supergraph of S. The underlying distribution of G can be expressed as p(G;S). Here, the notation of the present invention indicates a specific relationship between G and S, that is, a supergraph-subgraph relationship. The present invention also emphasizes that p(G;S) is a distribution of G alone; S acts as a parametric argument that explicitly defines the domain of the distribution. A molecular property is introduced as a condition, in which the model can define a conditional distribution p(G;S|y , y _S ), where y and y _S are the properties of the molecule and its scaffold, respectively. A vector containing property values. In other studies of molecular generation [25,27], a conditional distribution p(G|S) is defined, as often a substructural moiety is imposed as a condition. Unlike p(G;S), in the conditional distribution p(G|S), the space of G is not explicitly bounded by S, and due to its probabilistic nature, G which is not a supergraph of S is also allowed. On the other hand, the graph generated by the model of the present invention according to p(G;S) always includes S as a substructure. As can be seen below, the model of the present invention achieves this by sequentially adding nodes and edges to S. Before proceeding, the present invention recommends reference to Table 1 for the following notation.

Also, when a clear distinction is required, the present invention will refer to the "whole molecule" to distinguish the molecule from the scaffold.

The training objective of our model is a strategy to extend the graph to a larger graph with a distribution that follows the distribution of real molecules. The present invention achieves this by training our model to recover the molecules of the data set from its scaffold. Molecular scaffolds can be defined in a deterministic way as by Bemis and Murcko [39], which we used in our experiments.

The construction of the target graph is performed by successively determining node and edge additions. Decisions at each construction step are derived from node features and edge features of the graph at the step. Node features and edge features are iteratively updated to reflect the configuration history of previous steps. The construction process will be described in more detail below.

와 디코더(decoder)

로 구성되며, 각각

와 θ로 매개 변수화된다(parametrized). 인코더는 그래프 G를 인코딩 벡터 z에 인코딩하고 디코더는 z를 디코딩하여 G를 복구한다. 디코더는 추가적인 입력으로서 스캐폴드 그래프 S 를 필요로 하고, 실제 디코딩 과정은 노드 및 에지를 S 에 순차적으로 추가함으로써 실행된다. 처리되고 있는 일시적 그래프(transient graph)의 노드 및 에지 특징을 업데이트할 때 인코딩 벡터 z는 이의 정보에 지속적으로 영향을 미침으로써 그 역할을 수행한다. p(G;S)와 유사하게, 본 발명의 표기법

(G;S|z)는 우리 디코더의 후보 생성이 항상 S의 수퍼그래프임을 나타낸다. 인코더 표기법

(z|G;S)에 대해서는, 본 발명은 인코더가 또한

와

의 공동 최적화로 인해 스캐폴드에 의존도를 갖는다고 강조한다.The model of the present invention was realized with a VAE having the architecture shown in FIG. 1 . The architecture is an encoder

and decoder

consists of, each

and θ are parameterized. The encoder encodes the graph G into an encoding vector z and the decoder decodes z to recover G. The decoder requires a scaffold graph S as an additional input, and the actual decoding process is executed by sequentially adding nodes and edges to S . When updating the node and edge features of the transient graph being processed, the encoding vector z does its job by continuously influencing its information. Similarly to p(G;S), the notation of the present invention

(G;S| z ) indicates that our decoder's candidate generation is always a supergraph of S. encoder notation

For ( z |G;S), the present invention is that the encoder also

Wow

We emphasize that it is scaffold dependent due to the co-optimization of

As a latent variable model, VAE can be difficult to generate meaningful samples when the learned latent distribution of a manifold is insufficient to cover all information domains [40-42]. In generative models of molecules, this difficulty manifests itself in the form of chemically ineffectively generated molecules [15,43], and thus various SMILES-based [43-46] and graph-based [24,29,47- 49] models introduced explicit constraints or learning algorithms to improve the validity of their generation. The model of the present invention may include similar constraints or algorithms; Despite this absence, our model shows high chemical effectiveness in the resulting graph (see result 2 below).

Example 2: Graph Encoding

The goal of graph encoding is to generate a latent vector z of the entire graph G of the entire molecule. In the graph G = (V(G), E(G)) of any given whole molecule, the present invention first associates each node ν ∈ V(G) with a node feature vector h _ν , and each edge (u, ν) ∈ E(G) is associated with the edge feature vector h _{uν .} For initial node and edge features, the present invention selects the atomic type and bond type of the molecule. The present invention then embeds the initial feature vector into a new vector of a larger dimension so that the vector has sufficient capacity to represent deep information within and between nodes and edges. In order to fully encode the structural information of a molecule, the present invention wants every node embedding vector h _v to contain not only unique information of its own node v, but also the relationship of v and its neighborhood (its neighborhood). This can be done by propagating the information of each node to other nodes in the graph. As graph neural networks are specifically realized individually, various related methods have been devised [35].

를 구현하였다[36,50]. 본 발명의 네트워크의 알고리즘은 전파 단계(propagation phase)와 판독 단계(readout phase)로 구성되며, 다음과 같이 작성된다.In this study, the present invention is an encoder as a variant of an interactive network.

has been implemented [36,50]. The algorithm of the network of the present invention consists of a propagation phase and a readout phase, and is written as follows.

H' _V(G) = propagate ( H _V(G) , H _E(G) ) (1)

h _G = readout ( H' _V(G) ). (2)

The propagation phase itself consists of two phases. The first step computes the aggregated message between each node and its neighbors as

(3)

(3)

We use the message function M . The second step is to update the node vector using the aggregated message as follows:

(4)

(4)

Use the update function U . Updating all of the nodes in the feature vector _V H _(G) resulting in an updated set H _{'V (G),} as written in equation (1). Repeat the propagation step a fixed number of times each time it is applied, using different parameter sets in different iteration steps. After propagation, the read step (Equation (2)) computes the weighted sum of the node feature vectors, producing a single vector representation h _{G that summarizes the entire graph.} Then finally, the latent vector z is sampled from a normal distribution whose mean and variance are estimated from h _{G .}

The graph propagation can be adjusted by including an additional vector c to compute the aggregated message. In this case, the function M and (hence) propagte take c as additional arguments (i.e. they become M (·,·,·, c ) and propagate (·,·, c )). . When encoding the input graph, we choose c to be the concatenation of the property vectors y and y _{S to enable property-controlled generation.} Uses a graph for decoding (decoding), y, _S y and the potential of the connection vector z (concatenation) as a vector condition (condition vector) (see below).

Example 3: Graph Decoding

The goal of graph decoding is to reconstruct the graph G of the entire molecule from the latent vector z sampled in the graph encoding step. The graph decoding process of the present invention is described in Li et al. It is motivated by the sequential generation strategy of [25]. In the study of the present invention, the present invention is _{the entire molecular graph G by successively adding nodes and ashes from the scaffold graph G 0} (extracted from G by the Bemis-Murcko method [39] using RDKit software [51]). to build Here, G ₀ =S represents the initial scaffold graph, and we will write _{G t} to represent the transient (or completed) graph constructed from _{G 0 .}

The graph decoding of the present invention begins with preparing and propagating the initial node function _{of G 0 .} As with G, an initial feature vector of _{G 0} is prepared by embedding the atomic type and bond type of the scaffold molecule. The initial embedded (embedding) is performed by (embedded search (embed) in Figure 1), the same network used for the whole molecule. Then, _{the initial feature vector of G 0} is propagated a fixed number of times by another interaction network. Once propagation is complete, the decoder expands _{G 0} by processing through a loop of node additions and a subsequent loop of edge additions (inside). A detailed description of the process follows:

· Step 1: node addition. Select an atom type or terminate a building process with an estimated probability. If an atom type is selected, add a new node w with the type currently selected as the transient graph G _{t and proceed to step 2.} Otherwise, terminate the building process and return the graph.

· Step 2: edge addition. At a given new node, choose a join type or return to step 1 with an estimated probability. If the binding type is selected, proceed to step 3.

· Step 3: Node selection. With the estimated probability, a node ν is selected from existing nodes excluding w. Then, we add a new edge (v, w) to _{G t} with the type of bonding selected in step 2. Continue adding edges in step 2.

The flow of the entire process is shown on the right side of FIG. 1 . Excluded from steps 1-3 is the final step of selecting the appropriate isomer, which is described separately below.

In all steps, the model determines the next action by estimating a probability vector for a candidate action. Depending on whether the current step needs to add atoms (step 1), edges (step 2), or select atoms to connect to (step 3), the probability vector is computed by one of the following formulas:

은(na + 1)-길이 벡터이고, 여기서 엘리먼트(element)

내지

는 원자 유형에 대한 확률에 해당하고

은 종료 확률(termination probability)이다. 크기가 nb + 1 인 벡터인

는

내지

엘리먼트가 n_b 결합 타입에 대한 확률에 해당하며,

은 에지(edge) 추가를 중지할 확률이다. 마지막으로, 세번째 벡터

의 i 번째 엘리먼트는 i 번째 존재하는 노드와 마지막으로 추가된 노드를 연결할 확률이다.first probability vector

is (na + 1)-length vector, where element

inside

is the probability for the atomic type and

is the termination probability. a vector of size nb + 1

Is

inside

element corresponds to the probability for _{n b associative types,}

is the probability of stopping adding an edge. Finally, the third vector

The i-th element of is the probability of connecting the i-th existing node and the last added node.

가

에 추가되어야 한다. 이를 위해, 본 발명의 모델은 w의 원자 타입을 제시함으로써 초기 특성 벡터(initial feature vector)

를 준비한 후, 그것을

에서의 존재하는 노드 특성들과 통합하여, 적절한

를 계산한다. 유사하게, (v,w)라 불리는 새로운 에지(edge)가 추가될 때, 본 발명의 모델은

를

및

로부터 계산하여,

가 (v,w)의 결합 타입을 보여주는

를 업데이트한다. 새로운 노드 및 에지(edge)를 초기화하는 대응되는 모듈은 다음과 같다:When the model of the present invention decides to add a new node equal to w, the corresponding feature vector

go

should be added to To this end, the model of the present invention provides an initial feature vector by presenting the atomic type of w.

After preparing it

Integrating with the existing node properties in

to calculate Similarly, when a new edge called (v,w) is added, the model of the present invention is

cast

and

calculated from

showing the type of association of (v,w)

update The corresponding modules that initialize new nodes and edges are:

The graph building modules addNode , addEdge and selectNode include the preceding steps of propagating node properties. For example, the actual work performed by addNode is:

는 함수 합성(function composition)을 나타낸다. 우변에 따르면, 상기 모듈은 k 번의 그래프 전파(propagation)을 통해 노드 특성 벡터를 업데이트 한 뒤, 판독 벡터(readout vector)를 계산하고, 그 뒤 이를 z와 연결하고(concatenates), 최종적으로 다층 퍼셉트론 f를 통해

을 출력한다. 마찬가지로, addEdge 및 selectNode 모두는 propagete의 반복되는 적용으로 시작한다. 이러한 방식으로, 노드 특성들은 과도 그래프(transient graph)가 evolve 될 때마다 반복적으로 업데이트되고, 모든 구축 이벤트(building event)의 예측은 선행 이벤트(preceding event)의 이력(history)에 의존하게 된다. from above

denotes function composition. According to the right side, the module updates the node feature vector through k graph propagation, calculates a readout vector, then concatenates it with z, and finally multilayer perceptron f Through the

to output Similarly, both addEdge and selectNode start with repeated application of propagete . In this way, the node properties are iteratively updated whenever the transient graph evolves, and the prediction of all building events depends on the history of the preceding event.

As shown in equation (10), the propagation of the graph within addNode (and addEdge and selectNode ) incorporates a latent vector z, which encodes a full-molecular graph G . This makes the model of the present invention refer to z when making graph building decisions, and ultimately rebuilds G by decoding z. If our model is to be conditioned on the whole-molecular property y and the scaffold property y _s , then equations (5)-(7) and (10) incorporate =concat(z,y,y _s ) instead of z is understood to be

Example 4: Molecular Generation

를 생성한다. 지정된 분자 특성을 갖는 분자를 생성하는 것을 원하는 경우, 상응하는 특성 벡터 y 및 y_s는 구축 프로세스(building process)의 조건에 제공되어야 한다. When generating a new molecule, a scaffold S is required as input, and a latent vector z is sampled from a standard normal distribution. Then, the decoder creates a new molecular graph as a supergraph of S

to create If it is desired to generate molecules with specified molecular properties, the corresponding property vectors y and y _s must be provided to the conditions of the building process.

Example 5: Selection of isomers

가 구축된 후, 원자들 및 결합들의 입체화학적 배열이 결정된다[24]. 이성질체 선택 모듈 selectIsomer는 RDKit에 의해 열거된 모든 가능성있는 입체이성질체의 그래프 I를 만들어내고, 이들의 입체화학적 표지(stereochemical labels)가 없는 2D 구조들은

의 그것들과 동일하다. 모든 생성된 I는 노드 및 에지 특성들에서의 원자들 및 결합들의 입체화학적 배열을 포함한다. 그런 뒤, 모듈은 선택 확률을 다음과 같이 추정한다.Molecules may have stereoisomers. Stereoisomers of molecules have identical linkages between atoms, but different 3D geometries. Consequently, the complete creation of a molecule must also specify the stereoisomeric properties of the molecule. Molecular graph from z

After , the stereochemical arrangement of atoms and bonds is determined [24]. The isomer selection module selectIsomer produces a graph I of all possible stereoisomers enumerated by RDKit, and the 2D structures without their stereochemical labels are

are identical to those of Every generated I contains the stereochemical arrangement of atoms and bonds in node and edge properties. Then, the module estimates the selection probability as

의 엘리먼트들은 각각의 I를 선택하는 추정된 확률이다. here, vector

The elements of h are the estimated probability of choosing each I.

Example 6: objective function

The objective function of the present invention has the form of a log-likelihood of a general VAE [33]:

에 대해 상응하는 전체-분자 데이터세트 D(S)가 있다. 분자의 세트는 스캐폴드 세트 및 전체-분자 세트의 콜렉션을 생성할 수 있다: 미리 주어진 세트 내의 모든 분자의 스캐폴드가 정의되면, S를 생성하고, 미리 주어진 세트의 상기 분자들은 상기 콜렉션

내로 그룹핑될 수 있다. 이러한 S 및 D(S)를 사용하여, 본 발명의 목적은 식(12)의 우변을 최대화시키고, 이에 의해

를 최대화시키는, 파라미터

및 θ의 최적의 값을 찾는 것이다. 여기에서, 이중 기대(double expectation)는 스캐폴드 S에 대한 전체-분자 세트 D(S)의 의존성을 명시적으로 나타낸다. 즉, 본 발명의 정의에 따르면, 스캐폴드 세트 S를 정의하는 것이 첫 번째이고, 그 후, 각각의

는 전체-분자 세트 D(S)를 정의한다. ESI에서, 전체 프로세스의 알고리즘과 함께 모듈의 구현 및 정확한 작동에 대해 자세히 설명한다. Here, D _KL [····] is the Kullback-Leibler divergence, and p (z) is the prior distribution defined as standard normalized in the model of the present invention. The maximization of the first term on the right side maximizes the probability that the decoder p _θ of the present invention recovers the graph G from their latent representations z, and the maximization of the second term indicates that the encoder q _Φ of the present invention is not in the prior art. Try to get as close as possible. As mentioned in Example 1, the additional argument S in equation (12) represents the explicit constraint imposed by the model of the invention in which G is a supergraph of S. In actual training, there is a scaffold dataset S , and each scaffold

There is a corresponding full-molecular dataset D ( S ) for A set of molecules can produce a set of scaffolds and a collection of full-molecular sets: once the scaffolds of all molecules in a given set are defined, it produces S , and said molecules of a given set are said collection.

can be grouped into Using these S and D ( S ), it is an object of the present invention to maximize the right-hand side of equation (12), thereby

parameter that maximizes

and to find an optimal value of θ. Here, the double expectation explicitly indicates the dependence of the whole-molecule set D(S) on the scaffold S. That is, according to the definition of the present invention, it is first to define a scaffold set S, and then, each

defines the whole-molecular set D ( S ). In ESI, the implementation and correct operation of the module are described in detail along with the algorithm of the whole process.

Results and discussion

1. Datasets and Experiments

The dataset of the present invention is synthetic screening compounds (version March 2018) provided by InterBioScreen Ltd [52]. The data set (and hence the IBS data set) initially contained SMILES strings of organic compounds composed of H, C, N, O, F, P, S, Cl and Br atoms. We filter out strings that contain unconnected ions or fragments and those that cannot be read with RDKit. The preprocessing process of the present invention yielded 349,809 trained molecules and 116,603 test molecules. The average number of heavy atoms was 27, the maximum was 132, and the average molecular weight was 389 g/mol. The number of scaffold types in the training set was 85,318 and 42,751 in the test set. Experiments of the present invention involve training and evaluation of a scaffold-based graph generation model using the described data set. For conditional molecular generation, molecular weight (MW), topological polar surface area (TPSA) and octanol-water partition coefficient (log P ) were used. Models were conditioned either alone or jointly using one, two, or all of the three attributes. The learning rate was set to 0.0001, and all instances of the model were trained up to 20 epochs. Other hyper-parameters such as layer dimensions are specified in the ESI. Molecular properties were calculated using RDKit. In the following, the units of MW (g/mol) and TPSA (Å ² ) will be omitted for the sake of brevity.

2. Analysis of effectiveness, originality and novelty

The validity, originality and novelty of the generated molecule are basic evaluation metrics of the molecular generation model. For the precise meaning of the three metrics, the following definition follows [53]:

Validity = # of valid graphs / # of generated graphs

Uniqueness = non-overlapping, # of valid graphs / # of valid graphs

Novelty = # of eigengraph not in training set / # of eigengraph

A valid graph was defined when basic chemical requirements such as valency were met. We actually used RDKit to determine the validity of the generated graph. It is particularly important to identify the matrix in the model of the present invention, as molecules generated from the scaffold limit the space of candidate products. Models conditioned alone for MW, TPSA or log P were evaluated by randomly selecting 100 scaffolds from the dataset and generating 100 molecules from each scaffold. Target values (100 values for each feature) were randomly sampled from each feature distribution on the dataset. For the MW, it is unnatural for the MW to produce a molecule smaller than the MW of the scaffold, so such cases were excluded from the evaluation.

Table 2 summarizes the validity, originality and novelty of molecules generated from the model of the present invention, and the results of other molecular generation models for comparison.

Since values by different models were collected from their respective reports, the comparisons are approximate. Despite the strict limitations imposed by the scaffold, our model showed high validity, originality and novelty compared to the results of other models. The high originality and novelty were particularly significant, given that most of the scaffolds in the training set of the present invention have only very few whole-molecules. For example, out of 85,318 scaffolds in the training set, 79,700 scaffolds have less than 10 total-molecules. Therefore, it is unlikely that the model of the present invention achieved such high performance simply by memorizing the training set, and it can be concluded that the model of the present invention learns general chemical rules to extend any scaffold.

3. single-property control

For the next analysis, it was tested whether the scaffold-based graphing model of the present invention could generate molecules with the desired scaffold and desired properties simultaneously.

Although several molecular generation models have been developed to control the molecular properties of the generated molecules, it will be more difficult to control the molecular properties under the limiting conditions imposed by a given scaffold. The present inventors set target values of 80, 100, MW of 120, TPSA of 300, 350 and 400, and log P of 5, 6 and 7. For all 9 cases, we used the same 100 scaffolds used for results in Results 2, generating 100 molecules for each scaffold.

Figure 2 shows the distribution of properties of the resulting molecule. It is observed that the feature distributions are well concentrated around the burst value. This shows that, despite a narrowed search space, the model of the present invention successfully generated new molecules with desired properties. To see how our model expands given scaffolds according to specified property values, some of the molecules generated in Fig. 3 are pictorially shown. For target conditions with MW=400, TPSA=120, and log P=7, 9 examples were randomly sampled using 3 different scaffolds.

Our model created new molecules with specified properties by adding appropriate side chains: for example, our model added hydrophobic groups to the scaffold to create high-log P molecules , while polar functional groups were added to generate high-TPSA molecules.

4. Scaffold Dependencies

The molecular design process of the present invention starts with a given scaffold with successively added nodes and edges. Thus, the performance of the model of the present invention may be affected by the type of scaffold. Therefore, given a new scaffold, it was tested whether the model of the present invention retains its ability to generate desirable molecules. Specifically, in the present invention, a set of 100 new scaffolds not included in the training set (“unseen” scaffold) and an additional 100 scaffold set from the training set (hereinafter “seen” scaffold) were prepared. Then, 100 molecules were generated for each scaffold with randomly assigned property values. This process was repeated for MW, TPSA and log P.

Table 3 summarizes the effectiveness, originality and MAD of molecules generated from seen and unseen scaffolds.

Here, MAD represents the mean absolute difference between the specified property values and the property of the generated molecule. The results show that there is no significant difference in the three matrices between the two sets of scaffolds. This shows that the model of the present invention is generally applicable to any scaffold in generating effective molecules with controlled properties.

5. Multi-characteristic control

New molecular designs rarely require only controlling one specific molecular property. Above all, drug design particularly requires simultaneous control of multiple molecular properties. In view of this point, the performance of simultaneously controlling any two of W, TPSA and log P of the model of the present invention was first tested. We trained three instances of a model jointly controlling MW and TPSA, MW and log P , and log P and TPSA. Then, we specified each characteristic with two target values (MW of 350 and 450, TPSA of 50 and 100, and log P of 2 and 5), and combined them to prepare four production conditions for each pair did Under all production conditions, 100 randomly sampled scaffolds used in Results 2 and 3 were used, and 100 molecules were generated from each scaffold. Products with a target MW smaller than the MW of the scaffold used were excluded.

4 shows the results of the products conditioned with MW and TPSA, MW and log P , and log P and TPSA. The co-distribution of property values for the resulting molecules is indicated. Gaussian kernels were used to estimate the kernel density. It was observed that the modes of distribution are well located near the points of the target values. Exceptionally, the distribution by target (log P, TPSA)=(2,50) showed a relatively longer tail compared to the larger log P and TPSA values. This is because log P and TPSA are negatively correlated with each other by definition, and requiring small values for both can be an unphysical production task. Intrinsic correlations between molecular properties can result in even plausible targets exhibiting a dispersed property distribution. Such an example could be the result of another target (log P, TPSA)-(5, 50), but in most cases only a very small fraction of the total production.

Additionally, all three properties were incorporated to test the generation of conditioning. As described above, the same target values of W, TPSA and log P were used, and a total of 8 conditions were obtained. The remaining settings, including the number of scaffold sets and generations, were maintained. The results are shown in FIG. 5, and W, TPSA and log P values of the generated molecules are indicated. This plot shows that the distributions from different target conditions are well separated from each other. As with the double-characteristic results, all distributions are well-densely distributed around the target value.

We also quantitatively compared the generation performance of models for single- and multi-attribute control. Table 4 shows the performance statistics of single-, double- and triple-characteristic controls in terms of MAD, effectiveness and novelty.

The same 100 scaffolds were used to generate 100 molecules each time under randomly specified target conditions. The overall size of the descriptor is well preserved as the number of integrated features increases from one to two and three. The slight increase in the MAD value is due to the additional confinement of the chemical space enforced by the intrinsic correlations between multiple properties. Nevertheless, the magnitude of the deterioration is small compared to the average values of the characteristics (MW 389, TPSA 77, log P 3.6).

6. Design of EGFR inhibitors by semi-supervised learning

Deep learning methods often require millions of labeled data to fully unleash their performance. However, many practical applications suffer from lack of data. For example, in inhibitor design, the amount of binding affinity available for a target protein is only in the order of a few thousand. One possible approach to this case is semi-supervised learning, which incorporates large amounts of unlabeled data into small amounts of labeled data for training. In fact, there has been a previous report that semi-supervised learning for conditioned molecular design is efficient when only a small fraction of molecules in the dataset have characteristic values [16]. In order to confirm the applicability of the scaffold-based molecular graph generation model of the present invention in a similar situation, a system of semi-supervised learning was provided in the model of the present invention, and an inhibitor of human EGFR protein, which has a limited amount of data We tested how well we could design.

We applied a semi-supervised VAE [16,54] because it can be easily implemented by adding a label predictor to the normal VAE. In semi-supervised VAE, predictors were trained together to predict negative logarithm of half maximal inhibitory concentration (pIC50) values for human EGFR, along with the VAE portion.

_{Predictors were trained using the actual pIC 50} values for the labeled molecules and also used them in the VAE condition vector. For unlabeled molecules, we trained only the VAE portion using the _{pIC 50 values predicted by the predictors.} For the IBS training molecules previously used by the present inventors, 8,016 molecules were added from the ChEMBL database [55] to prepare the present dataset for semi-supervised learning. ChEMBL molecules were those with pIC ₅₀ values and were also labeled, whereas IBS molecules were unlabeled. Each molecule was paired with the scaffold from which it was extracted, and all scaffolds of ChEMBL as well as IBS molecules were treated to remove the label. ChEMBL molecules were divided into 6,025 training molecules and 1,991 test molecules, resulting in only 1.7% label percentage in the training set. To effectively train predictors, labeled and unlabeled inputs with a ratio of 1:5 were sampled from every batch. After 20 epochs of training, 100 scaffolds from the ChEMBL test set with _{predicted pIC 50 values between 5 and 6 were selected, and 100 molecules with predicted pIC 50} values of 8 were generated from each scaffold.

_{The MAD of the pIC 50} prediction for the ChEMBL test molecules was 0.58. In a total of 10,000 creations, our model showed 96.6% effectiveness, 44.9% originality, and 99.7% novelty. The relatively low originality is expected because the use of one fixed target value imposes strict conditions on the search space, thereby increasing the redundancy of the creation. 6 shows the distribution of _{predicted pIC 50} values for 100 scaffolds and generated molecules. Although the distribution of the generated molecules was concentrated at values lower than 8 and showed a relatively wide variance, _{the average improvement of the pIC 50} value compared to the scaffold was 1.29 (only unique molecules were counted), which _It is about 19.7 times strengthening of the inhibitory ability in terms of 50. Of more interest in practice, those with a pIC ₅₀ expected to be greater than 8 belong to 20% of the original products. These results show that the model of the present invention can be applied with minimal extension to many practical situations where data preparation is a problem.

Through the present invention, a scaffold-based molecular graph generation model has been proposed. Our model creates new molecules from the desired substructure or scaffold by sequentially adding new atoms and bonds to the graph of the scaffold. In contrast to other related methods, such as, for example, stochastic conditioning of substructures during generation, the strategy of the present invention naturally guarantees expansion of the scaffold of generated molecules.

Because molecular generation from scaffolds can utilize already existing properties, it may be easier to optimize or maintain target properties than by de novo molecule generation. When pharmacological activity is aimed at being optimized, it can be difficult to generate an optimal molecular structure from scratch due to the complexity of the molecular structure. In such cases, the use of a scaffold with appropriate activity may make the optimization more feasible. There are similar advantages when controlling multiple properties at the same time. For example, if the goal is to secure synthetic accessibility and improve activity of the generated molecule, using a scaffold that can be synthesized well will focus the goal more on the latter, increasing search efficiency.

The model of the present invention was evaluated by testing the effectiveness, originality and novelty of the resulting molecules. Despite the constraints on the search space imposed by the scaffold, our model showed similar results compared to previous SMILES-based and graph-based molecular generation models. Our model performed consistently well in terms of three metrics, given a new scaffold that was not included in the training set. This means that our model achieved good generalization rather than memorizing the pairings between scaffolds and molecules in the training set. Moreover, while maintaining a given scaffold, our model successfully generates new molecules with the desired degree of molecular properties, such as molecular weight, topologically polar surface area and octanol-water partition coefficient. create with Property-controlled generation can incorporate multiple molecular properties simultaneously. Our model can incorporate semi-supervised learning in applications such as protein inhibitor design, which is one of the common situations where only small amounts of labeled data are available. From all the results presented in the present invention, it can be seen that the scaffold-based molecular graph generation model of the present invention provides a practical method for optimizing the functionality of molecules with a conserved core structure.

Reference

1 P. G. Polishchuk, T. I. Madzhidov and A. Varnek, J. Comput.- Aided Mol. Des., 2013, 27, 675-679.

2 S. Kim, PA Thiessen, EE Bolton, J. Chen, G. Fu, A. Gindulyte, L. Han, J. He, S. He, BA Shoemaker, J. Wang, B. Yu, J. Zhang and SH Bryant, Nucleic Acids Res., 2016, 44, D1202-D1213.

3 K. H. Bleicher, Y. W¨uthrich, G. Adam, T. Hoffmann and A. J. Sleight, Bioorg. Med. Chem. Lett., 2002, 12, 3073-3076.

4 G. L. Card, L. Blasdel, B. P. England, C. Zhang, Y. Suzuki, S. Gillette, D. Fong, P. N. Ibrahim, D. R. Artis, G. Bollag, M. V. Milburn, S.-H. Kim, J. Schlessinger and K. Y. J. Zhang, Nat. Biotechnol., 2005, 23, 201-207.

5 M. E. Welsch, S. A. Snyder and B. R. Stockwell, Curr. Opin. Chem. Biol., 2010, 14, 347-361.

6 Y. Im, M. Kim, Y. J. Cho, J.-A. Seo, K. S. Yook and J. Y. Lee, Chem. Mater., 2017, 29, 1946-1963.

7 J. Dhar, U. Salzner and S. Patil, J. Mater. Chem. C, 2017, 5, 7404-7430.

8 A. Al Mousawi, F. Dumur, P. Garra, J. Toufaily, T. Hamieh, B. Graff, D. Gigmes, J. P. Fouassier and J. Lalev´ee, Macromolecules, 2017, 50, 2747-2758.

9 X. Sun, F. Wu, C. Zhong, L. Zhu and Z. Li, Chem. Sci., 2019, 10, 6899-6907.

10 B. Sanchez-Lengeling and A. Aspuru-Guzik, Science, 2018, 361, 360-365.

11 H. Chen, O. Engkvist, Y. Wang, M. Olivecrona and T. Blaschke, Drug Discovery Today, 2018, 23, 1241-1250.

12 M. H. S. Segler, T. Kogej, C. Tyrchan and M. P. Waller, ACS Cent. Sci., 2018, 4, 120-131.

13 A. Gupta, A. T. M¨uller, B. J. H. Huisman, J. A. Fuchs, P. Schneider and G. Schneider, Mol. Inf., 2018, 37, 1700111.

14 E. J. Bjerrum and B. Sattarov, Biomolecules, 2018, 8(4), 131.

15 R. G´omez-Bombarelli, JN Wei, D. Duvenaud, JM Hern´andez-Lobato, B. S´anchez-Lengeling, D. Sheberla, J. Aguilera-Iparraguirre, TD Hirzel, RP Adams and A. Aspuru -Guzik, ACS Cent. Sci., 2018, 4, 268-276.

16 S. Kang and K. Cho, J. Chem. Inf. Model., 2019, 59, 43-52.

17 J. Lim, S. Ryu, J. W. Kim and W. Y. Kim, J. Cheminf., 2018, 10, 31.

18 D. Polykovskiy, A. Zhebrak, D. Vetrov, Y. Ivanenkov, V. Aladinskiy, P. Mamoshina, M. Bozdaganyan, A. Aliper, A. Zhavoronkov and A. Kadurin, Mol. Pharm., 2018, 15, 4398-4405.

19 G. Lima Guimaraes, B. Sanchez-Lengeling, C. Outeiral, P. L. Cunha Farias and A. Aspuru-Guzik, arXiv e-prints, arXiv:1705.10843, 2017.

20 N. Jaques, S. Gu, D. Bahdanau, JM Hern´andez-Lobato, RE Turner and D. Eck, Proceedings of the 34 ^th International Conference on Machine Learning, Sydney, Australia, 2017, pp. 1645-1654.

21 M. Olivecrona, T. Blaschke, O. Engkvist and H. Chen, J. Cheminf., 2017, 9, 48.

22 D. Neil, M. H. S. Segler, L. Guasch, M. Ahmed, D. Plumbley, M. Sellwood and N. Brown, 6th International Conference on Learning Representations, Workshop Track Proceedings, Vancouver, BC, Canada, 2018.

23 M. Popova, O. Isayev and A. Tropsha, Sci. Adv., 2018, 4, eaap7885.

24 W. Jin, R. Barzilay and T. Jaakkola, Proceedings of the 35 ^th International Conference on Machine Learning, Stockholmsm¨assan, Stockholm Sweden, 2018, pp. 2323-2332.

25 Y. Li, O. Vinyals, C. Dyer, R. Pascanu and P. Battaglia, 6 ^th International Conference on Learning Representations, Workshop Track Proceedings, Vancouver, BC, Canada, 2018.

26 J. You, B. Liu, Z. Ying, V. Pande and J. Leskovec, in Advances in Neural Information Processing Systems 31, ed. S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi and R. Garnett, Curran Associates, Inc., 2018, pp. 6410-6421.

27 Y. Li, L. Zhang and Z. Liu, J. Cheminf., 2018, 10, 33.

28 R. Assouel, M. Ahmed, M. H. Segler, A. Saffari and Y. Bengio, arXiv e-prints, arXiv:1811.09766, 2018.

29 Q. Liu, M. Allamanis, M. Brockschmidt and A. Gaunt, in Advances in Neural Information Processing Systems 31, ed. S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi and R. Garnett, Curran Associates, Inc., 2018, pp.7795-7804.

30 M. Simonovsky and N. Komodakis, Artificial Neural Networks and Machine Learning - ICANN 2018, Cham, Switzerland, 2018, pp. 412-422.

31 N. De Cao and T. Kipf, arXiv e-prints, arXiv:1805.11973, 2018.

32 M. Skalic, J. Jim´enez, D. Sabbadin and G. De Fabritiis, J. Chem. Inf. Model., 2019, 59, 1205- 1214.

33 D. P. Kingma and M. Welling, 2nd International Conference on Learning Representations, Banff, AB, Canada, 2014.

34 D. Weininger, J. Chem. Inf. Model., 1988, 28, 31-36.

35 Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang and P. S. Yu, arXiv e-prints, arXiv:1901.00596, 2019.

36 J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals and G. E. Dahl, Proceedings of the 34th International Conference on Machine Learning, Sydney, Australia, 2017, pp. 1263-1272.

37 D. Polykovskiy, A. Zhebrak, B. Sanchez-Lengeling, S. Golovanov, O. Tatanov, S. Belyaev, R. Kurbanov, A. Artamonov, V. Aladinskiy, M. Veselov, A. Kadurin, S. Nikolenko , A. Aspuru-Guzik and A. Zhavoronkov, arXiv e-prints, arXiv:1811.12823, 2018.

38 J. You, R. Ying, X. Ren, W. Hamilton and J. Leskovec, Proceedings of the 35th International Conference on Machine Learning, Stockholmsm¨assan, Stockholm Sweden, 2018, pp. 5708-5717.

39 G. W. Bemis and M. A. Murcko, J. Med. Chem., 1996, 39, 2887-2893.

40 A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow and B. Frey, 4th International Conference on Learning Representations, Workshop Track Proceedings, San Juan, Puerto Rico, 2016.

41 J. He, D. Spokoyny, G. Neubig and T. Berg-Kirkpatrick, 7 ^th International Conference on Learning Representations, Conference Track Proceedings, New Orleans, LA, USA, 2019.

42 I. Goodfellow, arXiv e-prints, arXiv:1701.00160, 2016.

43 R.-R. Griffiths and J. M. Hern?andez Lobato, Chem. Sci., 2020, DOI: 10.1039/C9SC04026A.

44 M. J. Kusner, B. Paige and J. M. Hern?andez-Lobato, Proceedings of the 34th International Conference on Machine Learning, Sydney, Australia, 2017, pp. 1945-1954.

45 H. Dai, Y. Tian, B. Dai, S. Skiena and L. Song, 6 ^th International Conference on Learning Representations, Conference Track Proceedings, Vancouver, BC, Canada, 2018.

46 O. Mahmood and J. M. Hernndez-Lobato, arXiv e-prints, arXiv:1905.09885, 2019.

47 T. Ma, J. Chen and C. Xiao, in Advances in Neural Information Processing Systems 31, ed. S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi and R. Garnett, Curran Associates, Inc., 2018, pp. 7113-7124.

48 B. Samanta, A. De, G. Jana, P. K. Chattaraj, N. Ganguly and M. Gomez-Rodriguez, arXiv e-prints, arXiv:1802.05283, 2018.

49 Y. Kwon, J. Yoo, Y.-S. Choi, W.-J. Son, D. Lee and S. Kang, J. Cheminf., 2019, 11, 70.

50 P. Battaglia, R. Pascanu, M. Lai, D. Jimenez Rezende and K. Kavukcuoglu, in Advances in Neural Information Processing Systems 29, ed. D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon and R. Garnett, Curran Associates, Inc., 2016, pp. 4502-4510.

51 RDKit: Open-Source Cheminformatics, http://www.rdkit.org.

52 InterBioScreen Ltd, http://www.ibscreen.com.

53 N. Brown, M. Fiscato, M. H. Segler and A. C. Vaucher, J. Chem. Inf. Model., 2019, 59, 1096-1108.

54 D. P. Kingma, S. Mohamed, D. Jimenez Rezende and M. Welling, in Advances in Neural Information Processing Systems 27, ed. Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence and K. Q. Weinberger, Curran Associates, Inc., 2014, pp. 3581-3589.

55 A. Gaulton, LJ Bellis, AP Bento, J. Chambers, M. Davies, A. Hersey, Y. Light, S. McGlinchey, D. Michalovich, B. Al-Lazikani and JP Overington, Nucleic Acids Res., 2011 , 40, D1100-D1107.

Claims (6)

A method for generating an extended molecular structure having a desired scaffold structure comprising the steps of:
(a) learning a scaffold-based molecular generation model comprising: and
(a-1) Encoder

Encoder to take full-molecular graph G as input and encode G as a latent vector z

learning;
(a-2) decoder

A decoder to recover G from z by receiving an additional input of the scaffold graph S and sequentially adding nodes and edges to S

learning;
(b) A target molecule generation step, in which the expanded molecular structure is obtained by inputting the scaffold graph into the learned scaffold-based molecular generation model. The method of claim 1,
wherein the encoder comprises an algorithm consisting of:
As the propagation step, H' _V(G) = propagate (H _V(G) , H _E(G) ); and
As the read phase, h _G = readout (H' _V(G) ). 3. The method of claim 2,
A method, characterized in that the propagating step consists of the following steps:
(a) a first propagation step of calculating an aggregated message between each node and its adjacent node by Equation 1 below; and
[Equation 1]

(b) a second propagation step of updating a node vector using the aggregated message by Equation 2 below;
[Equation 2]

. 4. The method of claim 3,
The propagation step comprises calculating an aggregated message including an additional vector c to adjust the graph propagation. The method of claim 1,
the decoder

recover G from z by a decoding process comprising the following steps:
(a) (i) selecting the atom type to be added or (ii) terminating the building process, node addition;
(b) (i) selecting a binding type for the newly added node in step (a) or (ii) returning to step (a), an edge addition step; and
(c) select node v from existing nodes excluding w, then add the join type selected in step (b) and additional edges (v, w) and return to step (b), node selection (node selection) step. 6. The method of claim 5,
A method, characterized in that the probability vector for determining the next action of steps (a) to (c) is calculated by any one of Equations 3 to 5 below:
[Equation 3]

[Equation 4]

[Equation 5]

.

Images