DynamicBayesianNetwork

Full name SMRUCC.genomics.Analysis.BNLearn.DBN.DynamicBayesianNetwork Assembly SMRUCC.genomics.Analysis.BNLearn Members 34

Dynamic Bayesian Network for gene regulatory network simulation.

This DBN implements a 2-slice temporal Bayesian network (2TBN) where:

Gene/operon expression at time t+1 depends on TF and metabolite states at time t
TF states are provided as evidence from the ODE solver (TF protein/RNA abundance)
Metabolite concentrations are provided as evidence from the ODE solver

The DBN supports two modes:

Topology-only mode: Uses RegulatoryLink topology to initialize CPTs based on biological heuristics (activator/inhibitor effects via noisy-OR/AND gates). No RNAseq data required.
Data-fitting mode: Uses RNAseq time-series data to learn CPT parameters, with the topology-based CPTs serving as a Dirichlet prior.

Coupling with metabolic network ODEs:

DBN -> ODEs: Predicted gene states are mapped to RNA transcript abundance change rates (expected transcription rate), which serve as transcription rate terms in the ODEs: dR/dt = k_synthesis rate - k_degradation R
ODEs -> DBN: Metabolite concentrations and TF abundances from the ODEs are discretized (Low/Medium/High) and used as evidence for DBN inference.

01 Syntax

SMRUCC.genomics.Analysis.BNLearn.DBN.DynamicBayesianNetwork

02 Methods

Name	Overloads	Summary
.ctor	2	Default constructor with default configuration
BuildFromTopology	1	Build the DBN structure from a list of RegulatoryLink objects.
ApplyParentLimit	1	父节点数量上限保护：对父节点数超过 Config.MaxParents 的节点做确定性截断。截断策略：优先保留调控靶标更少（特异性更强）的父节点，同度时按 ID 排序，保证结果可复现；ParentIds / RegulatorTFs / TFEffectors 三者同步裁剪，避免状态不一致。默认上限 20 只用于隔离极端异常拓扑（正常情况下不…
BuildActivationModels	1	为所有节点预计算激活模型：把 TF / effector 在 ParentIds 中的下标以及调控方向解析成定长数组，供 DynamicBayesianNetwork.ComputeActivationScore()) 直接取用。
BuildActivationModel	1	构建单个节点的激活模型（TF 下标 / effector 下标 / 是否抑制）
LogTopologyScale	1	输出当前拓扑的规模诊断信息（父节点数分布 / CPT 行数估算 / hub 节点清单）。 CPT 的行数等于各父节点状态数的乘积，默认 3 态即 3^P（P = 父节点数），因此父节点数一旦不受限，初始化阶段的时间与内存都会指数级爆炸。该方法把规模显式输出到日志，用于定位导致爆炸的节点。
InitializeCPT	1	Initialize the CPT for a node based on its type and parent relationships.
GetParentStatesMap	1	构造父节点 ID → 状态列表的映射，供配置枚举与规模估算使用
ComputeDefaultDistribution	1	Compute the default probability distribution for a gene node given parent states.
ComputeActivationScore	1	Compute the activation score for a gene given parent (TF and metabolite) states.
StateToScore	1	Convert a state string to a numeric score (Low=0, Medium=0.5, High=1)
LearnParameters	1	Learn CPT parameters from RNAseq time-series data.
UpdateParametersOnline	1	Online parameter update using exponential moving average (EMA).
PredictNextState	1	Predict the next-state gene expression given current metabolite and TF abundances.
ComputeExpectedRNARate	1	Compute the expected RNA transcript abundance change rate from a distribution.
ComputeLogLikelihood	1	Compute the log-likelihood of the data given the current model.
GetThresholds	1	Get discretization thresholds for a node
DiscretizeValue	1	Discretize a continuous value into Low/Medium/High
Discretize	1	用指定节点自身的离散化阈值，把连续表达丰度离散为 Low/Medium/High。阈值解析顺序与内部完全一致：customThresholds → Config.NodeThresholds → 默认阈值（见 DynamicBayesianNetwork.GetThresholds())），从而保证外部调用方（例如设置野生型基线）与模型内部（参…
GetNode	1	Get a node by ID (returns Nothing if not found)
GetAllNodes	1	Get all nodes in the network
GetGeneNodes	1	Get all gene/operon nodes
GetTFNodes	1	Get all transcription factor nodes
GetMetaboliteNodes	1	Get all effector metabolite nodes
GetOperonGeneMapping	1	Get the operon-to-genes mapping
GetMarginalDistribution	1	Compute the marginal distribution of a node (averaging over all parent configs).
ResetToTopologyPrior	1	Reset all CPTs to topology-based defaults (discard learned parameters)
GetSummary	1	Get a summary string of the DBN structure (for debugging)
SaveToFile	1	Save the DBN (structure + learned parameters) to a text file.
LoadFromFile	1	Load the DBN from a text file (saved by SaveToFile)

03 Properties

Name	Overloads	Summary
Config	1	Configuration for the DBN (discretization, smoothing, rates, etc.)
topologySize	1

04 Fields

Name	Overloads	Summary
_activationModels	1	节点 ID → 预计算好的激活模型（父下标索引）。原始实现在每个父配置里都要对 ParentIds 做 IndexOf 字符串查找（O(P) 次、每次 O(P)），总复杂度为 O(3^P · P²)。预计算之后每个配置只需 O(P) 次数组取值。

05 Members

method .ctor #

#ctor

Default constructor with default configuration

method .ctor overload 2 #

#ctor(DBNConfig)

Constructor with custom configuration

method BuildFromTopology #

BuildFromTopology(IEnumerable(Of RegulatoryLink))

Build the DBN structure from a list of RegulatoryLink objects.

This method:

Creates nodes for all TFs, effector metabolites, and target operons
Sets up parent-child relationships (gene -> TF + effector parents)
Initializes CPTs based on biological heuristics (noisy-OR/AND gates)

After calling this method, the DBN is ready for prediction (topology-only mode) or parameter learning (data-fitting mode).

Parameters

Name	Type	Description
`links`	`IEnumerable(Of RegulatoryLink)`	List of regulatory links defining the network topology

method ApplyParentLimit #

ApplyParentLimit

父节点数量上限保护：对父节点数超过 Config.MaxParents 的节点做确定性截断。

截断策略：优先保留调控靶标更少（特异性更强）的父节点，同度时按 ID 排序，保证结果可复现；ParentIds / RegulatorTFs / TFEffectors 三者同步裁剪，避免状态不一致。默认上限 20 只用于隔离极端异常拓扑（正常情况下不会触发），不损失真实调控关系。

method BuildActivationModels #

BuildActivationModels

为所有节点预计算激活模型：把 TF / effector 在 ParentIds 中的下标以及调控方向解析成定长数组，供 DynamicBayesianNetwork.ComputeActivationScore() 直接取用。

method BuildActivationModel #

BuildActivationModel(DBNNode)

构建单个节点的激活模型（TF 下标 / effector 下标 / 是否抑制）

method LogTopologyScale #

LogTopologyScale

输出当前拓扑的规模诊断信息（父节点数分布 / CPT 行数估算 / hub 节点清单）。

CPT 的行数等于各父节点状态数的乘积，默认 3 态即 3^P（P = 父节点数），因此父节点数一旦不受限，初始化阶段的时间与内存都会指数级爆炸。该方法把规模显式输出到日志，用于定位导致爆炸的节点。

method InitializeCPT #

InitializeCPT(DBNNode)

Initialize the CPT for a node based on its type and parent relationships. For gene nodes, uses biological heuristics (noisy-OR/AND combination of activator and inhibitor effects). For other nodes, uses uniform distributions.

method GetParentStatesMap #

GetParentStatesMap(DBNNode)

构造父节点 ID → 状态列表的映射，供配置枚举与规模估算使用

method ComputeDefaultDistribution #

ComputeDefaultDistribution(DBNNode, List(Of String))

Compute the default probability distribution for a gene node given parent states. Uses the activation score to determine the distribution shape:

High activation score -> P(High) dominant
Moderate score -> P(Medium) dominant
Low score (inhibited) -> P(Low) dominant

method ComputeActivationScore #

ComputeActivationScore(DBNNode, List(Of String))

Compute the activation score for a gene given parent (TF and metabolite) states.

Uses a noisy-OR / noisy-AND combination model:

For activators: P(activation) = 1 - prod(1 - p_i) [noisy-OR]
For inhibitors: P(inhibition) = 1 - prod(1 - q_j) [noisy-OR]
Net score = activation - inhibition, mapped to [0, 1]

Interpretation of Effector enum (net effect on gene expression):

Activator: high effector concentration promotes gene activation
Inhibitor: high effector concentration promotes gene inhibition
Unknown: neutral effect (no contribution)

For TFs without effectors, uses the TF's DefaultRegulatoryDirection.

method StateToScore #

StateToScore(String)

Convert a state string to a numeric score (Low=0, Medium=0.5, High=1)

method LearnParameters #

LearnParameters(List(Of Dictionary(Of String, Double)), Dictionary(Of String, Tuple(Of Double, Double)))

Learn CPT parameters from RNAseq time-series data.

Uses Bayesian estimation with Dirichlet prior: P(s|parents) = (count(s) + alpha * prior(s)) / (total + alpha) where:

count(s) is the observed count of state s given the parent configuration
alpha is the smoothing parameter (Config.SmoothingAlpha)
prior(s) is the topology-based probability (serves as base distribution)

When alpha is large, the prior (topology) dominates. When alpha is small, the data dominates. When no data is available for a configuration, the prior is retained.

Parameters

Name	Type	Description
`rnaSeqTimeSeries`	`List(Of Dictionary(Of String, Double))`	List of time points. Each time point is a dictionary mapping node_id (gene/TF/metabolite) -> abundance value (continuous). Values should be normalized to [0, 1] for default thresholds, or custom thresholds should be provided.
`customThresholds`	`Dictionary(Of String, Tuple(Of Double, Double))`	Optional per-node discretization thresholds. Key = node ID, Value = Tuple(low_threshold, high_threshold). If not provided, uses Config.LowThreshold and Config.HighThreshold.

method UpdateParametersOnline #

UpdateParametersOnline(Dictionary(Of String, String), Dictionary(Of String, String))

Online parameter update using exponential moving average (EMA). Useful for incremental learning during simulation.

For each transition (current -> next), updates the CPT entry: P(s|parents) = (1 - lr) P_old(s|parents) + lr target(s) where target(s) = 1 if s == observed_next_state, else 0.

Parameters

Name	Type	Description
`currentStates`	`Dictionary(Of String, String)`	Discrete states at time t (node_id -> state string)
`nextStates`	`Dictionary(Of String, String)`	Discrete states at time t+1 (node_id -> state string)

method PredictNextState #

PredictNextState(Dictionary(Of String, Double), Dictionary(Of String, Double), Dictionary(Of String, String))

Predict the next-state gene expression given current metabolite and TF abundances.

This is the main interface for ODEs -> DBN coupling.

Algorithm:

Discretize metabolite concentrations and TF abundances into Low/Medium/High
For each gene node, look up CPT: P(gene[t+1] | parents[t])
Determine predicted state (argmax or sampled)
Compute expected RNA transcription rate from the distribution

Usage: Dim result = dbn.PredictNextState(metabolites, tfAbundances) ' result.GeneStates: operon_id -> "Low"/"Medium"/"High" ' result.RNAAbundanceChanges: operon_id -> transcription rate [0, 1]

Parameters

Name	Type	Description
`metaboliteAbundances`	`Dictionary(Of String, Double)`	Metabolite ID -> concentration (normalized to [0, 1] recommended). These come from the metabolic network ODEs.
`tfAbundances`	`Dictionary(Of String, Double)`	TF ID -> abundance (normalized to [0, 1] recommended). These come from the TF protein/RNA ODEs.
`currentGeneStates`	`Dictionary(Of String, String)`	Optional: current gene states for parent nodes that lack evidence. Used when a gene is also a TF (regulated by other TFs).

Returns

Prediction result with gene states, probabilities, and RNA abundance changes

method ComputeExpectedRNARate #

ComputeExpectedRNARate(Double(), List(Of String))

Compute the expected RNA transcript abundance change rate from a distribution.

E[rate] = P(High) k_high + P(Medium) k_basal + P(Low) * k_low

This value represents the expected transcription rate and can be used directly in ODE models: dR/dt = k_synthesis E[rate] - k_degradation R

method ComputeLogLikelihood #

ComputeLogLikelihood(List(Of Dictionary(Of String, Double)), Dictionary(Of String, Tuple(Of Double, Double)))

Compute the log-likelihood of the data given the current model. Useful for model evaluation, comparison, and convergence checking.

method GetThresholds #

GetThresholds(String, Dictionary(Of String, Tuple(Of Double, Double)))

Get discretization thresholds for a node

method DiscretizeValue #

DiscretizeValue(Double, Double, Double)

Discretize a continuous value into Low/Medium/High

method Discretize #

Discretize(String, Double)

用指定节点自身的离散化阈值，把连续表达丰度离散为 Low/Medium/High。

阈值解析顺序与内部完全一致：customThresholds → Config.NodeThresholds → 默认阈值（见 DynamicBayesianNetwork.GetThresholds()），从而保证外部调用方（例如设置野生型基线）与模型内部（参数学习、推理）使用同一套阈值。

Parameters

Name	Type	Description
`nodeId`	`String`	节点（基因）ID
`value`	`Double`	连续表达丰度

method GetNode #

GetNode(String)

Get a node by ID (returns Nothing if not found)

method GetAllNodes #

GetAllNodes

Get all nodes in the network

method GetGeneNodes #

GetGeneNodes

Get all gene/operon nodes

method GetTFNodes #

GetTFNodes

Get all transcription factor nodes

method GetMetaboliteNodes #

GetMetaboliteNodes

Get all effector metabolite nodes

method GetOperonGeneMapping #

GetOperonGeneMapping

Get the operon-to-genes mapping

method GetMarginalDistribution #

GetMarginalDistribution(String)

Compute the marginal distribution of a node (averaging over all parent configs). Useful for initialization and debugging.

method ResetToTopologyPrior #

ResetToTopologyPrior

Reset all CPTs to topology-based defaults (discard learned parameters)

method GetSummary #

GetSummary

Get a summary string of the DBN structure (for debugging)

method SaveToFile #

SaveToFile(String)

Save the DBN (structure + learned parameters) to a text file. Format: simple pipe-delimited text, human-readable.

method LoadFromFile #

LoadFromFile(String)

Load the DBN from a text file (saved by SaveToFile)

property Config #

Config

Configuration for the DBN (discretization, smoothing, rates, etc.)

field _activationModels #

_activationModels

节点 ID → 预计算好的激活模型（父下标索引）。

原始实现在每个父配置里都要对 ParentIds 做 IndexOf 字符串查找（O(P) 次、每次 O(P)），总复杂度为 O(3^P · P²)。预计算之后每个配置只需 O(P) 次数组取值。

property topologySize #

topologySize