/**
* @param tree
* @param node
* @return and array of the total amount of time spent in each of the discrete states along the
* branch above the given node.
*/
private double[] getProcessValues(final Tree tree, final NodeRef node) {
double[] processValues = null;
double branchTime = tree.getBranchLength(node);
if (mode == Mode.MARKOV_JUMP_PROCESS) {
processValues = (double[]) trait.getTrait(tree, node);
} else if (mode == Mode.PARSIMONY) {
// an approximation to dwell times using parsimony, assuming
// the state changes midpoint on the tree. Does a weighted
// average of the equally parsimonious state reconstructions
// at the top and bottom of each branch.
if (treeChanged) {
fitchParsimony.initialize(tree);
// Debugging test to count work
// treeInitializeCounter += 1;
// if (treeInitializeCounter % 10 == 0) {
// System.err.println("Cnt: "+treeInitializeCounter);
// }
treeChanged = false;
}
int[] states = fitchParsimony.getStates(tree, node);
int[] parentStates = fitchParsimony.getStates(tree, tree.getParent(node));
processValues = new double[fitchParsimony.getPatterns().getStateCount()];
for (int state : states) {
processValues[state] += branchTime / 2;
}
for (int state : parentStates) {
processValues[state] += branchTime / 2;
}
for (int i = 0; i < processValues.length; i++) {
// normalize by the number of equally parsimonious states at each end of the branch
// processValues should add up to the total branch length
processValues[i] /= (states.length + parentStates.length) / 2;
}
} else if (mode == Mode.NODE_STATES) {
processValues = new double[dataType.getStateCount()];
// if (indicatorParameter != null) {
// // this array should be size #states NOT #rates
// processValues = new double[indicatorParameter.getDimension()];
// } else {
// // this array should be size #states NOT #rates
// processValues = new double[rateParameter.getDimension()];
// }
// if the states are being sampled - then there is only one possible state at each
// end of the branch.
int state = ((int[]) trait.getTrait(tree, node))[traitIndex];
processValues[state] += branchTime / 2;
int parentState = ((int[]) trait.getTrait(tree, tree.getParent(node)))[traitIndex];
processValues[parentState] += branchTime / 2;
}
return processValues;
}
@Override
public List<Integer> getAncestors(int i, boolean b) {
List<Integer> ancestors = new ArrayList<Integer>();
if (b) ancestors.add(i);
for (NodeRef n = tree.getParent(tree.getNode(i)); n != null; n = tree.getParent(n))
ancestors.add(n.getNumber());
return ancestors;
}
@Override
public int getSibling(int i) {
NodeRef n = tree.getNode(i);
if (tree.isRoot(n)) return RootedTree.NULL;
NodeRef p = tree.getParent(n);
int c1 = tree.getChild(p, 0).getNumber();
int c2 = tree.getChild(p, 1).getNumber();
return n.getNumber() == c2 ? c1 : c2;
}
public void getTransitionProbabilities(
Tree tree, NodeRef node, int rateCategory, double[] matrix) {
NodeRef parent = tree.getParent(node);
final double branchRate = branchModel.getBranchRate(tree, node);
// Get the operational time of the branch
final double startTime = tree.getNodeHeight(parent);
final double endTime = tree.getNodeHeight(node);
final double branchTime = branchRate * (startTime - endTime);
if (branchTime < 0.0) {
throw new RuntimeException("Negative branch length: " + branchTime);
}
double distance = siteModel.getRateForCategory(rateCategory) * branchTime;
int matrixCount = 0;
boolean oneMatrix = (getEpochWeights(startTime, endTime, weight) == 1);
for (int m = 0; m < numberModels; m++) {
if (weight[m] > 0) {
SubstitutionModel model = modelList.get(m);
if (matrixCount == 0) {
if (oneMatrix) {
model.getTransitionProbabilities(distance, matrix);
break;
} else model.getTransitionProbabilities(distance * weight[m], resultMatrix);
matrixCount++;
} else {
model.getTransitionProbabilities(distance * weight[m], stepMatrix);
// Sum over unobserved state
int index = 0;
for (int i = 0; i < stateCount; i++) {
for (int j = 0; j < stateCount; j++) {
productMatrix[index] = 0;
for (int k = 0; k < stateCount; k++) {
productMatrix[index] +=
resultMatrix[i * stateCount + k] * stepMatrix[k * stateCount + j];
}
index++;
}
}
// Swap pointers
double[] tmpMatrix = resultMatrix;
resultMatrix = productMatrix;
productMatrix = tmpMatrix;
}
}
}
if (!oneMatrix) System.arraycopy(resultMatrix, 0, matrix, 0, stateCount * stateCount);
}
@Override
public int getParent(int i) {
NodeRef n = tree.getParent(tree.getNode(i));
return n != null ? n.getNumber() : RootedTree.NULL;
}
/**
* Traverse the tree calculating partial likelihoods.
*
* @param tree tree
* @param node node
* @param operatorNumber operatorNumber
* @param flip flip
* @return boolean
*/
private boolean traverse(Tree tree, NodeRef node, int[] operatorNumber, boolean flip) {
boolean update = false;
int nodeNum = node.getNumber();
NodeRef parent = tree.getParent(node);
if (operatorNumber != null) {
operatorNumber[0] = -1;
}
// First update the transition probability matrix(ices) for this branch
if (parent != null && updateNode[nodeNum]) {
final double branchRate = branchRateModel.getBranchRate(tree, node);
final double parentHeight = tree.getNodeHeight(parent);
final double nodeHeight = tree.getNodeHeight(node);
// Get the operational time of the branch
final double branchLength = branchRate * (parentHeight - nodeHeight);
if (branchLength < 0.0) {
throw new RuntimeException("Negative branch length: " + branchLength);
}
if (flip) {
substitutionModelDelegate.flipMatrixBuffer(nodeNum);
}
branchUpdateIndices[branchUpdateCount] = nodeNum;
branchLengths[branchUpdateCount] = branchLength;
branchUpdateCount++;
update = true;
}
// If the node is internal, update the partial likelihoods.
if (!tree.isExternal(node)) {
// Traverse down the two child nodes
NodeRef child1 = tree.getChild(node, 0);
final int[] op1 = {-1};
final boolean update1 = traverse(tree, child1, op1, flip);
NodeRef child2 = tree.getChild(node, 1);
final int[] op2 = {-1};
final boolean update2 = traverse(tree, child2, op2, flip);
// If either child node was updated then update this node too
if (update1 || update2) {
int x = operationCount[operationListCount] * Beagle.OPERATION_TUPLE_SIZE;
if (flip) {
// first flip the partialBufferHelper
partialBufferHelper.flipOffset(nodeNum);
}
final int[] operations = this.operations[operationListCount];
operations[x] = partialBufferHelper.getOffsetIndex(nodeNum);
if (useScaleFactors) {
// get the index of this scaling buffer
int n = nodeNum - tipCount;
if (recomputeScaleFactors) {
// flip the indicator: can take either n or (internalNodeCount + 1) - n
scaleBufferHelper.flipOffset(n);
// store the index
scaleBufferIndices[n] = scaleBufferHelper.getOffsetIndex(n);
operations[x + 1] = scaleBufferIndices[n]; // Write new scaleFactor
operations[x + 2] = Beagle.NONE;
} else {
operations[x + 1] = Beagle.NONE;
operations[x + 2] = scaleBufferIndices[n]; // Read existing scaleFactor
}
} else {
if (useAutoScaling) {
scaleBufferIndices[nodeNum - tipCount] = partialBufferHelper.getOffsetIndex(nodeNum);
}
operations[x + 1] = Beagle.NONE; // Not using scaleFactors
operations[x + 2] = Beagle.NONE;
}
operations[x + 3] = partialBufferHelper.getOffsetIndex(child1.getNumber()); // source node 1
operations[x + 4] =
substitutionModelDelegate.getMatrixIndex(child1.getNumber()); // source matrix 1
operations[x + 5] = partialBufferHelper.getOffsetIndex(child2.getNumber()); // source node 2
operations[x + 6] =
substitutionModelDelegate.getMatrixIndex(child2.getNumber()); // source matrix 2
operationCount[operationListCount]++;
update = true;
if (hasRestrictedPartials) {
// Test if this set of partials should be restricted
if (updateRestrictedNodePartials) {
// Recompute map
computeNodeToRestrictionMap();
updateRestrictedNodePartials = false;
}
if (partialsMap[nodeNum] != null) {}
}
}
}
return update;
}
/**
* Traverse the tree calculating partial likelihoods.
*
* @return whether the partials for this node were recalculated.
*/
protected boolean traverse(Tree tree, NodeRef node) {
boolean update = false;
boolean rootUpdated = false;
int nodeNum = node.getNumber();
NodeRef parent = tree.getParent(node);
// First update the transition probability matrix(ices) for this branch if it is a normal branch
if (parent != null && updateNode[nodeNum]) {
final double branchRate = branchRateModel.getBranchRate(tree, node);
// Get the operational time of the branch
final double branchTime =
branchRate * (tree.getNodeHeight(parent) - tree.getNodeHeight(node));
if (branchTime < 0.0) {
throw new RuntimeException("Negative branch length: " + branchTime);
}
cenancestorlikelihoodCore.setNodeMatrixForUpdate(nodeNum);
for (int i = 0; i < categoryCount; i++) {
double branchLength = siteModel.getRateForCategory(i) * branchTime;
siteModel.getSubstitutionModel().getTransitionProbabilities(branchLength, probabilities);
cenancestorlikelihoodCore.setNodeMatrix(nodeNum, i, probabilities);
}
update = true;
} else if (parent == null
&& cenancestorHeight != null
&& updateNode[nodeNum]) // The root has to be updated
{
// First update the transition probability matrix(ices) for the root-cenancestor fake branch
rootUpdated = true;
// Get the operational time of the fake branch from the root to the cenancestor
double rootHeight = treeModel.getNodeHeight(treeModel.getRoot());
double branchRate =
branchRateModel.getBranchRate(
rootHeight,
getCenancestorHeight()); // TODO: Could this be easily improved? I would do to adapt
// the tree structure and abstact tree likelihood
double branchTime =
branchRate
* getCenancestorBranch(); // TODO: Could this be easily improved? The same as before
for (int i = 0; i < categoryCount; i++) {
double branchLength = siteModel.getRateForCategory(i) * branchTime;
siteModel.getSubstitutionModel().getTransitionProbabilities(branchLength, probabilities);
cenancestorlikelihoodCore.setNodeMatrix(nodeNum, i, probabilities);
}
}
// If the node is internal, update the partial likelihoods.
if (!tree.isExternal(node)) {
// Traverse down the two child nodes
NodeRef child1 = tree.getChild(node, 0);
final boolean update1 = traverse(tree, child1);
NodeRef child2 = tree.getChild(node, 1);
final boolean update2 = traverse(tree, child2);
// If either child node was updated then update this node too
if (update1 || update2 || rootUpdated) {
if (update1 || update2) {
final int childNum1 = child1.getNumber();
final int childNum2 = child2.getNumber();
cenancestorlikelihoodCore.setNodePartialsForUpdate(nodeNum);
if (integrateAcrossCategories) {
cenancestorlikelihoodCore.calculatePartials(childNum1, childNum2, nodeNum);
} else {
cenancestorlikelihoodCore.calculatePartials(
childNum1, childNum2, nodeNum, siteCategories);
}
if (COUNT_TOTAL_OPERATIONS) {
totalOperationCount++;
}
}
if (parent == null) {
// No parent this is the root of the tree
double[] partials;
int nodeNumCenan = getCenancestorIndex();
if (cenancestorHeight != null) {
if (rootUpdated) {
// Calculate the partials at the cenancestor. The transition matrix of the root was
// calculated before.
cenancestorlikelihoodCore.setNodePartialsForUpdate(nodeNumCenan);
if (integrateAcrossCategories) {
cenancestorlikelihoodCore.calculatePartials(nodeNum, nodeNumCenan);
} else {
cenancestorlikelihoodCore.calculatePartials(nodeNum, nodeNumCenan, siteCategories);
}
}
partials = getCenancestorPartials();
} else { // Using the cenancestor model without cenancestor date. It assumes that the root
// of the tree is the cenancestor. Not tested. It shouldn't be normally used
// either.
partials = getRootPartials();
}
// calculate the pattern likelihoods
double[] frequencies = frequencyModel.getFrequencies();
cenancestorlikelihoodCore.calculateLogLikelihoods(
partials, frequencies, patternLogLikelihoods);
}
update = true;
}
}
return update;
}