msbooster_params.txt

## MSBooster parameter template (defaults match Constants.java for v1.6.4)
## Lines without '=' are ignored. A value of 'null' means "use the code default"
## (the parser skips that key); empty values become empty strings, which is only
## valid for String fields. Edit the (REQUIRED) entries for your run.

############################################################
# I/O and runtime
############################################################
mzmlDirectory =                                  # (REQUIRED) path(s) to mzML/mgf, space-separated
pinPepXMLDirectory =                             # (REQUIRED) path(s) to .pin (and .pepXML), space-separated
fragger = null                                   # path to MSFragger fragger.params; null skips
outputDirectory = null                           # default: <pin parent>/MSBooster
editedPinSuffix = edited
renamePin = 1
deletePreds = false
loadingPercent = 10
numThreads = 0                                   # 0 -> Runtime.availableProcessors() - 1
plotExtension = png

############################################################
# Decoy / digestion (most are auto-filled when 'fragger' is set)
############################################################
fasta =
decoyPrefix = rev_
cutAfter = KR
butNotAfter = P
digestMinLength = 7
digestMaxLength = 50
digestMinMass = 500.0
digestMaxMass = 5000.0
minPrecursorCharge = 1
maxPrecursorCharge = 8
keepDecoys = 1

############################################################
# Prediction models
############################################################
spectraModel =
rtModel =
imModel =
auxSpectraModel =
useSpectra = true
useRT = true
useIM = false
findBestSpectraModel = false
findBestRtModel = false
findBestImModel = false
rtCollection = general                           # general | isolabel
ms2Collection = general                          # general | isolabel | hla | timstof
imCollection = general
rtSearchModelsString =
ms2SearchModelsString =
imSearchModelsString =
rtBestModelSearchMetric = top
imBestModelSearchMetric = top
spectraBestModelSearchMetric = median

############################################################
# Local predictors (name one with spectraModel/rtModel/imModel;
# DIA-NN is what runs when none is named)
############################################################
FragCast = null                                  # path to the FragCast executable (selectable via spectraModel/rtModel/imModel = FragCast)
FragCastModelDir =                               # dir with FragCast-{RT,IM,Spec}(-Spec-Fast).onnx; empty -> resolved beside the exe
fragCastTopN = 20                                # build-library: top-N fragments per precursor by intensity
fragCastMinFragMz = 200.0                        # build-library: drop fragments below this m/z
fragCastMinRelIntensity = 0.01                   # build-library: drop fragments below this relative intensity
fragCastMinFragSize = 2                          # build-library: minimum fragment length in residues
DiaNN = null                                     # path to the DIA-NN executable (only when spectraModel/rtModel/imModel = DIA-NN)

############################################################
# FragCast local transfer learning (CPU only, no server)
#
# Fine-tunes FragCast on a spectral library you supply, producing one zip that
# holds every model it wrote. Everything runs through the same FragCast
# executable on this machine; nothing is uploaded. Every fine-tuned model is
# kept, whatever it scored.
#
# Fine-tuning is its own run, and rescoring with the result is another:
#   java -cp MSBooster.jar transferlearn.FragCastLocalWorkflow \
#        --paramsList msbooster_params.txt --library library.tsv
# names a FragCastModelZip; give that to a later run to rescore with it. The zip
# goes where --output-dir says, or beside the library. For server-based transfer
# learning instead, see transferlearn.SecondPassWorkflow.
#
# One library is all it needs. FragCast holds out its own slice of it and
# selects the best epoch against that - so there is no second library to supply
# and no holdout fraction to choose.
#
# Nothing about HOW the fine-tune runs is settable here: epochs, learning rate,
# batch size, warmup, training depth and seed are FragCast's own, and MSBooster
# omits those flags so each task trains the way the executable considers right.
#
# The library's headers must be the ones FragCast's fine-tune reads:
# ModifiedPeptideSequence and PrecursorCharge are required; NormalizedRetentionTime,
# PrecursorIonMobility and LibraryIntensity/FragmentType/FragmentCharge/
# FragmentSeriesNumber supply the rt, im and spec targets respectively. Exports
# that name these columns differently must be renamed, not guessed at.
############################################################

# Custom FragCast weights for prediction (--rt-onnx/--im-onnx/--spec-onnx).
# Empty = the pretrained models in FragCastModelDir. Set them to rescore with a
# model a fine-tune produced, without fine-tuning again - usually you want
# FragCastModelZip below instead, which names all three at once.
FragCastRtOnnx =
FragCastImOnnx =
FragCastSpecOnnx =                               # supersedes FragCast-Fast: the weights file decides which MS2 model runs

# The same three models as one file, and what a fine-tune produces: it gathers
# the ONNX files it exported into this zip and then removes them, so the zip is
# the whole product. Naming it here unpacks it and fills in whichever of the
# three parameters above was left empty. Anything named above wins over it, and
# the run says so rather than overriding it quietly.
FragCastModelZip =

############################################################
# Reusing existing prediction files (skip if regenerating)
############################################################
spectraPredFile = null
RTPredFile = null
IMPredFile = null

############################################################
# Koina (only required when a Koina model is selected)
############################################################
KoinaURL =
useKoina = false
numKoinaAttempts = 5
initialKoinaMillisecondsToWaitRtIm = 30000
initialKoinaMillisecondsToWaitMs2 = 60000
numPSMsToCalibrate = 1000
diannPeptidePredictionLimit = 2000000

############################################################
# MS/MS spectral matching
############################################################
ppmTolerance = 0.0                               # 0 -> use highResppmTolerance (or fragger params)
lowResppmTolerance = 300.0
highResppmTolerance = 20.0
matchWithDaltons = null
DaTolerance = 0.05
useTopFragments = true
topFragments = 20
useBasePeak = true
percentBasePeak = 1.0
splitPredInputFile = 1                           # legacy; FragPred no longer splits

############################################################
# Fragmentation / NCE
############################################################
FragmentationType = auto                         # HCD, CID, ETD, ETHCD, ECD, EID, UVPD, ETCID, or auto
instrument =
autoSwitchFragmentation = true
NCE = 25
minNCE = 20
maxNCE = 40
calibrateNCE = true
ignoredFragmentIonTypes =
onlyFragmentIonTypes =

############################################################
# RT/IM calibration (LOESS)
############################################################
loessEscoreCutoff = 3.1622776E-4                 # 10^-3.5
rtLoessRegressionSize = 5000
imLoessRegressionSize = 1000
minLoessRegressionSize = 100
minLinearRegressionSize = 10
loessBandwidth = 0.01,0.05,0.1,0.2
robustIters = 2
regressionSplits = 5
massesForLoessCalibration =
massOffsets =
massOffsetsDetailed =
massDiffToVariableMod = 0
loessScatterOpacity = 0.35
realMinuteFilter = 10000.0
percentRTgradientFilter = 100.0
removeWashGradient = false
washGradientBins = 100

############################################################
# Modifications
############################################################
unimodObo = null
additionalModsFile =

############################################################
# Features to compute
############################################################
features = unweightedSpectralEntropy,weightedSpectralEntropy,hypergeometricProbability,intersection,top6matchedIntensity,deltaRTLOESS,deltaRTLOESSreal