Discourse graph view — a rendering of record fields, no schema change

Can Raman spectra distinguish yeast strains and species, and do batch effects confound the answer?

Can spontaneous Raman spectra be used to classify yeast strain and species identity, and to what extent do experimental batch effects confound such predictions?

Source: Leave-one-batch-out cross-validation reveals strong batch effects in Raman spectroscopy of yeast cultures (doi:10.57844/arcadia-xdmk-yq0w). Path record: arcadia-raman-path.json. Headline outcome partial. Species identity is recoverable from the spectra (median MCC 0.97 under leave-one-plate-out cross-validation). Strain identity is not: it appears recoverable under standard cross-validation (0.79) but collapses when whole plates are held out (0.32), and an adversarial classifier predicts plate identity almost perfectly, showing that batch effects dominate. All graphs · Back to the pilot

How to read these graphs

A discourse graph breaks a piece of research into typed one-sentence nodes and draws how they bear on one another. The grammar is that of Discourse Graphs (Chan and Akamatsu). Read each graph left to right: a question, the experiments that addressed it, the results they produced, and the hypothesis every result is drawn against.

QUE Question: What the work set out to answer.
HYP Hypothesis: The answer the authors expected. Every result earns one edge against it.
EXP Experiment: One attempt at the question: a trial arm, an assay, a survey, an analysis.
RES Result: One finding from an experiment, with where in the source it comes from.
ISS Issue: An open question or untried branch the work leaves for someone else.

addresses: An experiment addresses the question.
produces: An experiment produces a result.
supports: A positive result that backs the hypothesis.
opposes: A negative result that counts as evidence against the hypothesis, or a refutation of an earlier claim.
informs: A partial or mixed result that bears on the hypothesis without settling it.
consistent with: A negative or inconclusive result that does not count either way: no positive control passed, or there is no sensitivity or power statement, so the absence of an effect cannot be told from a measurement that failed.
relation: A thin dotted line: a link the source itself states between two experiments, such as one being derived from another.

A dashed box marks a result the rule above counts as uninformative, or an issue. In the drawing, click a node to jump to its text below.

The graph

Read left to right. The question (QUE) is addressed by 3 experiments (EXP), each of which produces results (RES), and each result bears on the hypothesis (HYP) with one of four edges. 5 support it, 1 oppose it, 0 inform it, and 2 are merely consistent with it because the informativeness rule would not count them as evidence either way. Click a node to jump to its full text below.

supports positive finding
opposes informative negative, or refutation
informs partial finding
consistent with inconclusive or uninformative: no positive control passed, or no sensitivity statement
thin grey a relation the record itself states between two attempts
dashed box a result the informativeness rule marks uninformative, or an issue

The grammar is that of Discourse Graphs (Chan and Akamatsu): QUE, HYP, EXP, RES and ISS nodes with supports, opposes and informs edges. Every node and edge is derived from fields already in the Negative-Results Record; the edge a finding earns is decided by its outcomeClass and informativeness. Issues (ISS) hang below the graph as a strip: the open questions and untried branches the path leaves for someone else to pick up.

Nodes

Question and hypothesis

QUE: Can spontaneous Raman spectra be used to classify yeast strain and species identity, and to what extent do experimental batch effects confound such predictions?
HYP: Evaluate whether spontaneous Raman spectra can distinguish strains and species of yeast, and determine the extent to which batch effects from experimental variation confound such predictions.

Experiments (3 attempts)

Node	Attempt	Type	Headline	JSON
EXP	An adversarial task shows plate identity is almost perfectly predictable, confirming a strong batch effect median MCC 1.0 (range 0.50-1.0); plate identity is very reliably predictable After correction: median MCC 0.10 (range -0.49-0.64); the plate-level batch effect has been removed	`computational`	`positive`	record
EXP	Species identity is recoverable from Raman spectra and survives holding out whole replicates median MCC 0.97 without batch correction and 0.93 with it; near-perfect under leave-one-plate-out cross-validation	`computational`	`positive`	record
EXP	Strain identity appears predictable from Raman spectra until whole replicates are held out, and then it is not median MCC 0.79 (range 0.71-0.93); strong diagonal confusion matrix, but strain identities are confounded with plate identities in this analysis median MCC 0.32 (range 0.19-0.42); only a few strains remain distinguishable After batch correction: median MCC 0.39 versus 0.32 before correction; not meaningfully improved	`computational`	`inconclusive-controls-failed`	record

Results (8 findings)

Node	Finding	Id	Outcome and informativeness	Edge	JSON
RES	Can plate identity be predicted from the spectra, given that the three plates are end-to-end replicates with no true biological difference? target: plate identity, a biologically meaningless label; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 1.0 (n = 3)	`plate-loso` in arcadia-raman-batch-effect	`positive` not-applicable rule applies to negative and inconclusive findings only	supports HYP	record
RES	Can plate identity be predicted from the spectra after batch correction, given that the three plates are end-to-end replicates with no true biological difference? target: plate identity, a biologically meaningless label; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 0.1 (n = 3)	`plate-loso-corrected` in arcadia-raman-batch-effect	`positive` not-applicable rule applies to negative and inconclusive findings only	supports HYP	record
RES	Can species identity (S. cerevisiae versus S. pombe) be predicted from Raman spectra under leave-one-plate-out cross-validation? target: yeast species identity from Raman spectra; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 0.97 (n = 3)	`species-lopo` in arcadia-raman-species-classification	`positive` not-applicable rule applies to negative and inconclusive findings only	supports HYP	record
RES	Can species identity (S. cerevisiae versus S. pombe) be predicted from Raman spectra under leave-one-plate-out cross-validation after batch correction? target: yeast species identity from Raman spectra; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 0.93 (n = 3)	`species-lopo-corrected` in arcadia-raman-species-classification	`positive` not-applicable rule applies to negative and inconclusive findings only	supports HYP	record
RES	Can strain identity be predicted from Raman spectra under standard five-fold cross-validation? target: yeast strain identity from Raman spectra; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 0.79 (n = 3)	`strain-standard-cv` in arcadia-raman-strain-classification	`refuted` not-applicable an affirmative refutation is judged on the refuting analysis, not on the controls of the run it refutes	opposes HYP	record
RES	Can strain identity be predicted from Raman spectra when whole experimental replicates are held out? target: yeast strain identity from Raman spectra; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 0.32 (n = 3)	`strain-lopo` in arcadia-raman-strain-classification	`inconclusive-controls-failed` uninformative negative control not clean; no sensitivity or power statement	consistent-with HYP	record
RES	Can strain identity be predicted from Raman spectra after the plate-level batch effect has been removed? target: yeast strain identity from Raman spectra; multiclass Matthews correlation coefficient (MCC), bounded in [-1, +1], where +1 is perfect prediction, 0 is random guessing and negative values indicate systemic misclassification; computed per fold and reported as the median across folds = 0.39 (n = 3)	`strain-lopo-corrected` in arcadia-raman-strain-classification	`inconclusive-underpowered` uninformative no sensitivity or power statement	consistent-with HYP	record
RES	Do these conclusions depend on the choice of classifier or batch-correction method? target: robustness of the conclusions to model and batch-correction choice; median MCC across every combination of dataset, model and task = not meaningfully different between random forest and support vector machine, or between the linear mixed model and ComBat	`strain-robustness` in arcadia-raman-strain-classification	`positive` not-applicable rule applies to negative and inconclusive findings only	supports HYP	record

Relations between attempts

arcadia-raman-batch-effect comparedWith arcadia-raman-species-classification species signal survives the batch effect that destroys the strain signal
arcadia-raman-strain-classification wasDerivedFrom arcadia-raman-batch-effect the adversarial plate task is the negative control for these findings
arcadia-raman-strain-classification wasDerivedFrom arcadia-raman-species-classification the species task is the positive control for these findings

Issues (6 open questions and untried branches)

ISS How much of the residual strain-level signal loss is stochastic sample-level batch effect and how much is genuine absence of strain-specific Raman signatures? (open-question)
ISS Incorporate at least three end-to-end replicates from the outset. (proposed-by-authors)
ISS Include leave-one-replicate-out cross-validation in analysis workflows. (proposed-by-authors)
ISS Include adversarial prediction tasks in which a model is trained to predict biologically meaningless variables such as replicate or plate identifiers. (proposed-by-authors)
ISS Repeat the strain task on a panel with greater genetic diversity; the authors state the negative result may not generalise to more divergent strains. (proposed-by-authors)
ISS Repeat with a sample-preparation protocol that avoids desiccation, which the authors identify as a likely source of the plate-level batch effect. (proposed-by-authors)