Neural signatures of glioma progression using implanted neural interfaces

Key results

1. SuggestedIn an adult glioblastoma (GBM; U-87, nude mice) and a pediatric DIPG6 xenograft (NSG mice), gamma-band local field potential (LFP) power rose relative to saline controls as tumors progressed, and day-wise linear support vector machines (SVMs) told tumor-bearing from control mice apart.

   GBM, 5 cancer vs 6 healthy mice: cancer-favoring SVM weights concentrated at 40–80 Hz, and gamma power went from below controls early to above them later. DIPG6, 4 vs 4: cancer-favoring weights above 80 Hz, with gamma bands rising earliest and most. Leave-one-animal-out accuracy beat pooled permutation nulls (461 and 69 unique shuffles) on several, but not all, recording days. · 2.2 Neural biomarkers of cancer progression in a…; 2.3 Neural biomarkers of cancer progression in a…; Permutation testing for classifier performance; Figure legends

2. SuggestedLower frequencies diverged between the models. In GBM, theta, alpha and beta power moved opposite to gamma and ended below controls. In DIPG6, the lower bands rose progressively, giving a broad-band increase that plateaued late in the study.

   GBM: control-favoring SVM weights sat near 20–40 Hz; very low frequencies were higher in cancer mice early, then consistently lower. DIPG6: after a transient dip at all frequencies on the first recording day, power rose first above about 80 Hz and then across all frequencies for roughly the final 3 weeks of a 55-day series of weekly recordings. · 2.2 Neural biomarkers of cancer progression in a…; 2.3 Neural biomarkers of cancer progression in a…; 4.2.2 Study 2

3. ShownFrom a single recording session, the LFP spectrum predicted bioluminescence imaging (BLI) tumor burden in held-out U-87 mice. Power at every frequency correlated positively with BLI, most strongly up to about 70 Hz.

   Study 3, n=18: an elastic-net linear model on 10 sampled frequencies plus all their interactions (best fit pure L2, α = 0.02) reached a leave-one-out cross-validated R^2 of 0.61. Refits on 10,000 BLI-label shuffles were significantly worse (p = 0.002). The strongest single correlation was at 35 Hz. · 2.4 Predicting tumor burden using neural…; 4.2.3 Study 3; 4.3.3 Predicting BLI scores based on neural…; Figure legends

4. ShownOver 38 days in untreated and temozolomide (TMZ)-treated U-87 mice, power in a single band, combined with day and treatment-group terms, predicted the log BLI of held-out mice over time; low gamma gave the best fit.

   Study 1: 80 mouse-day points from 18 mice (8 untreated, 10 TMZ), in a linear model with band power, recording day, treatment group, and day-by-group and power-by-day interactions (L2 alpha = 10). Leave-one-mouse-out R^2 was 0.32 for low gamma (30–80 Hz). Permutations swapping whole BLI trajectories gave p = 0.004 (low gamma), 0.019 (beta) and 0.022 (high gamma). · 2.4 Predicting tumor burden using neural…; 4.2.1 Study 1; 4.3.3 Predicting BLI scores based on neural…; 4.4 Statistics

5. SuggestedBand-power trajectories were fitted with peaked (quadratic) and plateau (square-root) curves. Untreated tumor mice favored the peak; healthy and TMZ mice fit both about equally, though most channels in both cancer groups leaned to the peak. Low-gamma rise rate predicted BLI growth rate without treatment labels.

   Only mice with at least 5 paired BLI and recording days were used (6 cancer, 5 healthy, 6 TMZ in Figure 6b). Group-averaged TMZ trajectories looked more saturating, like healthy ones. Rise rate correlated with the slope of log BLI above about 15 Hz. Low-gamma (30–80 Hz) rise rate predicted BLI growth rate with a leave-one-out R^2 of 0.63; against 10,000 label shuffles, p = 0.01. · 2.5 Predicting tumor growth rate across groups…; 4.3.4 Predicting tumor growth rate based on…; 4.4 Statistics; Figure legends