Anterior-posterior gradient of plasticity in primate prefrontal cortex

The functional organization of the primate prefrontal cortex has been a matter of debate with some models speculating dorso-ventral and rostro-caudal specialization while others suggesting that information is represented dynamically by virtue of plasticity across the entire prefrontal cortex. To address functional properties and capacity for plasticity, we recorded from different prefrontal sub-regions and analyzed changes in responses following training in a spatial working memory task. This training induces more pronounced changes in anterior prefrontal regions, including increased firing rate during the delay period, selectivity, reliability, information for stimuli, representation of whether a test stimulus matched the remembered cue or not, and variability and correlation between neurons. Similar results are obtained for discrete subdivisions or when treating position along the anterior-posterior axis as a continuous variable. Our results reveal that anterior aspects of the lateral prefrontal cortex of non-human primates possess greater plasticity based on task demands.


Database of recordings used in different analyses
We collected data using some variations of tasks and stimulus sets, including sessions that used different shaped stimuli; sessions using a variable duration of the delay period; and sessions in which the sample stimulus was always a match. The selection of task and stimulus set was done prior to the onset of recordings and the properties of neurons recorded were not known ahead of time. We therefore used the largest available set that was appropriate for each analysis in order to increase statistical power. Our conclusions would not differ had we limited all analyses to a common dataset (S 5 ) but insufficient power would be available for some comparisons, in some areas.
In total, we recorded from 3908 neurons (set S 1 ) before and after training, while the monkeys were either passively viewing stimuli or performing the active task, across all task and stimulus sets (Posterior-dorsal region, n=182 neurons pre-training/211 neurons post-training; Mid-dorsal, n=756/690; Anterior-dorsal, n=359/222; Posterior-ventral, n=633/612; Anteriorventral, n=122/121; see also Supplemental Table 1). We used this dataset in analyses testing responsiveness to the first stimulus in the trial, in Fig. 3A and S1.
A subset, S 2 , included 3704 neurons (95% of the total) that were tested with white squares, appearing at the nine stimulus locations of Fig. 1 and was used in the analyses comparing maximum firing rate, and spatial selectivity of neuronal responses to the first stimulus in the trial, in A subset of the preceding set, S 3 , using delay periods of a fixed duration equal to 1.5 s, included 3508 neurons (90% of the total). These neurons were used in the analyses plotting PSTH across the trial in Fig. 2  A subset of the S 2 set, S 4 , included neurons tested with both match and nonmatch stimuli displayed during the sample period. We had 3132 neurons (80% of total) thus available, and these were used for the Match-Nonmatch difference analysis in Fig. 7  Finally a common subset of all preceding sets, S 5 , included neurons tested with the stimuli at all spatial locations, with a fixed delay period duration equal to 1.5 s, and included both match and nonmatch stimuli. This set included 3040 neurons (78% of the total). This was used for the ROC analysis of

Analysis based on same 4 monkeys tested before and after training
The results presented in the main text relied on all available data from 6 monkeys prior to training and 4 of these monkeys after they were trained in working memory tasks. To ensure that any changes between stages were not due to individual differences driven by the monkeys not tested after training, we repeated our analysis, restricting our sample to the same animals In terms of percentage of neurons activated during the delay period, a significant increase was present overall after training, which was greatest for the anterior-dorsal region.

The highest Selectivity Index values (Supplementary
Prior to training, no significant difference was present in terms of mutual information between dorsal areas (Supplementary Figure 11C) was observed (1-way ANOVA, F 2,967 =0.19, p=0.827). Following training, mutual information increased significantly for all dorsal areas, however the increase was proportionally greater for the anterior dorsal areas (t-test, p=0.018, p=9.7x10 -9 , p=5.0x10 -14 , for the posterior, mid, and anterior areas, respectively). As a result, mean information values between dorsal areas was now significantly different (1-way ANOVA, F 2,1063 =16.4, p=9.8x10 -8 ). In the ventral regions, we saw a similar increase in Mutual Information after training, however the increase for the anterior-ventral region did not reach significance (ttest, t 183 =0.439, p=0.439).
Finally, we examined the absolute difference between match and nonmatch responses between training stages (Supplementary Figure 11D). Prior to training, the posterior-dorsal region had a lower absolute difference than the mid-dorsal and anterior-dorsal (1-way ANOVA, F 2,991 =3.82, p=0.02). This difference was not present when we included the data from all 6 monkeys. When we examined the effects of training using a 2-way ANOVA, we observed that there was a significant main effect of training stage (F 1,1751 =40.23, p=2.9x10 -10 ) and interaction between training stage and dorsal region (F 2,1751 =6.39, p=0.0017). There were no differences in match/nonmatch absolute difference response rates prior to training in the ventral region, but there were significant main effects of training stage (F 1,1153 =19.92, p=8.9x10 -6 ) and interaction of training stage and ventral region (F 2,1153 =13.03, p=3.2x10 -4 ).
In conclusion, analysis restricted to the same 4 subjects before and after training confirmed that the anterior areas exhibited the highest increase in the percentage of active neurons during the delay period, in firing rate during this period, mutual information, and in selectivity index. Restricting data to 4 monkeys also revealed a difference in |Match-Nonmatch| preference between dorsal areas prior to training, however this finding did not change that the highest increase was observed in the anterior dorsal area after training.

Analysis based of Neurons with Excellent Spike Isolation
In order to discount the possibility that changes were observed between training stages due to systematic differences in spike isolation, or levels of noise in recordings, we repeated our analysis based on a subset of neurons, for which excellent isolation was achieved, based on  (Fig. 2).

Recording Depth Analysis
Recording depths could be estimated approximately, based on the initial detection of neural activity as electrodes were advanced into the cortex and the relative distance traversed from that point to the recording site of each neuron. Our recordings targeted the supra-granular layers, based on anatomical expectations about neurons active during the delay period 1, 2 . Prior to training, mean recording depths in the dorsal areas were 0.45, 0.50, and 0.28 mm for the posterior, middle, and anterior region, respectively. These means included penetrations that descended into the principal sulcus, where greater depths do not necessarily imply deeper layers. After training, average depths also corresponded to the supra-granular layers: 0.38, 0.29, and 0.38 mm for the posterior, middle, and anterior region, respectively.
Depths of neurons in the ventral areas (which did not include sulci) were even more superficial. Prior to training, the mean depth of recordings was 0.34 and 0.25 mm for the posterior and anterior ventral PFC, respectively. After training the corresponding depths were 0.35 and 0.19 mm, for the two areas respectively.

Analysis of Individual-Monkey-Derived Means
We performed regression analyses on mean values obtained separately from each monkey, using the average of neuronal values from each monkey as a single observation, in order to ensure that results obtained by pooling all data were not skewed by data from 1-2 individuals.
The dependence of neuronal measures on AP position determined in this fashion (Fig.   8 3H,I,J,K,P,Q,R,S) tended to agree with the results based from all data (Fig. 3D,E,F,G,L,M,N,O).
No significant dependence of mean delay-period discharge rate on AP position was observed in dorsal areas prior to training (β pooled = -0.2 spikes/percentiled-location), when all data were pooled together (Fig. 3D). A similar, negative slope (β individual =-5.5) was also obtained from the individual-monkey-derived means (Fig. 3H). In contrast, a significant positive slope was observed after training from pooled data (β pooled = 9.0, Fig. 3E) and a positive slope was obtained from individual-monkey data (β individual =10.6, Fig. 3I Similarly, differences between match and non-match responses showed the same dependence on AP position when obtained from pooled data, or from individual-monkey-derived means. Regressing the absolute difference between match and nonmatch responses |M-NM| on AP position produced slopes that were not significantly different than zero in the dorsal prefrontal cortex prior to training (β pooled = 0.04, β individual = -1.6). These were positive after training (β pooled = 5.1, β individual = 4.9). The same pattern of changes was observed in the ventral prefrontal cortex, and with similar values between pooled and individual-monkey-derived data before training (β pooled = 0.05, β individual = -1.6) and after training (β pooled = 8.0, β individual = 11.2).

SUPPLEMENTARY DISCUSSION
Although our analysis relied on thousands of neurons sampled from six monkeys, it was not practical to sample all areas, in all animals. Data from the most anterior regions in particular were only drawn from a few animals. It is important to emphasize however, that the effects of training we reported in the anterior regions were consistent with previous reports in the literature that have described robust modulation of prefrontal neurons by task variables [3][4][5][6][7] , which presumably come about after training. It was the relative lack of changes in the posterior areas after training (which were sampled densely, in multiple monkeys) that were responsible for the gradient of plasticity that we observed across the anterior-posterior axis of the prefrontal cortex ( Fig. 3-7).
Why is the anterior prefrontal cortex more plastic after training in a spatial working memory task? Our neurophysiological experiments could not address this question, however, anatomical evidence points to an organization of inputs that funnels increasingly more complex information into anterior areas. The most posterior aspect of PFC, (posterior-dorsal in our study), receives sensory afferents [8][9][10] , which allow a faithful representation of stimuli even in naïve animals. More anterior, hierarchically superior, auditory areas of the temporal lobe project to more anterior prefrontal subdivisions [11][12][13] . Face and color information appears to be concentrated in cortical patches, also suggestive of an anterior progression of more complex information 14 . Successively more anterior areas integrate inputs from posterior ones, appearing to have little specialization for stimulus features or sensorimotor mappings but have the capacity to be activated by higher order cognitive operations, suggesting a "rostral-caudal axis of cognitive control" 15