Spikes Versus BOLD: What Does Neuroimaging Tell Us About Neuronal Activity? David J

Spikes versus BOLD: what does neuroimaging tell us about neuronal activity? David J. Heeger, Alex C. Huk, Wilson S. Geisler and Duane G. Albrecht

By demonstrating that fMRI responses in human MT+ increase linearly with motion coherence and comparing these responses with slopes of single-neuron firing rates in monkey MT, a new paper provides the best evidence so far that fMRI responses are proportional to firing rates.

Functional magnetic resonance imaging to infer something about the changes in signal5, and two recent studies have (fMRI) is a noninvasive technique for mea- underlying neuronal activity. demonstrated a qualitative match .com suring changes in cerebral blood flow and The vascular source of the fMRI sig- between fMRI measurements in human oxygenation that reflect the underlying nal places important limits on the use- visual cortex and multi-unit electrode neuronal activity. This new technology fulness of the technique. At best, the recordings in homologous areas of the offers a powerful method for exploring the fMRI signal (at each image position and macaque monkey cortex6,7. However, the osci.nature neuronal basis of human cognition, per- temporal frame) would be proportional fMRI signal might reflect not only the fir- ception, and behavior, but its ultimate suc- to the local firing rate, averaged over a ing rates of the local neuronal population cess will depend to a large extent on the small region of cortex and a short period but also subthreshold activity, for exam- relationship between the fMRI signal (the of time. If that were true, then it would ple, simultaneous excitation and inhibi- http://neur ¥ blood-oxygenation level dependent or allow us to make direct inferences about tion that would not result in an action BOLD signal) and the underlying spiking firing rates from fMRI data. There is potential but which would nevertheless activity of neurons. Despite the many fMRI some indirect empirical support for a deplete blood oxygen. If that were true, papers appearing in top journals, we still proportional relationship between aver- the interpretation of fMRI data would be have only a rudimentary understanding of age firing rates of neurons and the fMRI confounded. this relationship. Although it is known that the fMRI signal is triggered by oxygen depletion due to the metabolic demands of Motion stimulus 0% Coherence 50% Coherence 100% Coherence increased neuronal activity, the details of 2000 Nature America Inc.

this process are only partially under- © stood1–3. In this issue of Nature Neuro- science, Rees et al.4 have taken a major step forward in establishing a quantitative link between neuronal responses and fMRI sig- Responses of nals; their findings offer the most com- pelling support to date for the hypothesis MT neurons that fMRI responses are directly proportional to average neuronal firing rates. In a typical fMRI experiment, one col- lects a time-series of images while a stimulus or cognitive task is systematically varied. If the stimulus or task variations preferred direction evoke a large enough change in metabolic demand in a certain brain region, then Average neuronal the image intensity in that region will response modulate over time about its mean intensity value. From these changes, which are typically less than 5%, researchers hope spikes/s

David Heeger and Alex Huk are at the Fig. 1. Average firing rate in monkey MT increases linearly with motion coherence. Top, coherent Department of Psychology, Stanford University, motion stimuli. Middle, responses of four hypothetical MT neurons. At 0% coherence (column 1), Stanford, California 94305, USA. Wilson all 4 neurons respond with the same baseline firing rate. At 50% and 100% coherence (columns 2 Geisler and Duane Albrecht are at the and 3), the MT neuron that prefers upward motion increases its firing rate, the responses of neu- Department of Psychology, University of Texas, rons with rightward and leftward preferred directions do not change their firing rates appreciably, Austin, Texas 78712, USA. and the neuron that prefers downward motion decreases its firing rate slightly. Bottom, average e-mail: [email protected] neuronal activity increases linearly with motion coherence.

Fig. 2. fMRI responses in human 2.8 fits and found that whereas the linear fit V1 are proportional to average 1.0 single neurons accounted for a statistically significant firing rates in monkey V1. Data 2.4 points, human V1 activity as a proportion of the variance in the data, the 0.8 function of stimulus contrast, ctivity 2.0 higher-order polynomials did not fit signal)

averaged across subjects and g nificantly better. The reader might be averaged across two different 0.6 1.6 slightly disappointed, however, that the stimulus patterns (adapted from 1.2 graphs in Figs. 4 and 5 of their paper ref. 10). Error bars, ± 1 s.e.m. 0.4 show only the best-fit 3rd-order polyno-

fMRI Response fMRI

Red curve, average firing rate in BOLD si (% (spikes/s/neuron) 0.8 mials, not the actual measurements of monkey V1, as a function of con- Average Neural A Neural Average 0.2 cortical activity. trast, estimated from a large 0.4 Because of the simple proportional database of microelectrode recordings (333 neurons)11. The 0 0 relationship observed, Rees et al. were ordinates were scaled to obtain 0 10203040506070 able to compare the slopes of the average the best match between the single neuron’s firing rates with their Contrast (%) fMRI responses and the average fMRI data. Assuming that monkey MT firing rates. The average firing rates were determined in the following fashion. Responses as a function and human MT+ are truly homologous of contrast were measured using drifting sine wave gratings of the optimal spatiotemporal frequency, and that firing rates are similar in the two orientation and direction of motion. The responses for each neuron were fitted with a descriptive species, they estimated that a change in function, and the smooth curves were summed across the population. After subtracting the total the fMRI signal of 1% corresponds to a .com maintained activity and dividing by the number of neurons, the average response was then scaled, change in average firing rate of based upon the measured distributions of selectivity of V1 neurons for spatial frequency, temporal frequency, orientation and direction of motion11. The average response for the contrast-reversing plaid 9 spikes/second per neuron. patterns used in ref. 10 was determined by adjusting for the differences in spatiotemporal contrast Motivated by this report, we per- between sine waves and contrast-reversing plaids; this adjustment contributed another scaling factor, formed a similar analysis on some of our osci.nature but had almost no effect on the shape of the average contrast response. The average maximum previously reported measurements of the response of V1 neurons in our sample for optimal drifting sine waves is approximately 45 spikes per activity in human and monkey primary second. The relatively low average response rates in Fig. 2 are the result of the high degree of selec- visual cortex (V1). In our fMRI experi- tivity of V1 neurons. A contrast-reversing plaid pattern of a given spatial frequency and orientation ments, we measured human V1 activity http://neur

¥ produces a response in only a small minority of V1 neurons. Therefore, given that most of the cells are as a function of stimulus contrast for a not responding, the average spike rate per neuron for the population as a whole is rather low. particular stimulus pattern10. In our electrophysiological experiments, we measured the responses of macaque V1 neurons to stimuli of varying contrasts, Rees and colleagues measured fMRI extensively characterized8,9. Figure 1 (mid- orientations, directions and spatiotem- responses in human MT+ (the MT com- dle) shows the responses of four hypo- poral frequencies11. From this large data- plex, also known as V5), a motion-respon- thetical MT neurons, each with a different base of individual V1 neurons, we sive cortical region that is widely thought preferred motion direction, as a function estimated the average firing rate in mon- 2000 Nature America Inc.

to be homologous to monkey cortical areas of motion coherence in the upward direc- key V1, as a function of contrast, for the © MT (also known as V5) and MST. Neurons tion. Across the full range of motion specific stimulus pattern that had been in macaque MT are direction selective; they coherences (from 0% to 100% coherence), used in the fMRI experiments. increase their firing rates when a stimulus the responses of the upward-preferring Our results, like those of Rees et al., moves in a certain ‘preferred’ direction. neuron increase linearly with motion imply a proportional relationship between Their responses are suppressed when the coherence, the responses of the rightward- fMRI response and average firing rate. same stimulus moves opposite to the pre- and leftward-preferring neurons do not The red curve in Fig. 2 represents the aver- ferred direction. Motion in directions change appreciably, and the responses of age neuronal activity (left ordinate), and orthogonal to this axis has little effect on the downward-preferring neuron decrease the data points represent the fMRI mea- the firing rate. Across the population of linearly. Crucially, the slope with which surements (right ordinate). As is evident neurons in MT, different neurons have dif- the upward neuron’s responses increase is in the figure, the two data sets are strik- ferent preferred directions, so that differ- steeper than the slope with which the ingly similar; both increase steeply at low ent directions of motion stimulate different downward neuron’s responses decrease. contrasts and then saturate (level off) at subpopulations of neurons. Because of this asymmetry, the average high contrasts. Assuming again that MT neurons are also sensitive to the neuronal response across these four hypo- humans and monkeys show comparable strength, or coherence, of the motion sig- thetical neurons increases linearly with firing rates in V1, we estimate the constant nal. Using a stimulus that consists of a coherence (Fig. 1, bottom). of proportionality to be approximately field of moving dots, motion coherence Rees et al. reasoned that if the fMRI 0.4 spikes/second per neuron for each 1% can be varied parametrically from 0% signal were proportional to average firing change in the fMRI signal (see legend). (each dot moves in a random direction) rate, then it too would increase linearly Thus, in two different cortical areas (V1 to 100% (all dots move in the same direc- with motion coherence. They then mea- and MT), using different stimulus condi- tion). A motion coherence of 50% means sured fMRI responses in human MT+ tions, different behavioral protocols and that 50% of the dots move in the same across a range of motion coherences and different data analysis methods, both we direction while the other 50% move in confirmed that the fMRI responses and Rees et al. found that fMRI responses random directions (Fig. 1, top). increased linearly with motion coherence. are proportional to average firing rates The responses of MT neurons to such Technically, they compared a linear fit to (although with different proportionality stochastic motion stimuli have been the data with higher-order polynomial constants, 0.4 versus 9).

There are many possible explanations celed out by the decreased activity in the major step forward in bridging this gap, for the discrepancy between the two esti- inhibited postsynaptic cells. We would allowing neuroscientists to move beyond mates of the proportionality constant, argue, however, that this is unlikely for sev- colored spots on pictures of brains toward including differences in the methodology eral reasons. First, we have observed a qual- quantitative measurements of neuronal as well as actual differences in the physi- itative match between fMRI measurements activity in meaningful units of spikes/sec- ology between the two cortical areas. It is and average firing rates even under condi- ond per neuron. not likely that the differences in the fMRI tions that seem to involve strong inhibito- 1. Shulman, R. G. & Rothman, D. L. Proc. Natl. signals are due to differences in the neu- ry interactions6. Second, because excitatory Acad. Sci. USA 95, 11993–11998 (1998). ronal firing rates. The response ranges of and inhibitory neurons are tightly inter- 2. Buxton, R. B., Wong, E. C. & Frank, L. R. the cortical neurons in the two areas are connected in the local cortical circuits, it is Magn. Reson. Med. 39, 855–864 (1998). quite similar: the average maximum fir- unlikely that one could observe a large 3. Vanzetta, I. & Grinvald, A. Science 286, ing rates are 40 spikes/second for MT8 and increase in inhibition without a concomi- 1555–1558 (1999). 45 spikes/second for V111, and the tant increase in excitation. After all, for an 4. Rees, G., Friston, K. & Koch, C. Nat. Neurosci. strongest stimuli used in the two studies inhibitory neuron to increase its firing rate, 3, 716–723 (2000). (100% coherence and 100% contrast it must be receiving more excitatory input, 5. Boynton, G. M., Engel, S. M., Glover, G. H. & Heeger, D. J. J. Neurosci. 16, 4207–4221 respectively) would be expected to evoke and most of the excitatory input comes (1996). 13,14 these maximum rates. The difference in from the local cortical neighborhood . 6. Heeger, D. J., Boynton, G. M., Demb, J. B., the proportionality constant may be relat- Hence, excitation and inhibition are likely Seidemann, E. & Newsome, W. T. J. Neurosci. ed to the fact that the ranges of the fMRI to be balanced so that the firing rates of 19, 7162–7174 (1999). .com signal measured in the two laboratories excitatory and inhibitory neurons will 7. Wandell, B. A et al. Neuron 24, 901–909 differ by a factor of ten: 0.25 % in MT+ increase and decrease in unison15. To what (1999). for 100% coherence and 2.4% in V1 for extent these assumptions hold true for 8. Britten, K. H., Shadlen, M. N., Newsome, 70% contrast. Regardless of what the final other brain regions is an open question, W. T. & Movshon, J. A. Vis. Neurosci. 10, osci.nature 1157–1169 (1993). explanation may turn out to be, the and it will be important to find out 9. Britten, K. H. & Newsome, W. T. important finding that the fMRI signal whether the linear relationship between J. Neurophysiol. 80, 762–770 (1998). and the neuronal spike rate appear to be spike rates and fMRI signals applies to sub- 10. Boynton, G. M., Demb, J. B., Glover, G. G. & linearly related. cortical structures. Heeger, D. J. Vision Res. 39, 257–269 (1999). http://neur ¥ To fully understand why the fMRI sig- Functional MRI offers an unprece- 11. Geisler, W. S. & Albrecht, D. G. Vis. Neurosci. nal is proportional to average firing rate dented opportunity to study the neuronal 14, 897–919 (1997). will require a great deal of further research, underpinings of human behavior. (There 12. Wong-Riley, M. T. T. in Cerebral Cortex but the explanation might be based on the is something particularly captivating about Volume 10: Primary Visual Cortex in Primates (eds. Peters, A. & Rockland, K. S.) 141–200 following observations. First, it is reason- coming into the lab each day to literally (Plenum, New York, 1994). able to assume that the cerebral blood flow watch your own brain at work.) The full 13. Braitenberg V. & Schuz, A. Anatomy of the response is linked to metabolic demand. promise of this technique will not be real- Cortex (Springer, Berlin, 1991). Second, synaptic activity is responsible for ized, however, unless fMRI measurements 14. Peters, A. & Sethares, C. J. Comp. Neurol. 306, nearly all of the energy metabolism in the can be specifically related to conventional 1–23 (1991). 2000 Nature America Inc. cortex, both for neurotransmitter recy- electrophysiological measurements of neu- 15. Shadlen, M. N. & Newsome, W. T. J. Neurosci. © cling1 and for repolarizing the dendritic ronal activity. Rees et al. have taken a 18, 3870–3896 (1998). membranes12. Third, synaptic activity is of course highly correlated with the firing rates of the presynaptic neurons. Finally, the majority of synapses (both excitatory and inhibitory) can be traced to a local Growth of the NMDA network of connections originating in the nearby cortical neighborhood, leaving only receptor industrial complex a small minority of inputs from more Morgan Sheng and Sang Hyoung Lee remote cortical and subcortical struc- tures13,14. Taken together, these observa- Grant et. al. use mass spectroscopy to identify new components of tions imply that there should be a functional relationship between average the NMDA receptor-associated protein complex, revealing a more firing rates and fMRI responses. The complex postsynaptic organization than previously thought. results reported by Rees et al., as well as our own results (Fig. 2), suggest that this The NMDA subtype of glutamate recep- plasticity. Neurobiologists have been try- relationship is linear: an increase in aver- tor is critical in excitatory synaptic trans- ing to unravel the molecular mecha- age firing rate causes a proportional mission and activity-dependent synaptic nisms of NMDA receptor signaling for increase in local synaptic activity, which many years, most recently by analyzing causes a proportional increase in meta- The authors are in the Department of the specific intracellular signaling pro- bolic demand, which causes a proportional Neurobiology and Howard Hughes Medical teins that are physically associated with 1 change in the vascular response. Institute, Massachusetts General Hospital and NMDA receptors . In this issue of Nature One might think that synaptic inhibi- Harvard Medical School, 50 Blossom Street, Neuroscience, Grant and colleagues2 have tion would complicate matters because, for Boston, Massachusetts 02114, USA. combined the latest mass spectrometry example, the energy required to recycle e–mail: [email protected]; (MS) techniques with large-scale inhibitory neurotransmitter might be can- [email protected] immunoblotting (looking at 106 candi-

nature neuroscience • volume 3 no 7 • july 2000 633