neuroscience

The Path to Scalable Psychiatric Gene Therapy and a Future of Cures for Widespread Mental Illnesses — Restoring Joy to a Billion Lives

By No Comments
The Path to Scalable Psychiatric Gene Therapy and a Future of Cures for Widespread Mental Illnesses — Restoring Joy to a Billion Lives

PDF version: The Path to Scalable Psychiatric Gene Therapy and a Future of Cures for Widespread Mental Illnesses – By Logan Thrasher Collins

The Promise of Psychiatric Gene Therapy

Current psychiatric interventions remain insufficient to address the highly prevalent mental illnesses which plague more than a billion people (1 in 7) across the world.1 Widespread and debilitating diseases like major depressive disorder (MDD),2 anxiety disorders,3 schizophrenia,4 bipolar disorders,5 post-traumatic stress disorder (PTSD),6 substance abuse disorders,7,8 and personality disorders9,10 pose an enormous global health burden and are one of the most central causes of human suffering. Patients frequently do not respond to pharmacological interventions for these conditions, resulting in vast numbers of people struggling through life without options for proper management. As a result, at least 800K people die by suicide annually.11 Today’s neuropharmacology industry employs small molecule drugs which modulate neurochemical states, utilizing strategies like neurotransmitter reuptake inhibition and receptor agonism. Mechanistic underpinnings of many neuropharmacological treatments are not well understood.2,12 Furthermore, small molecules suffer from extensive off-target binding13 and frequently come with side effects, many of which can be extremely debilitating and/or dangerous. Though it has been cemented into place by partial successes,12 the way we currently treat mental illnesses is woefully inadequate.

Why might gene therapy solutions eventually offer better treatment options for common psychiatric disorders compared to traditional small molecule pharmaceuticals? Gene therapies possess a number of intrinsic advantages like cell-type-specific targetability, direct in situ expression of therapeutic proteins or RNAs, and the capacity to dynamically respond to environmental conditions. They also have potential for spatiotemporal programmability via emerging sonogenetics14–17 and chemogenetics18 approaches. (Sonogenetics has particular promise, which will be discussed in more detail later). In addition, gene therapies may be engineered to downregulate (RNAi or CRISPRi)19 or even ablate (CRISPR knockout)20 expression of almost any gene in the genome, offering an unprecedented array of new therapeutic targets. CRISPRa might also be employed to upregulate target genes without altering the genome.19 Engineering gene therapies which express multiple proteins or RNAs at once may synergistically improve efficacy.21,22 Importantly, gene therapies can be engineered to persist for long periods of time23,24 or to only provide a burst of short term expression. Depending on the genetic payload, one or the other of these durations may represent the most optimal choice. Gene therapies altogether provide a much larger space of possibilities for precision alteration of brain states than has been possible for small molecule treatments. I would argue that this space’s capabilities remain severely underexplored primarily because of a lack of delivery system capabilities.

Better Delivery Systems are Needed

Gene therapy promises to open a new world of precision psychiatric treatments, yet its potential has gone unrealized. This makes sense as there are a number of obstacles which render psychiatric gene therapy a difficult target. Among these, the challenges of brain delivery, safety, and manufacturing scalability represent particularly recalcitrant bottlenecks. Adeno-associated virus (AAV) gene therapies have progressed furthest in the brain delivery field. Industry players like 4DMT, Dyno Therapeutics, Apertura Therapeutics, and Capsida Biotherapeutics have made efforts via directed evolution, rational design, and machine learning towards optimizing AAV capsids for blood-brain-barrier (BBB) crossing efficiency. However, safety concerns stemming from several patient deaths in systemically administered AAV therapies over the past few years have slowed this progress. Additionally, limitations in AAV manufacturing capacity pose a problem for scaling the vector to populations of 1M+ patients.25 This will be discussed in more detail further on. It should be noted as well that AAVs are limited by their small DNA packaging capacity of 4.7 kb. To unlock the potential of psychiatric gene therapy, we need safer and more scalable delivery systems.

Although there exist multiple obstacles to overcome before gene therapy can realize its potential as a psychiatric modality, I propose a lack of delivery systems represents a foundational missing piece. Without strong delivery vehicles to feasibilize solutions, the field of psychiatric gene therapy will not be credible enough to receive substantial investment. It is a “tools problem”. As mentioned earlier, there is a particular need for vectors which at once possess high safety, scalability, and efficacy. I strongly suspect that the emergence of vectors with these qualities would seed an explosion of efforts towards gene therapies for brain diseases, which would eventually allow us to tackle psychiatric ailments. While psychiatric diseases are unlikely to represent the initial targets of brain gene therapies, opening the door to brain delivery will in my view be necessary to take steps in the direction of modernizing psychiatry through precision genetic medicines.

Examination of the Current Landscape

As mentioned earlier, AAV gene therapies are the current frontrunner for brain delivery yet possess both scalability and safety limitations. I will explain the scalability issue using publicly available information on AAV manufacturing: Final yields (after purification) of AAVs have been reported or modeled in scientific literature with values ranging from around 7.5×1012 vg/L to 7.5×1013 vg/L.26–28 I will thus assume here that 5×1013 vg/L is a typical yield. For this rough calculation, I will also assume that the yield scales linearly with bioreactor volume. As such, a 2,000 L bioreactor would make batches of around 1017 vg and a 200 L bioreactor would make batches of around 1016 vg. According to a 2024 report, a 200 L AAV production run at cGMP quality costs about $2M (including analytics).29A 2022 meta-analysis study of clinical AAV doses states that per patient systemic delivery amounts range from 3.5×1013 vg total to 1.5×1017 vg total.30 Huang et al.’s highly promising AAV BIhTFR1 capsid (the basis for Apertura Therapeutics) was originally administered to mice at a higher dose of 1014 vg/kg and a lower dose of 5×1012 vg/kg.31 Generously (perhaps too generously) assuming that the lower dose is sufficient, this would equate to about 4×1014 vg total in an average 80.7 kg North American adult human.32 Dividing 1016 vg from a $2M (200 L) batch by 4×1014 vg per dose, this means each batch would provide 25 doses for about $80,000 each. Even if substantial improvements in manufacturing yield and in lowering required dosage happen, I am skeptical that systemic AAV approaches will scale to disease indications with 1M+ patients. Despite this, AAVs remain still a central point in the gene therapy industry for a reason and paradigm-shifting approaches to manufacturing25 and/or efficacy might still change the current scalability challenges.

Another existing modality is transient focused ultrasound BBB opening (BBBO). I would argue that BBBO is extremely promising for some applications but not a universal solution. Treatment of many brain diseases necessitates brain-wide delivery. By contrast, BBBO is generally a localized delivery technique.33 Although some common psychiatric diseases fit these parameters, most common ailments with clean clinical endpoints do not. Some work has been done to extend BBBO ultrasound to larger-volume delivery through multiple sonication34,35 or raster scanning,36 but this remains much less well-developed by comparison to localized BBBO approaches and may exhibit greater safety concerns. Indeed, while single-site BBBO possesses a fairly strong safety profile, there is still evidence it can cause problematic inflammatory responses and occasional microhemorrhages.37–39 Also, the level of risk may increase if delivery of vectors with large diameters (e.g. 100 nm) is needed.40,41 As BBBO involves a device, an injection of microbubbles, a procedure, and its own set of safety concerns, it adds complexity which can increase regulatory burden. But I do not think BBBO should be discounted. In some situations, it possesses enormous advantages. These situations may indeed include potential treatments for certain psychiatric disorders. Though BBBO does not universally solve the problem of safe and scalable delivery, I expect it may still play a major role in the field of brain gene therapy.

Intranasal delivery represents a highly promising alternative to intravenous injections which maintains minimal invasiveness. It circumvents the BBB by allowing delivery vectors to migrate through the olfactory (and to a lesser degree trigeminal) nerves into the brain.42,43 This minimizes toxicity by vastly reducing exposure of peripheral organs to the delivery vector. Additionally, much lower doses of delivery vector can be used for intranasal delivery, which might bring AAVs back into the equation as a potentially scalable option. The main drawback of the intranasal route is that the vast majority of delivered DNA accumulates in the olfactory bulb and adjacent brain regions.42,44 Roughly, as the distance from these regions increases, the amount of DNA delivered decreases.44 In a study by Chukwu et al., intranasal delivery of AAV9 was shown to achieve 15% transduction efficiency and 9% gene expression efficiency on average across the brain compared to intravenous delivery.44 Remarkably, this intranasal delivery decreased exposure of peripheral organs by a factor of 13,400 compared to intravenous injection. It should be noted that AAV9 has limited BBB crossing efficiency compared to optimized capsids like AAV BIhTFR1.31 Indeed, intravenous AAV BIhTFR1 transduces brain 40-50 times more efficiently than intravenous AAV9 in humanized TfR1 mice. Yet overall, I would speculate that novel intranasal delivery systems have strong potential for safer and more scalable gene therapies. The intranasal route deserves serious consideration.

Translational Strategies

The path to psychiatric gene therapy may require a detour focusing on “easier” high-prevalence brain disease indications with more clearly defined clinical endpoints. This detour will allow the field to consolidate, cultivating enough successes to justify the financial risk of pursuing psychiatric diseases. Additionally, manufacturing, regulatory, and clinical infrastructure for brain gene therapy in large patient populations may establish itself in this way. That said, I do think it would be beneficial for companies to pursue psychiatric indications in parallel. Even if these initial attempts do not pan out, they may help strengthen the field’s knowledge base and infrastructure. What are some high-prevalence brain indications with clear clinical endpoints which represent strong potential targets for early brain delivery? I will start by nominating stroke, Parkinson’s disease (PD), and epilepsy (open to suggestions here). Though it will by no means be easy to develop efficacious gene therapies for such conditions, I remain optimistic that this foundation of successes in brain treatment is attainable.

In my view, sonogenetic genes have immense potential as payloads for psychiatric gene therapy. Sonogenetics broadly speaking involves delivery of genes encoding mechanosensitive proteins, often ion channels.45 The mechanosensitive proteins change state (e.g. channel opening) in response to ultrasound waves, allowing neuromodulation through transcranial focused ultrasound (tFUS). Sonogenetic gene therapy should thus enable both millimeter-scale spatial resolution and cell-type-specific targeting for neurostimulation, offering unprecedented possibilities for treatment of psychiatric diseases.14–17 This technological convergence could radically transform how mental illness is treated. But delivery nonetheless remains among the most central challenges which must be overcome before sonogenetics reaches clinical feasibility. Neuromedicine cannot explore sonogenetic therapies without a strong foundation of enabling delivery systems.

From a translational perspective, one of the greatest strengths of sonogenetics is that the different effects of stimulating a chosen neuronal cell type across numerous human brain regions may rapidly be tested. While this could have benefits at the preclinical level as well, the most important outcomes will likely occur during clinical stage testing. As an example, consider an anxiety disorder patient who has received a gene therapy which expresses mechanosensitive ion channels in a brain-wide fashion across GABAergic neurons. A clinician might leverage tFUS stimulation in the patient’s lateral amygdala for a few weeks.46,47 If that strategy did not improve the patient’s symptoms, the clinician may easily switch the tFUS stimulation to the central amygdala region46 or the bed nucleus of the stria terminalis (BNST)48 or the lateral septum.49 Many distinct brain regions could be explored without needing to develop a new therapeutic. This would allow fast clinical screening of strategies for modulating neural circuits towards better mental health. Since the challenges of clinical trials represent a massive limiting factor for therapeutics in general, the ability to quickly explore this space of possibilities could dramatically accelerate discovery. Combining the incredible precision of sonogenetic tFUS with such a rapid screening strategy may reveal superior therapeutic targets for combating mental illness.

Conclusion

As I have discussed in this essay, a lack of scalable delivery systems is a central roadblock to the promise of psychiatric gene therapy. Because of this, I have begun developing strategies for overcoming the scalable delivery problem through a number of ideas. I have done early-stage experiments towards a couple of these ideas while others remain in the ideation stage. (I cannot publicly describe the details of my ideas on the internet because public disclosure of IP precludes patentability). I should note that I am currently a graduate student and expect to complete my PhD in around six months. This article represents part of my efforts to lay the groundwork for my upcoming scalable brain delivery plans.

Mental illness represents one of the most profound challenges facing humanity. It affects how we live our lives and interact with the world. It affects people we love. It takes away precious time from people who otherwise could have been experiencing the extraordinary beauty of life and the universe. I believe that positive emotional experiences represent the most fundamental form of value in the cosmos. Mental illness blocks people off from experiencing joy, which is in my view an immeasurable tragedy that needs to be righted. It is time to do the science needed to reset our brains to live life to the fullest.

If you are interested in discussing anything related to this space, please reach out to: logan (dot) phospholipid (at) gmail (dot) com!

If you would like to read more about me and my scientific and entrepreneurial background, please check out my bio at: https://logancollinsblog.com/

References:

1.        Mental disorders (World Health Organization). https://www.who.int/news-room/fact-sheets/detail/mental-disorders (2025).

2.        Marx, W. et al. Major depressive disorder. Nat. Rev. Dis. Prim. 9, 44 (2023).

3.        Craske, M. G. et al. Anxiety disorders. Nat. Rev. Dis. Prim. 3, 17024 (2017).

4.        Leucht, S. et al. Schizophrenia. Nat. Rev. Dis. Prim. 11, 83 (2025).

5.        Vieta, E. et al. Bipolar disorders. Nat. Rev. Dis. Prim. 4, 18008 (2018).

6.        Yehuda, R. et al. Post-traumatic stress disorder. Nat. Rev. Dis. Prim. 1, 15057 (2015).

7.        Strang, J. et al. Opioid use disorder. Nat. Rev. Dis. Prim. 6, 3 (2020).

8.        MacKillop, J. et al. Hazardous drinking and alcohol use disorders. Nat. Rev. Dis. Prim. 8, 80 (2022).

9.        De Brito, S. A. et al. Psychopathy. Nat. Rev. Dis. Prim. 7, 49 (2021).

10.      Gunderson, J. G., Herpertz, S. C., Skodol, A. E., Torgersen, S. & Zanarini, M. C. Borderline personality disorder. Nat. Rev. Dis. Prim. 4, 18029 (2018).

11.      An, S. et al. Global prevalence of suicide by latitude: A systematic review and meta-analysis. Asian J. Psychiatr. 81, 103454 (2023).

12.      Hyman, S. E. Revitalizing Psychiatric Therapeutics. Neuropsychopharmacology 39, 220–229 (2014).

13.      Roth, B. L. Molecular pharmacology of metabotropic receptors targeted by neuropsychiatric drugs. Nat. Struct. Mol. Biol. 26, 535–544 (2019).

14.      Liu, T. et al. Sonogenetics: Recent advances and future directions. Brain Stimul. 15, 1308–1317 (2022).

15.      Xian, Q. et al. Modulation of deep neural circuits with sonogenetics. Proc. Natl. Acad. Sci. 120, e2220575120 (2023).

16.      Hahmann, J., Ishaqat, A., Lammers, T. & Herrmann, A. Sonogenetics for Monitoring and Modulating Biomolecular Function by Ultrasound. Angew. Chemie Int. Ed. n/a, e202317112 (2024).

17.      Tang, J., Feng, M., Wang, D., Zhang, L. & Yang, K. Recent advancement of sonogenetics: A promising noninvasive cellular manipulation by ultrasound. Genes Dis. 11, 101112 (2024).

18.      Miyakawa, N. et al. Chemogenetic attenuation of cortical seizures in nonhuman primates. Nat. Commun. 14, 971 (2023).

19.      Bendixen, L., Jensen, T. I. & Bak, R. O. CRISPR-Cas-mediated transcriptional modulation: The therapeutic promises of CRISPRa and CRISPRi. Mol. Ther. 31, 1920–1937 (2023).

20.      Doudna, J. A. The promise and challenge of therapeutic genome editing. Nature 578, 229–236 (2020).

21.      Zhao, M., Zhao, Z., Koh, J.-T., Jin, T. & Franceschi, R. T. Combinatorial gene therapy for bone regeneration: Cooperative interactions between adenovirus vectors expressing bone morphogenetic proteins 2, 4, and 7. J. Cell. Biochem. 95, 1–16 (2005).

22.      Won, Y.-W. et al. Synergistically Combined Gene Delivery for Enhanced VEGF Secretion and Antiapoptosis. Mol. Pharm. 10, 3676–3683 (2013).

23.      Wang, J. & Vos, J.-M. H. Infectious Epstein-Barr virus vectors for episomal gene therapy. in Gene Therapy Methods (ed. Phillips, M. I. B. T.-M. in E.) vol. 346 649–660 (Academic Press, 2002).

24.      Alba, R., Bosch, A. & Chillon, M. Gutless adenovirus: last-generation adenovirus for gene therapy. Gene Ther. 12, S18–S27 (2005).

25.      Collins, L. T., Ponnazhagan, S. & Curiel, D. T. Synthetic Biology Design as a Paradigm Shift toward Manufacturing Affordable Adeno-Associated Virus Gene Therapies. ACS Synth. Biol. 12, 17–26 (2023).

26.      Reid, C. A., Hörer, M. & Mandegar, M. A. Advancing AAV production with high-throughput screening and transcriptomics. Cell Gene Ther. Insights (2024).

27.      Cameau, E., Glover, C. & Pedregal, A. Cost modelling comparison of adherent multi-trays with suspension and fixed-bed bioreactors for the manufacturing of gene therapy products. Cell Gene Ther. Insights (2020).

28.      Smith, J., Grieger, J. & Samulski, J. Overcoming bottlenecks in AAV manufacturing for gene therapy. Immuno-oncology Insights (2018).

29.      Gangurde, R. & Winitsky, S. Gene therapy: are high costs and manufacturing complexities impeding progress? https://www.parexel.com/insights/blog/gene-therapy-are-high-costs-and-manufacturing-complexities-impeding-progress (2024).

30.      Au, H. K. E., Isalan, M. & Mielcarek, M. Gene Therapy Advances: A Meta-Analysis of AAV Usage in Clinical Settings. Front. Med. Volume 8-, (2022).

31.      Huang, Q. et al. An AAV capsid reprogrammed to bind human transferrin receptor mediates brain-wide gene delivery. Science (80-. ). 384, 1220–1227 (2024).

32.      Walpole, S. C. et al. The weight of nations: an estimation of adult human biomass. BMC Public Health 12, 439 (2012).

33.      Gorick, C. M. et al. Applications of focused ultrasound-mediated blood-brain barrier opening. Adv. Drug Deliv. Rev. 191, 114583 (2022).

34.      Batts, A. J. et al. A multifunctional theranostic ultrasound platform for remote magnetogenetics and expanded blood-brain barrier opening. Brain Stimul. Basic, Transl. Clin. Res. Neuromodulation 18, 1939–1951 (2025).

35.      Nouraein, S. et al. Acoustically targeted noninvasive gene therapy in large brain volumes. Gene Ther. 31, 85–94 (2024).

36.      Felix, M.-S. et al. Ultrasound-Mediated Blood-Brain Barrier Opening Improves Whole Brain Gene Delivery in Mice. Pharmaceutics vol. 13 1245 at https://doi.org/10.3390/pharmaceutics13081245 (2021).

37.      McMahon, D. & Hynynen, K. Acute Inflammatory Response Following Increased Blood-Brain Barrier Permeability Induced by Focused Ultrasound is Dependent on Microbubble Dose. Theranostics 7, 3989–4000 (2017).

38.      Kovacs, Z. I. et al. Disrupting the blood–brain barrier by focused ultrasound induces sterile inflammation. Proc. Natl. Acad. Sci. 114, E75–E84 (2017).

39.      Patwardhan, A. et al. Safety, Efficacy and Clinical Applications of Focused Ultrasound-Mediated Blood Brain Barrier Opening in Alzheimer’s Disease: A Systematic Review. J. Prev. Alzheimer’s Dis. 11, 975–982 (2024).

40.      Chen,  Hong & Konofagou,  Elisa E. The Size of Blood–Brain Barrier Opening Induced by Focused Ultrasound is Dictated by the Acoustic Pressure. J. Cereb. Blood Flow Metab. 34, 1197–1204 (2014).

41.      Shumer-Elbaz, M. et al. Low-frequency ultrasound-mediated blood-brain barrier opening enables non-invasive lipid nanoparticle RNA delivery to glioblastoma. J. Control. Release 385, 114018 (2025).

42.      Patel, D., Patel, B. & Wairkar, S. Intranasal delivery of biotechnology-based therapeutics. Drug Discov. Today 27, 103371 (2022).

43.      Crowe, T. P., Greenlee, M. H. W., Kanthasamy, A. G. & Hsu, W. H. Mechanism of intranasal drug delivery directly to the brain. Life Sci. 195, 44–52 (2018).

44.      Chukwu, C., Yuan, J. & Chen, H. Intranasal versus intravenous AAV delivery: A comparative analysis of brain-targeting efficiency and peripheral exposure in mice. Gene Ther. (2025) doi:10.1038/s41434-025-00585-y.

45.      Legon, W. & Strohman, A. Low-intensity focused ultrasound for human neuromodulation. Nat. Rev. Methods Prim. 4, 91 (2024).

46.      Babaev, O., Piletti Chatain, C. & Krueger-Burg, D. Inhibition in the amygdala anxiety circuitry. Exp. Mol. Med. 50, 1–16 (2018).

47.      Benarroch, E. E. The amygdala: Functional organization and involvement in neurologic disorders. Neurology 84, 313–324 (2015).

48.      Giardino, W. J. & Pomrenze, M. B. Extended Amygdala Neuropeptide Circuitry of Emotional Arousal: Waking Up on the Wrong Side of the Bed Nuclei of Stria Terminalis. Front. Behav. Neurosci. Volume 15–2021, (2021).

49.      Wang, D. et al. Lateral septum-lateral hypothalamus circuit dysfunction in comorbid pain and anxiety. Mol. Psychiatry 28, 1090–1100 (2023).

Amygdala Structure, Function, and Clinically Relevant Pathways

By No Comments
Amygdala Structure, Function, and Clinically Relevant Pathways

PDF version:

Amygdala Structure, Function, and Clinically Relevant Pathways – by Logan Thrasher Collins

Anatomy

The amygdala consists of nuclei which can be grouped into (i) the basolateral nuclear group (BLA), (ii) the superficial cortex-like laminated region (sCLR) which contains the cortical nuclei (Co), and (iii) the centromedial nuclear group.1 The BLA consists of the lateral nucleus (LA) and basal nucleus (BA). In turn, the BA consists of the basolateral nucleus and the basomedial nucleus. The centromedial nuclear group consists of the central nucleus (Ce), medial nucleus (Me), and intercalate cell mass (IC). In turn, Ce consists of a lateral (CeL) subdivision and a medial (CeM) subdivision. The centromedial nuclear group (Ce, Me, and IC) along with the bed nucleus of the stria terminalis (BNST) and sublenticular substantia innominata together comprise the centromedial extended amygdala.

The cellular composition of the BLA nuclei and the sCLR’s Co nuclei resembles that of the cerebral cortex in that the majority of the neurons are pyramidal-like glutamatergic cells while the rest are local GABAergic inhibitory interneurons.1 The inhibitory interneurons include parvalbumin-containing neurons which mainly synapse on the soma and proximal dendrites of the pyramidal cells and somatostatin-containing neurons which mainly synapse on the distal dendrites of the pyramidal neurons. By contrast, the composition of the Ce and Me nuclei resembles the striatum in that many of the neurons are similar to GABAergic medium spiny neurons.

Overview of Amygdala Connectivity

Signals flow into the amygdala primarily via synapses in the BLA. Inputs from cortical sensory areas and from the thalamus (relaying subcortical signals) synapse on neurons in the LA.1 These inputs in the LA facilitate coincidence detection and associative learning tying together the sensory cortical representations of the world with subcortical information coming in via the thalamus. The LA pyramidal-like neurons send excitatory signals to the BA’s projection neurons and to the BA’s interneurons. (It should be noted that the subnuclei of the BA are also interconnected with the prefrontal cortex, hippocampus, and striatum). Next, the BA projects glutamatergic inputs to the CeM’s GABAergic projection neurons. These BA glutamatergic projections additionally synapse on the inhibitory interneurons of the IC and the CeL, both of which regulate the CeM neurons. (An additional layer of complexity comes from further inhibitory interneuron circuits within the LA, BA, IC, CeL, and CeM). Finally, the CeM’s GABAergic projection neurons send output signals to the hypothalamus and brainstem.

Sensory inputs to the amygdala’s LA come from several sources.1 Sensory association areas of the temporal cortex carry visual and auditory information. These areas are part of the ventral stream of sensory processing, which encodes analyses of complex features to facilitate face recognition and auditory recognition. The insular cortex, which encodes somatosensory and visceral sensations, also sends inputs to the LA. Subcortical sensory inputs to the LA come via the thalamus. In addition to LA inputs, the CeM receives visceral and nociceptive inputs directly from the pons. The sCLR receives olfactory input from the olfactory bulb as well as from higher olfactory areas.

Interestingly, the amygdala sends outputs back to cortical sensory association areas as well as primary sensory areas.1 These modulate the valence of specific sensory stimuli, which can be thought of as a way to assign emotional value to particular stimuli.

The amygdala has strong bidirectional interactions with the orbitofrontal cortex (OFC).1 In particular, the OFC receives strong inputs from the BA and targets the IC’s GABAergic neurons. The amygdala also interacts with the dorsal anterior cingulate cortex (dACC) and ventral anterior cingulate cortex (vAAC). The BA sends outputs to the dACC while the vACC projects back to the BA. The BA projects to the entorhinal cortex and receives inputs from the hippocampus as well, which may help tie emotional significance of particular events undergoing processing to associated memories. Finally, the amygdala receives subcortical inputs from arousal systems, including basal forebrain cholinergic inputs, ventral tegmental area (VTA) dopaminergic inputs, noradrenergic locus coeruleus inputs, and rostral raphe serotonergic inputs. The amygdala also projects back to all of these neuromodulatory regions and can influence the arousal systems.

CeM outputs to the hypothalamus and brainstem facilitate visceral behavioral responses to fear. These projections trigger various endocrine and autonomic peripheral nervous system responses such as secretion of adrenocorticotropic hormone (ACTH) into the blood and increased activation of the sympathetic nervous system.

Fear Learning

As mentioned earlier, the convergence of cortical inputs and subcortical inputs onto LA neurons facilitates associative learning between neutral stimuli and unpleasant stimuli. The neutral stimulus is often referred to as the “conditioned stimulus” (CS) while the unpleasant stimulus is referred to as the “unconditioned stimulus” (UC). In classical animal studies, the CS might take the form of a neutral sound (e.g. a tone) while the UC is often an electrical shock to the feet. It is variable as to which input pathways carry information about the CS and which input pathways carry information about the UC.2

In auditory fear learning, both the subcortical thalamic pathway afferents and the auditory cortex afferents have been shown to carry sensory CS information into the LA.2 When an animal must discriminate between two distinct CS sounds to learn which sound is associated with a foot shock UC, the auditory cortical pathway is thought to be necessary because plasticity in the auditory cortex facilitates the discriminative learning. Interestingly, the primary auditory cortex has been shown to carry information about complex multifrequency sounds into the LA while the more ventral associative areas of the auditory cortex bring information about simpler tone sounds.

Both the cortical and subcortical pathways have also been shown to carry parts of the UC. In particular, the parabrachial nucleus of the brainstem has been shown to encode nociceptive UC information. To transfer this information, the parabrachial nucleus projects to the CeM and CeL nuclei of the amygdala.2–4 However, the parabrachial nucleus does not project to the LA.2 It remains unknown if there is a separate (probably glutamatergic) input to the LA which carries aversive information for associative fear learning.

It should also be noted that evidence implicates neuromodulatory systems as carrying part of the UC signal during fear learning. Locus coeruleus noradrenergic projections have been shown to contribute about half of the strength of the fear learning signal.5 That is, when silenced during fear learning in rats, a 50% decrease in learned fear occurred. Additionally, a subpopulation of dopaminergic neurons from the VTA which projects to the BA has been shown to contribute about 30% of the strength of the fear learning signal.6 That is, when silenced during fear learning in mice, a 30% decrease in learned fear occurred. Acetylcholine inputs from the basal forebrain into the BLA have also been demonstrated to be necessary for efficient fear learning.7 These neuromodulators may also be released, though to a lesser degree, during fear memory recall. Finally, serotoninergic neurons (especially from the raphe nuclei) have been implicated to sometimes act on the 5-HT1A receptors of GABAergic interneurons of the LA to inhibit fear learning in LA pyramidal cells.1 That said, serotonin can have other effects in the BLA and its influence is not fully understood.8

Amygdala, Emotion, and Anxiety

The amygdala represents a central part of circuits relating to fear and anxiety as well as of circuits of general emotional valence. Elevated amygdala activity with decreased top-down regulation from the vmPFC has been shown in people with higher anxiety.1,9 It is important to note that the vmPFC overlaps with the ACC and OFC, which were discussed earlier. The vmPFC can facilitate the process of fear extinction: the decline of a learned fear via repeated exposure of a neutral CS without the associated aversive UC. As such, decreased functional connectivity between the amygdala and vmPFC is common in people with anxiety disorders.

The vmPFC facilitates fear extinction by sending excitatory input from the OFC to GABAergic neurons in the IC, which then inhibit the BA’s inputs to the CeM. The BA itself also sends excitatory projections up to the vmPFC which can induce the vmPFC’s fear extinction circuits. A distinct group of excitatory neurons in the BA target the dACC, which then sends excitatory projections back to the amygdala’s BLA to facilitate fear learning (in contrast to the vmPFC projections).10,11 Indeed, LTP occurring via this circuit within the dACC contributes to the formation and maintenance of fear memory. In this way, the dACC is a direct part of the learning network which creates fear memories.

Inhibitory interneurons within the amygdala act as important regulators of anxiety responses.12 In the BLA, inhibitory interneurons can suppress the magnitude of anxiety by releasing GABA onto the pyramidal projection neurons. Inhibitory interneurons in the CeL can also constrain the activity of amygdala output projection neurons of the CeM, which leads to decreased fear behavior. But it should be noted that the BLA can receive sensory input associated with either threatening or rewarding stimuli. Because of this, its projection neurons trigger different behavioral responses (threat or reward behaviors) depending on the nature of the stimulus.

There exist non-overlapping populations of putative projection neurons in the BLA which are thought to fire in response to threat and reward stimuli separately.12,13 These populations are thought to develop via the Hebbian associative learning described previously, which leads to formation of fear pathways for some stimuli, but can also promote association of rewarding stimuli with neutral stimuli and thus form learned emotional pathways of positive valence.13 Additionally, inhibitory interneurons of the BLA suppress threat-related projection neurons when reward-related projection neurons are active and vice versa. With anxiety disorders, these interneuron circuits are frequently dysregulated in that negative valence is assigned to neutral or reward stimuli, leading to activation of only the threat-related projection pathway.

Extended Networks of the Amygdala and Anxiety

As has been discussed to some degree so far, the amygdala does not function in isolation. It makes numerous reciprocal connections with other brain areas to facilitate its operation. Some of the most important of these include cortical regions like the vmPFC, OFC, and ACC, which were discussed earlier. But extended subcortical structures like the BNST and hippocampus (which have so far only been mentioned briefly) also play major roles.

The BNST is a collection of nuclei nearby to the amygdala which is recruited during sustained fear and anxiety responses.13 It is thought that the BNST specifically activates during prolonged stressful periods of greater than 10 minutes in duration.14 The BLA sends glutamatergic projections into the BNST’s anterodorsal (ad) nucleus. Interestingly, these excitatory inputs to the BNST ad nucleus promote anxiolytic outcomes. Additionally, local inhibition of the ad nucleus from the BNST’s oval (ov) nucleus promotes anxiogenic outcomes. The BNST’s ad nucleus facilitates anxiolytic states by sending its own (predominantly) GABAergic projections to the VTA to increase positive emotional valence, to the lateral hypothalamus (LH) to decrease risk avoidance, and to the parabrachial nucleus of the brainstem to decrease respiration rate.15

BNST ad projections to the VTA are mostly GABAergic neurons synapsing onto VTA inhibitory interneurons.16 It should be noted that there are also ventral BNST (vBNST) GABAergic and glutamatergic projections which synapse onto different populations of VTA inhibitory interneurons, triggering anxiogenic phenotypes and anxiolytic phenotypes respectively.17 A major population of BNST ad projections to the lateral hypothalamus are GABAergic neurons preferentially synapsing onto GABAergic target neurons. Among these is a subpopulation of GABAergic projection neurons targeting GABAergic lateral hypothalamus neurons which also produce orexin (a neuropeptide which stimulates food intake behaviors and promotes wakefulness).16 BNST ad GABAergic projections to the parabrachial nucleus probably inhibit glutamatergic neurons which themselves would otherwise signal for increased respiratory rate.18

The amygdala also interacts with the hippocampus. As mentioned earlier, the BLA sends excitatory inputs to the hippocampal formation by first synapsing at the entorhinal cortex (EC), which then sends its own excitatory inputs to the hippocampus.13 These inputs are necessary for acquisition of contextual fear memories, likely mediated by the BLA amygdala’s fear learning mechanism in combination with hippocampal memory representations.

In addition, the BLA sends glutamatergic synapses directly onto pyramidal cells in the ventral hippocampus (vHPC) CA1 region, increasing anxiety-like behavior when these BLA projections are active. In part, the vHPC mediates its effects on anxiety through glutamatergic projections to the lateral septum, which sends its own projections onwards to the hypothalamus. The vHPC glutamatergic projections stimulate activation of corticotropin releasing factor receptor 2 (CRFR2) expressing GABAergic projection neurons in the lateral septum through a mechanism which is not fully understood.13,19 These GABAergic projection neurons inhibit the anterior hypothalamic area (AHA), which itself inhibits the paraventricular nucleus (PVN) of the hypothalamus as well as the periaqueductal gray (PAG). In this way, the lateral septum disinhibits the paraventricular nucleus and the periaqueductal gray, which leads to neuroendocrine and behavioral outcomes associated with persistent anxiety.

Conclusion

While this writeup serves as an initial primer on the amygdala, there remain a plethora of relevant neural circuits to explore beyond what has been described here. Nonetheless, I hope that the information provided will offer a useful starting point for learning about the amygdala’s structure, function, and effects on mammalian emotions. As further reading, I specifically recommend references #1, #9, #12, and #13.

References

1.      Benarroch, E. E. The amygdala: Functional organization and involvement in neurologic disorders. Neurology 84, 313–324 (2015).

2.      Palchaudhuri,  Shriya, Osypenko,  Denys & Schneggenburger,  Ralf. Fear Learning: An Evolving Picture for Plasticity at Synaptic Afferents to the Amygdala. Neurosci. 30, 87–104 (2022).

3.      Han, S., Soleiman, M. T., Soden, M. E., Zweifel, L. S. & Palmiter, R. D. Elucidating an Affective Pain Circuit that Creates a Threat Memory. Cell 162, 363–374 (2015).

4.      Herry, C. & Johansen, J. P. Encoding of fear learning and memory in distributed neuronal circuits. Nat. Neurosci. 17, 1644–1654 (2014).

5.      Uematsu, A. et al. Modular organization of the brainstem noradrenaline system coordinates opposing learning states. Nat. Neurosci. 20, 1602–1611 (2017).

6.      Tang, W., Kochubey, O., Kintscher, M. & Schneggenburger, R. A VTA to Basal Amygdala Dopamine Projection Contributes to Signal Salient Somatosensory Events during Fear Learning. J. Neurosci. 40, 3969 LP – 3980 (2020).

7.      Jiang, L. et al. Cholinergic Signaling Controls Conditioned Fear Behaviors and Enhances Plasticity of Cortical-Amygdala Circuits. Neuron 90, 1057–1070 (2016).

8.      Bocchio, M., McHugh, S. B., Bannerman, D. M., Sharp, T. & Capogna, M. Serotonin, Amygdala and Fear: Assembling the Puzzle. Front. Neural Circuits Volume 10–2016, (2016).

9.      Zhang, W.-H., Zhang, J.-Y., Holmes, A. & Pan, B.-X. Amygdala Circuit Substrates for Stress Adaptation and Adversity. Biol. Psychiatry 89, 847–856 (2021).

10.    Toyoda, H. et al. Interplay of Amygdala and Cingulate Plasticity in Emotional Fear. Neural Plast. 2011, 813749 (2011).

11.    Jhang, J. et al. Anterior cingulate cortex and its input to the basolateral amygdala control innate fear response. Nat. Commun. 9, 2744 (2018).

12.    Babaev, O., Piletti Chatain, C. & Krueger-Burg, D. Inhibition in the amygdala anxiety circuitry. Exp. Mol. Med. 50, 1–16 (2018).

13.    Calhoon, G. G. & Tye, K. M. Resolving the neural circuits of anxiety. Nat. Neurosci. 18, 1394–1404 (2015).

14.    Hammack, S. E., Todd, T. P., Kocho-Schellenberg, M. & Bouton, M. E. Role of the bed nucleus of the stria terminalis in the acquisition of contextual fear at long or short context-shock intervals. Behavioral Neuroscience vol. 129 673–678 at https://doi.org/10.1037/bne0000088 (2015).

15.    Kim, S.-Y. et al. Diverging neural pathways assemble a behavioural state from separable features in anxiety. Nature 496, 219–223 (2013).

16.    Giardino, W. J. & Pomrenze, M. B. Extended Amygdala Neuropeptide Circuitry of Emotional Arousal: Waking Up on the Wrong Side of the Bed Nuclei of Stria Terminalis. Front. Behav. Neurosci. Volume 15–2021, (2021).

17.    Jennings, J. H. et al. Distinct extended amygdala circuits for divergent motivational states. Nature 496, 224–228 (2013).

18.    Kaur, S. et al. Glutamatergic Signaling from the Parabrachial Nucleus Plays a Critical Role in Hypercapnic Arousal. J. Neurosci. 33, 7627 LP – 7640 (2013).

19.    Anthony, T. E. et al. Control of Stress-Induced Persistent Anxiety by an Extra-Amygdala Septohypothalamic Circuit. Cell 156, 522–536 (2014).

Dopaminergic Neurons in the Ventral Tegmental Area as a Target for Treatment-Resistant Depression

By No Comments
Dopaminergic Neurons in the Ventral Tegmental Area as a Target for Treatment-Resistant Depression

Prelude

This essay is intended as an intuitive examination of a reward system neural circuit which may serve as a useful target for new therapies aimed at fighting treatment-resistant depression (TRD). My purpose here is not to introduce an entirely novel concept, but rather to compile in one place a set of important explanations on how information flow in the reward system relates to TRD and how these reward system mechanisms may have clinical relevance.

Treatment Resistant Depression

Treatment resistant depression (TRD) is a widespread and debilitating condition. Patients with TRD are defined to have failed to adequately respond to two or more treatments for depression.1  As a broader category, depression affects about 280 million people worldwide.2 TRD affects roughly 30% of these patients3 (~84 million). In the USA, it has been estimated that about 2.8 million adults suffer from TRD.1 A common symptom associated with TRD is anhedonia, the inability to feel positive emotions. It is thought that defects in the brain’s reward pathway are central to the neurobiology of TRD since this pathway contains the neural circuitry necessary to encode positive emotional experiences.

Reward Circuits

When sensory recognition of a potential reward occurs, various pathways inhibit activity of the lateral habenula (LHb), which in turn inhibits the rostromedial tegmental nucleus (RMTg). This disinhibits the ventral tegmental area (VTA).4 The VTA’s dopaminergic projections then spike in phasic bursts, sending dopamine to the nucleus accumbens (NAc) (mesolimbic pathway, a part of the medial forebrain bundle or MFB) and prefrontal cortex (PFC) (mesocortical pathway).5 NAc GABAergic medium spiny neurons (MSNs) generally express either the dopamine 1 receptor (D1R) or express the dopamine 2 receptor (D2R). D1R MSNs are excited by dopamine while D2R MSNs are inhibited by dopamine. Mesolimbic inputs bias the NAc to output from the D1R MSNs, which stimulate the direct motor pathway to respond to the reward. The GABAergic MSN activity furthermore inhibits the ventral pallidum (VP), which in turn lifts its own GABAergic inhibition on targets such as the mediodorsal thalamus, lateral hypothalamus, and VTA.6 This increases arousal and helps with motor processes. The mediodorsal thalamus projects to the PFC and triggers circuits that represent the value of the reward.7

These circuits facilitate reward learning by a comparative mechanism called reward prediction error (RPE). The pedunculopontine tegmental nucleus (PPTg) receives inputs about the actual reward from brainstem sensory signals (and other brain areas) and projects glutamatergic and cholinergic synapses into the VTA (and elsewhere) to activate the dopaminergic neurons.8 If the reward is less valuable than expected, the LHb activates, which triggers firing of GABAergic neurons in the RMTg onto the VTA dopamine neurons, shutting down the mesolimbic activity.4 If the actual reward remains valuable, then the mesolimbic activity continues. This process where the inhibitory LHb-RMTg signal is “subtracted” from the stimulatory PPTg signal determines the RPE comparison’s outcome and whether the VTA continues its dopaminergic signals.9 All of this facilitates reward learning, where mesolimbic long-term potentiation (LTP) occurs if the reward is as strong as expected (or stronger) and mesolimbic long-term depression (LTD) (not the same as psychiatric depression) occurs if the reward is not as strong as expected.

Dopaminergic Neurons and Treatment Resistant Depression

As a central driver within the reward system, dopaminergic VTA neurons have high potential as a target for combatting TRD. Activation of these neurons may alleviate anhedonia and increase motivation. There already exists clinical evidence that stimulation of VTA dopaminergic neurons has significant benefits. As mentioned, the mesolimbic pathway projections of VTA dopaminergic neurons make up a major part of the MFB. Multiple clinical studies on deep brain stimulation (DBS) of the MFB (specifically the supero-lateral MFB or slMFB) have shown long-term beneficial effects for patients with TRD.10–12 Functional imaging evidence suggests this works indirectly through activation of descending glutamatergic fibers from the PFC which activate the VTA’s dopamine neurons.10 Dopamine axons themselves are small in diameter, which make them not as responsive to conventional DBS. It should be noted that the VTA is a highly heterogeneous structure with dopaminergic, GABAergic, and glutamatergic neurons,13 so DBS of the VTA in general might have off-target effects and/or partially mitigate the benefits of the stimulation. Activation of the VTA’s GABAergic and glutamatergic neurons can have markedly different effects compared to activation of only its dopaminergic neurons.14 In mice, GABAergic VTA neuronal activity particularly has been found to occur in response to aversive stimuli and stimuli predicting the absence of reward.15,16 In rats, optogenetic stimulation of VTA dopamine neurons promotes motivated behavior while optogenetic stimulation of VTA GABA neurons disrupts reward and promotes aversion.17 Clinical and animal model evidence thus supports the idea that selective activation of VTA dopamine neurons might act as a potent therapy for TRD.

Conclusion

Based on the literature, raising the basal level of VTA dopaminergic neuron activity might demonstrate a strong ameliorative effect on TRD. Extensive preclinical and clinical testing will of course be crucial to establish safety. Possible addictiveness of treatments which activate this circuit will need careful examination in particular. Depending on the modality of treatment, different forms of neurological adaptation may occur, so ways of mitigating this issue should be explored. VTA dopaminergic neurons represent a promising target for next-generation therapies aimed at overcoming TRD.

References

1.        Zhdanava, M. et al. The Prevalence and National Burden of Treatment-Resistant Depression and Major  Depressive Disorder in the United States. J. Clin. Psychiatry 82, (2021).

2.        World Health Organization – Depressive disorder (depression). https://www.who.int/news-room/fact-sheets/detail/depression (2023).

3.        McIntyre, R. S. et al. Treatment-resistant depression: definition, prevalence, detection, management, and investigational interventions. World Psychiatry 22, 394–412 (2023).

4.        Hong, S., Jhou, T. C., Smith, M., Saleem, K. S. & Hikosaka, O. Negative Reward Signals from the Lateral Habenula to Dopamine Neurons Are Mediated by Rostromedial Tegmental Nucleus in Primates. J. Neurosci. 31, 11457 LP – 11471 (2011).

5.        Juarez, B. & Han, M.-H. Diversity of Dopaminergic Neural Circuits in Response to Drug Exposure. Neuropsychopharmacology 41, 2424–2446 (2016).

6.        Root, D. H., Melendez, R. I., Zaborszky, L. & Napier, T. C. The ventral pallidum: Subregion-specific functional anatomy and roles in motivated behaviors. Prog. Neurobiol. 130, 29–70 (2015).

7.        Haber, S. N. & Knutson, B. The Reward Circuit: Linking Primate Anatomy and Human Imaging. Neuropsychopharmacology 35, 4–26 (2010).

8.        Skvortsova, V. et al. A Causal Role for the Pedunculopontine Nucleus in Human Instrumental Learning. Curr. Biol. 31, 943-954.e5 (2021).

9.        Eshel, N. et al. Arithmetic and local circuitry underlying dopamine prediction errors. Nature 525, 243–246 (2015).

10.      Fenoy, A. J. et al. Deep brain stimulation of the “medial forebrain bundle”: sustained efficacy of antidepressant effect over years. Mol. Psychiatry 27, 2546–2553 (2022).

11.      Schlaepfer, T. E., Bewernick, B. H., Kayser, S., Mädler, B. & Coenen, V. A. Rapid Effects of Deep Brain Stimulation for Treatment-Resistant Major Depression. Biol. Psychiatry 73, 1204–1212 (2013).

12.      Fenoy, A. J., Quevedo, J. & Soares, J. C. Deep brain stimulation of the “medial forebrain bundle”: a strategy to modulate the reward system and manage treatment-resistant depression. Mol. Psychiatry 27, 574–592 (2022).

13.      Faget, L. et al. Afferent Inputs to Neurotransmitter-Defined Cell Types in the Ventral Tegmental Area. Cell Rep. 15, 2796–2808 (2016).

14.      Root, D. H. et al. Distinct Signaling by Ventral Tegmental Area Glutamate, GABA, and Combinatorial Glutamate-GABA Neurons in Motivated Behavior. Cell Rep. 32, (2020).

15.      van Zessen, R., Phillips, J. L., Budygin, E. A. & Stuber, G. D. Activation of VTA GABA Neurons Disrupts Reward Consumption. Neuron 73, 1184–1194 (2012).

16.      Tan, K. R. et al. GABA Neurons of the VTA Drive Conditioned Place Aversion. Neuron 73, 1173–1183 (2012).

17.      Tong, Y., Pfeiffer, L., Serchov, T., Coenen, V. A. & Döbrössy, M. D. Optogenetic stimulation of ventral tegmental area dopaminergic neurons in a female rodent model of depression: The effect of different stimulation patterns. J. Neurosci. Res. 100, 897–911 (2022).

Notes on Ultrasound Physics and Instrumentation

By No Comments
Notes on Ultrasound Physics and Instrumentation

PDF version: Notes on Ultrasound Physics and Instrumentation – by Logan Thrasher Collins

Fundamentals of ultrasound waves

Sound waves such as those in ultrasound are longitudinal waves where particles oscillate backwards and forwards along the wave’s direction of propagation. This forms regions of higher (compression) and lower (rarefaction) density. Sound waves occur through the exchange of kinetic energy from molecular movement with the potential energy from elastic compression and stretching of bonds.

The speed of sound c depends on the medium. For example, in air c = 330 m/s while in water c = 1480 m/s. The frequency of a sound wave depends on the source producing it. Frequency is measured in Hertz (Hz) where 1 Hz = 1 cycle/second. Sound waves with frequencies greater than or equal to 20 kHz are referred to as ultrasound. The wavelength of a sound wave is λ = c/f. It is measured in mm (or other units of length). A sound wave’s phase describes what position the wave starts at within a cycle of oscillation and is measured in degrees or radians. Phase shifts are typically measured relative to a phase of 0° (or 0 rad). A sound wave’s amplitude corresponds to its “loudness” and describes the height of the wave’s peaks (or depth of its troughs). Cosine-based or complex exponential equations can describe sound waves in a similar way as they describe electromagnetic waves. These equations include amplitude as a coefficient, frequency as a factor multiplied by time, and phase shift as a term added to time.

Ultrasound pressure, power, and intensity

A sound wave’s excess pressure is the difference between its peak amplitude and the normal ambient pressure of the medium. When a medium is compressed, the excess pressure is positive and when it is rarified, the excess pressure is negative. In practice, the ambient pressure is often quite low and can be ignored, in which case pressure is used rather than excess pressure. Excess pressure and pressure are measured in Pascals (Pa), which are equivalent to N/m2.

When an ultrasound wave passes through a medium, it deposits energy (measured in Joules) into said medium. The rate at which a source produces this energy is the power, which is measured in Watts or J/s. An ultrasound wave’s power is unevenly distributed across the beam and often is more concentrated near the beam’s center. Intensity is a measure of the power flowing through a unit area perpendicular (or normal) to the wave’s direction of propagation and is measured in W/m2 or W/cm2. Intensity is proportional to the wave’s pressure squared (I ∝ p2).

Ultrasound and its medium

As mentioned, the medium (material) an ultrasound wave passes through determines the sound’s speed of propagation. Specifically, the material’s density and stiffness determine the speed of sound propagation. Density ρ is often measured in kg/m3. For instance, bone’s density ρbone = 1850 kg/m3 and water’s density ρwater = 1000 kg/m3 (or 1 g/cm3). Stiffness k describes how much a material resists deformation by force F. It equals the amount of pressure (stress) needed to change the thickness of a material by a given fraction and is measured in Pa. The ratio of change in thickness ΔL over original thickness L0 is called tensile strain ε. The increase (or decrease) in length of a material due to applied force is denoted ΔL = L – L0. Stress σ or fractional change in thickness is given by change in thickness divided by original thickness of the sample. Stress over strain is the elastic modulus E (also called Young’s modulus). Here, A is the cross-sectional area of a sample perpendicular to the applied force. Note that these formulas apply only for tensile stress (which induces tensile strain), the type of stress where a material is compressed or elongated.

The relationship between a sound wave’s speed and the properties of density and thickness can be modeled by masses on springs where ρ = m = density (corresponding to masses in the model) and k = stiffness (corresponding to the spring stiffness in the model). The relationship between c, ρ, and k is given by the following equation.

As a result of these properties, the speed of sound varies in different tissues. A table of the approximate speeds of sound across selected tissues is provided.

Acoustic impedance z measures the response of a medium’s particles in terms of their velocity v to a sound wave of a given pressure p. Acoustic impedance can also be expressed in terms of density and stiffness or in terms of density and the sound’s speed. The latter form of definition is called the characteristic acoustic impedance of the tissue. It has units of kg m–2 s–1, also called rayl.

Ultrasound reflection and refraction

When an ultrasound wave traveling through a medium of acoustic impedance z1 encounters a new medium of different acoustic impedance z2, some of the wave is transmitted and some is reflected back. The equations in the following discussion will initially assume that the wave approaches the interface at a 90° angle (perpendicular) until stated otherwise.

To maintain continuity across the interface, the following equations must hold. Here, pi and vi are initial pressure and velocity, pr and vr are reflected pressure and velocity, and pt and vt are transmitted pressure and velocity.

Additionally, the intensity transmitted It across the interface equals the incident intensity Ii minus the reflected intensity Ir.

By using the equation for the definition of acoustic impedance z = p/v, the following two equations are true. Algebraically manipulating these equations leads to the third equation below. Rp is called the amplitude (or pressure) reflection coefficient of the interface. It is important because it decides the amplitude of the echoes produced at various interfaces within the tissue.

Note that if the acoustic impedance of the first medium is greater than the acoustic impedance of the second medium (i.e. z1 > z2), then Rp is negative and the reflected wave is inverted (across the x axis). For most interfaces between soft tissues, Rp is quite small and so most of the wave is transmitted to produce further echoes deeper within the tissue. This is useful for ultrasound imaging. Since the interface between air and soft tissue has a much larger Rp, the ultrasound source must be placed in direct contact with the patient’s skin to avoid air blocking transmission. This also means that gas-containing tissues like lungs and gut effectively block imaging beyond the region with the gas.

Another way to describe reflection is in terms of the intensity reflection coefficient RI. Since I ∝ p2, this means that RI = Rp2.

Since the incident intensity is Ii = It + Ir and the transmitted intensity is It = Ii – Ir, one can also define an intensity transmission coefficient TI as the first equation below. Since the transmitted pressure is pt = pi + pr, one can also define a pressure transmission coefficient Tp as the second equation below.

It is useful to realize that, because energy at interfaces must be conserved, the following equation holds true.

But ultrasound waves often do not approach interfaces at a 90° angle, so the equations in this section must be modified via trigonometry to account for other angles. First, note that the incident angle θi will equal the backscattered angle θr (this is true in the 90° case as well).

When a sound wave in a less dense tissue (slower sound wave) crosses into a tissue of greater density (faster sound wave), the transmitted wave bends away from the normal and thus θt > θi. 

Likewise, when a sound wave in a tissue of greater density (faster sound wave) crosses into a tissue of less density (slower sound wave), the transmitted wave bends towards the normal and thus θt < θi.

Ultrasound scattering

When ultrasound encounters an object which is small compared to the wavelength, it scatters in all directions, though slightly more energy is typically backscattered towards the transducer than away from it. Scattering specifics depend on the shape, size, and acoustic properties (z, k, ρ) of the object.

When many similar small objects are close together (e.g. red blood cells), constructive interference can occur, which is useful. In the case of the blood, this is the basis for Doppler ultrasound, which measures blood flow. By comparison, when many small objects are far apart, complex interference patterns occur. This leads to a phenomenon in images known as “speckle”, which is usually (but not always) considered a form of undesirable noise.

Absorption and relaxation of ultrasound

As ultrasound moves through tissue, it loses energy due to absorption, resulting in heat. There are two ways this occurs: relaxation absorption and classical absorption. The effects of relaxation absorption are typically much more dominant.

Relaxation absorption depends on the elastic properties of tissue, occurring when the tissue returns to its original state after rarefaction or compression by ultrasound. This is quantified by the relaxation time τ = 1/fr, which is how long the tissue takes to return to its original state after the ultrasound’s effect. 

Relaxation is characterized by a relaxation absorption coefficient βr which is given by the first two equations below. Here, fr is the frequency of relaxation, f is the frequency of ultrasound, and B0 is a material-specific constant. In practice, tissues contain a range of values of τ and fr, so the third equation below is a more general formula where the overall relaxation absorption coefficient is proportional to the sum of the various contributions. Higher values of βr mean more energy is absorbed into the tissue.

It is useful to note that in tissues, the relationship between the relaxation absorption coefficient βr and the frequency f is approximately linear.

Classical absorption is less important in tissues since, at clinical frequencies, the relaxation absorption is strongly dominant as mentioned. That said, overall absorption does consist of a combination of relaxation and classical absorption (though the latter may be approximated away sometimes). Classical absorption occurs because of friction between particles as they are displaced by ultrasound, causing loss of energy to heat. This loss is characterized by the classical absorption coefficient βclass ∝ f2.

Attenuation coefficients

When an ultrasound beam propagates through tissue, the sum of the absorption and scattering is described as attenuation, which causes an exponential decrease in the pressure and intensity of the ultrasound as a function of the propagation distance x through tissue.

The following equations describe the loss of ultrasound intensity and pressure as the wave moves through tissue. Here, µ is the intensity attenuation coefficient and α is the pressure attenuation coefficient. The value of µ is equal to twice the value of α (so, µ = 2α). Both have units of cm–1, though the value of µ is often given in units of decibels (dB) per cm, where the conversion factor is µ(dB cm–1) = 4.343µ(cm–1). It is useful to note that each 3 dB decrease corresponds to a decrease in intensity by a factor of 2.

Approximate frequency dependences of µ are given in the table below. As an example, in soft tissue, the value of µ = 1 dB cm–1 for 1 MHz ultrasound and µ = 2 dB cm–1 for 2 MHz ultrasound. Note that the value of µ for fat is calculated differently than the others via the equation µ(f) = 0.7f1.5 dB.

Ultrasound transducers

Ultrasound imaging is performed using ultrasound transducers. A gated frequency generator first produces short periodic voltage pulses, which are then amplified and fed into the transducer via a transmit-receive switch. Because the transducer transmits high power pulses and receives low intensity signals from the reflected ultrasound waves, the transmit and receive circuits must be isolated from each other. Amplified voltage is converted by a shaped piezoelectric material (typically lead zirconate titanate, which is abbreviated PZT) in the transducer into a mechanical pressure wave which is transmitted into the tissue. 

After reflecting and scattering from boundaries within the tissue, pressure wave signals return to the transducer and are converted back into voltages by the piezoelectric material. The voltages must pass through low-noise preamplifier before digitization. Further amplification and signal processing facilitates display of images on a computer.

When oscillating voltage is applied to one end of shaped PZT material, the thickness of the PZT element oscillates at the same frequency as the voltage. By placing this element in contact with skin, mechanical pressure waves are transmitted into tissue. The element has a resonant frequency f0 which is determined by its thickness T and the speed of the ultrasound wave in the PZT material cPZT. The value of cPZT is ~4000 m/s.

For most ultrasound devices, the transducer element’s thickness must be designed to equal one half the wavelength of ultrasound in the PZT material λPZT, so T = λPZT/2. This facilitates use of the resonant frequency.

Because of the much higher acoustic impedance of PZT material compared to skin (~18 times higher), a large amount of the energy would be reflected if the PZT was placed directly onto the skin’s surface. This would mean the mechanical wave traveling into the tissue would lose most of its energy.

To prevent this energy loss from happening, transducers possess a matching layer with a zmatching value between zskin and zPZT as given by the equation below. The thickness of the matching layer is typically made to be 1/4 of the ultrasound’s wavelength in its material T = λmatching/4. All this improves the transmission and reception efficiency. Sometimes multiple matching layers are used to further improve efficiency.

At the back of a PZT element, there is a damping layer, typically made of some backing material and epoxy. This damping layer prevents the PZT from continuing to oscillate (at a decaying rate) after each voltage pulse. This continued oscillation would blur the boundaries between the short pulses, which can decrease axial resolution (as will be described soon).

Although transducers have a central frequency f0, they typically cover a range of frequencies (e.g. a 3 MHz transducer might cover a range of 1-5 MHz). Higher mechanical damping leads to broader transducer bandwidth. Transducer bandwidth is described as the frequency range over which the sensitivity is greater than half the maximum sensitivity level. The relationship between bandwidth and f0 is often quantified by the quality factor Q, which is the ratio of f0 to the bandwidth. Low values of Q mean larger bandwidths. Note that 2f0 is the second harmonic frequency.

Beam geometry and resolution

FUS transducers produce a very complicated wave pattern close to the face of the transducer (the near-field or Fresnel zone). This complicated pattern is not usually useful since it has many parts where the intensity is zero. Beyond the near-field zone, the wave pattern is much simpler and decays exponentially with distance (the far-field or Fraunhofer zone). The boundary between the two zones is called the near-field boundary (NFB) and occurs at a distance ZNFB away from the face of the transducer. The following equation (where r is the radius of the transducer) can be used to calculate the ZNFB value.

After the NFB, the FUS beam diverges (spreads out laterally) with an angle of deviation θ which is given by the equation below.

For the far-field zone, the lateral shape of the beam approximates a Gaussian function. The full width at half maximum (FWHM) defines the lateral resolution of the beam. It is given by the equation below, where σ is the standard deviation of the Gaussian. This value is unique to each FUS beam at the specific desired depth. It can be calculated by the following equation.

Single element transducers also produce ultrasound side lobes where the first zero of the side lobe at angle φ is the same equation as the FUS beam’s divergence angle equation. In ultrasound imaging, the side lobes can cause artifacts if they are backscattered from tissue outside of the imaged region.

Axial resolution is the closest distance two boundaries can be relative to each other (in a direction parallel to the FUS beam’s propagation) while still allowing them to be resolved as two distinct features. It is given by the equation below where pd is the pulse duration and c represents the speed of the ultrasound in the tissue.

The reason that axial resolution works this way is because the echoes of beams returning from two different boundaries are distinguishable so long as these boundaries are spaced widely enough that they do not overlap in time.

Some typical values of axial resolution are 1.5 mm at a frequency (1/c) of 1 MHz or 0.3 mm at a frequency of 5 MHz. But it should be noted that attenuation of the FUS increases at higher frequencies, so there is an important tradeoff between penetration depth and axial resolution. (Very high frequencies such as 40 MHz can be used for imaging the skin at high resolution).

Single flat ultrasound transducers possess relatively poor lateral resolution. Concave curved transducers can achieve better resolutions. (It should be noted that the transducer equations above may be somewhat altered in the case of curved transducers rather than flat transducers). To make a curved transducer, one can add a curved plastic lens in front of the piezoelectric element or the piezoelectric element itself can be made in a concave curved shape.

The shape of a transducer’s curvature can be described by an “f-number”, which is a value equal to the focal distance divided by the aperture dimension where the aperture dimension is determined by the size of the transducer element.

Lateral resolution for a bowl-shaped curved (focused) transducer is calculated using the following equation below where λ is the ultrasound wavelength, F is the focal distance, and D is the diameter of the transducer. The focal distance F is where the lateral beamwidth is narrowest and is approximated as the radius of curvature (ROC) of the lens or PZT element. This approximation is valid except in the case of very high curvature.

When deciding on the focusing power of a transducer, there is a compromise between high spatial resolution and depth over which good spatial resolution is achievable. For a strongly focused transducer, locations further away from the beam’s focal plane diverge more sharply than for a weakly focused transducer. This can be quantified by the on-axis depth-of-focus (DOF) which equals the axial distance over which the beam’s intensity is at least 50% of its maximum value.

Transducer arrays

Contemporary FUS systems typically use arrays consisting of many small piezoelectric elements (rather than single-element transducers). These arrays allow 2D imaging via electronic steering of the beam through tissue while the transducer is held at a fixed position. Sophisticated electronics produce a dynamically changing focus during pulse transmission and signal reception, which maintains high resolution throughout the image. Linear and phased array transducers represent the two main types of arrays.

A linear array consists of many (often 128-512) rectangular piezoelectric elements where the space between elements is called kerf and the distance between their centers is referred to as pitch. Each element is mechanically and electrically isolated from its neighbors by filling the kerf regions with acoustically isolating material. The elements are not focused. Pitch is designed to range from λ/2 to 3λ/2 (where λ is ultrasound wavelength in tissue). Linear arrays are usually about 1 cm wide and 10-15 cm long.

Linear arrays work by using separate voltage pulses to excite a small number of elements at slightly different times where the outer elements are excited first and the inner elements are excited after a short delay. This creates an effectively curved wavefront with a focal point at a certain distance from the array. After all backscattered echoes have been received, another beam consisting of a distinct subset of elements performs the same steps. This is repeated in sequence until all of the groups of elements have completed the procedure. If even numbers of elements were used for each group, the entire process can be repeated again with odd numbers of elements to cover the focal points between those acquired before.

Linear array focusing occurs only in one dimension. By contrast, the elevation plane (the direction perpendicular to the image plane) cannot be focused unless a curved lens is included to produce focus in this dimension. Linear arrays are most often used for applications involving large fields of view and relatively low penetration depth.

Phased arrays are typically around 1-3 cm in length and 1 cm in width. They are used in applications like cardiac imaging where there is only a small region of the body through which the ultrasound can enter without running into bone or air.

As with linear arrays, phased arrays apply voltage pulses at slightly varying times to excite elements and produce an effective wavefront with a certain focal length. But phased arrays must employ beam steering to reconstruct a full 2D image. This occurs by changing the pattern of excitation to sweep the effective wavefront beam across a range of directions to cover the image plane.

Phased arrays also employ a process called dynamic focusing to optimize lateral resolution over the full depth of imaged tissue. This involves dynamically changing the number of elements used to produce a wavefront with varying focal lengths. At deeper regions in the tissue, the number of elements needed to position the focus (where optimal lateral resolution is achieved) is higher than at shallower regions in the tissue. Dynamic focusing allows high lateral resolution across the full depth of the scan. However, dynamic focusing is relatively slow since multiple scans are needed to build up a single line of the image. It should be noted that the length of each element determines the “slice thickness” for the image’s elevation dimension.

There are also multidimensional transducer arrays which include extra rows of transducer elements. These multidimensional arrays can focus in the elevation dimension without the need for curved elements or lenses (though they are more complicated devices). Multidimensional arrays with a small number of extra rows (e.g. 3-10) are referred to as 1.5 dimensional arrays. These 1.5D arrays can facilitate some level of focusing in the elevation dimension, though to a limited extent. When multidimensional arrays possess a large number of extra rows (up to the number of elements in each row), they are referred to as 2 dimensional arrays. These can acquire full 3D image data without needing to be moved from their initial position.

Annular arrays represent another class of transducer array. They are useful at very high frequencies (>20 MHz) since linear or phased arrays are quite difficult to create for these frequencies. Annular arrays consist of concentric rings of piezoelectric material alternating with rings acoustically isolating material. Beam forming is accomplished using an analogous strategy to that of phased arrays. The outermost rings are excited first and the innermost rings last, producing an effective focus. Because annular arrays require mechanical motion to sweep the beam through tissue for reconstruction of images, commercially available devices have been developed to precisely control this motion.

When transducer arrays receive signals, they pass through an amplifier to strengthen them before digitization. However, such amplifiers do not provide linear gain for signals with a dynamic range that exceeds 40-50 dB. This is an issue because very strong signals appear from tissue boundaries near the transducer while much weaker signals appear from tissue boundaries deeper in the body. Weak signals can thus be lost when attempting to receive over larger dynamic ranges. A process called time-gain compensation (TGC) is employed to circumvent this issue. TGC increases the amplification factor as a function of time after transmission of an ultrasound pulse. As a result, the weaker backscattered echoes which come later are amplified to a greater degree than the stronger backscattered echoes which come sooner. TGC is controlled by the operator of the instrument, which usually comes with a variety of preset values for clinical imaging protocols.

Parameters for focused ultrasound in practice

There is no universally accepted definition of FUS dose, so various metrics of exposure are used to quantify how much ultrasound is delivered during a therapeutic session. Examples of such metrics (which will be discussed further below) include acoustic pressure or peak negative pressure, mechanical index, frequency, pulse repetition frequency (PRF), and intensity.

Peak negative pressure is often measured in MPa and describes the degree of rarefaction caused by the ultrasound wave in tissue. For low intensity focused ultrasound (LIFU), the greatest mechanical safety risk is from cavitation (bubble formation and collapse). To measure cavitation risk, the mechanical index (MI) is used. MI can be computed using the following equation where Pn is the peak negative pressure, f0 is the fundamental frequency, and the derating constant of 0.3 adjusts for tissue attenuation (~7% loss per cm per MHz) and has units dBcm–1MHz–1. After derating, 0.3Pn has units of MPa. FDA guidelines specify that MI should not exceed 1.9. MI itself is unitless.

An ultrasound wave’s pulse’s duration (PD) is the number of cycles divided by the frequency. For instance, a pulse with 500 cycles of 500 kHz ultrasound would last for 1 ms. The pulse repetition interval (PRI) is the amount of time between the start of one pulse and the start of the next pulse (so it includes both the pulse and the pause after the pulse). Pulse repetition rate (PRR) also known as the pulse repetition frequency (PRF) equals 1/PRI. The pulse duty cycle (PDC) equals PD/PRI and is expressed as a percentage. PD typically ranges from microseconds to seconds, PRI from milliseconds to seconds, and PDC from <1% up to 70%.

One’s choice of a particular PD and PDC comes from two main factors: (i) the duty cycle can have varying neuromodulatory effects (excitatory or inhibitory) depending on its value and (ii) lower PDC values can be leveraged to limit total energy and heat deposition.

A pulse train is a series of pulses, for which the pulse train duration (PTD) equals the total number of pulses times the PRI. Typical PTDs range from less than 1 second to several minutes. The amount of time between the start of one pulse train and the start of the next is called the pulse train repetition interval (PTRI). The amount of time between pulse trains is called the interstimulus interval (ISI). The pulse train duty cycle (PTDC) equals PTD/PTRI.

It should be noted that the PTDC does not have a major influence on neuromodulatory effect, so the ratio is driven by safety such that the ISI is long enough to limit cumulative heating to reasonable levels.

Multiplying PDC by PTDC gives an overall duty cycle equal to (PD/PRI)(PTD/PTRI) which can be further multiplied by the average intensity of the pulses Iavg to obtain average temporal intensity Iavg_tp. FDA diagnostic safety guidelines state that average intensity should fall below 720 mW/cm2. So, Iavg_tp = (PD/PRI)(PTD/PTRI)Iavg generally should not exceed 720 mW/cm2.

Total ultrasound application time is the sum of the durations of all pulse trains plus ISIs. It typically ranges from less than 1 minute to over 60 minutes. Longer total ultrasound application time is thought to usually improve efficacy by depositing more energy, though this may not always be true. Energy per unit time might play a more significant role in efficacy, but this is an ongoing area of investigation.

Frequency (or fundamental frequency f0) is the primary frequency of FUS passing through the tissue. It is typically measured in kilohertz (kHz) or megahertz (MHz). In human neuromodulation applications, frequency typically ranges from 200-700 kHz (or 0.2-0.7 MHz), providing an acceptable tradeoff between amount of energy entering the brain and the size of the focal region. The reason that the upper limit of frequency for human neuromodulation is typically ~700 kHz is because FUS energy attenuation by the skull at 700 kHz is ~75% (though this varies depending on skull morphology) and keeps increasing at higher frequencies.

Recall from the equations in the earlier discussion that lateral resolution involves wavelength λ (and f = c/λ), so frequency influences the focal region’s size. Also discussed earlier, frequency influences the distance from the transducer to the near-field boundary (ZNFB). Frequency itself is generally not believed to contribute to neuromodulatory effects in a direct fashion, though this is still under investigation. So, the f0 value is usually selected to create a focal volume of a desired size at a given depth.

Intensity is defined as power per unit area (and recall the unit of power is Watts or J/s) and is the rate at which energy is transferred by the FUS wave. For ultrasound at any given point in time during the wave cycle, intensity is proportional to the square of the acoustic pressure as described by the equation below where P is the acoustic pressure, ρ is the density of the medium, and c is the ultrasound speed in the medium. Recall that ρc = z, the acoustic impedance. Acoustic intensity is usually measured in watts per square centimeter (W/cm2).

Beyond instantaneous intensity, FUS is often measured by spatial peak pulse average intensity (ISPPA) and by spatial peak temporal average intensity (ISPTA). ISPPA is the average intensity experienced during a single ultrasound pulse. Note that I does not equal Pn2/2z in the case of a ramped pulse. To determine average intensity (ISPPA) for ramped pulses, the integral of intensity across the pulse is divided by the pulse’s duration PD. Ramped pulses distribute energy more smoothly and help mitigate auditory confounds for LIFU applications.

ISPTA represents the average intensity of the FUS beam at the point where it is strongest averaged over the pulse duration while accounting for any off periods. It is described by the following equation consisting of the ISPPA multiplied by the PDC (which is the fraction of PRI that the pulse is turned on).

References:

1.      Legon, W. & Strohman, A. Low-intensity focused ultrasound for human neuromodulation. Nat. Rev. Methods Prim. 4, 91 (2024).

2.      Smith, N. B. & Webb, A. Introduction to Medical Imaging: Physics, Engineering and Clinical Applications. (Cambridge University Press, 2010).

Reexamining the neurobiological correlates of subjective experience for whole-brain emulation (slides)

By No Comments
Reexamining the neurobiological correlates of subjective experience for whole-brain emulation (slides)

I presented these slides (PDF and images below) during the Workshop on Philosophy and Ethics of Brain Emulation (January 28th-29th, 2025) at the Mimir Center for Long Term Futures Research in Stockholm, Sweden. In my talk, I explored how various biological phenomena beyond standard neuronal electrophysiology may exert noticeable effects on the computations underlying subjective experiences. I emphasized the importance of the large range of timescales that such phenomena operate over (milliseconds to years). If we are to create emulations which think and feel like human beings, we must carefully consider the numerous tunable regulatory mechanisms the brain uses to enhance the complexity of its computational repertoire. At the workshop, we also discussed how the peripheral nervous system, enteric nervous system, endocrine system, musculoskeletal system, sensory systems, and perhaps even immune system may or may not play roles in subjective experience. To better understand what is needed to support proper emulations, I recommend recruitment of more fundamental neurobiology specialists to the whole-brain emulation community.

PDF: Reexamining the neurobiological correlates of subjective experience for whole-brain emulation

My abstract:

How much biological detail must a whole-brain emulation (WBE) incorporate to accurately preserve human subjective experience such that living as a WBE would truly feel like existing as a human? I ask this question independently of whether a nonbiological substrate can support subjective experience. Sandberg and Bostrom’s whitepaper Whole Brain Emulation: A Roadmap, briefly explores how levels of emulation detail ranging from abstract brain modules to quantum interactions between molecules may influence success criteria. They estimate that the necessary detail may at most involve an emulation of a connectome with multicompartmental models of neurons, dynamical synaptic states, and concentrations of metabolites and neurotransmitters in each compartment. For simplicity, I will refer to this as a multicompartmental emulation.

Although a multicompartmental emulation might produce an approximation of a human mind, I argue that the lack of additional biological layers of regulation could result in a subjective experience which has significant inaccuracies or missing pieces. Biological systems possess a massive number of tunable regulatory phenomena extending beyond multicompartmental electrophysiology. Some of these phenomena include but are not limited to morphological plasticity of brain cells, glial influence on computation, neurovascular coupling, adult neurogenesis, intercellular RNA transport, gap junctions, volume transmission, influence of perineuronal nets and other extracellular matrix (ECM) components, mechanical influences (e.g. crowding of synapses) on neuronal computation, ephaptic coupling, temporal evolution of genomic-transcriptomic state, and co-transmission of multiple neurotransmitters from the same synaptic bouton. In particular, experiences which depend on long-timescale changes across the brain may not be properly captured by a model which focuses on the fast electrophysiological dynamics of multicompartmental models with fixed connectomic and morphological properties.

To move towards accurately reproducing the feeling of humanness, I propose a first step of rigorously surveying neuroscience literature and evaluating how biological regulatory phenomena contribute to subjectively observed conscious experiences. This may facilitate construction of a draft catalogue where putative links between aspects of conscious experience and neurobiological phenomena can be established. Such an examination of the neural correlates of subjective experiences may serve as an initial guide for future efforts towards WBEs which preserve the feeling of humanness.