Examples of suitable ETH Grant proposals

Applicant has a background in:
Structural biology of membrane receptors using Cryo-EM methods
(A similar logic of assessing the step out nature of a project holds true for developers or expert users of methods such as imaging, NMR, mass spectrometry, omics, etc., as well as experts on a given system such a pathway, organism, ecosystem)

Project description, and why it is a step out:
With my established track record in studying the structural biology of cell membrane receptors using Cryo-Electron Microscopy (Cryo-EM), I am now proposing a new research direction into the field of immunotherapy, specifically, the development of next-generation Chimeric Antigen Receptor T-cell (CAR-T) therapies. While this proposal represents a novel area for me, my expertise in receptor structural biology is highly relevant and equips me with a unique perspective to tackle challenges in CAR-T therapy design.

This project proposes to leverage my expertise in receptor structural biology and Cryo-EM to create novel CAR cell designs that address these limitations:

1. Structural Analysis of CAR Cell Interactions: Using Cryo-EM, we will analyze the structures of CARs in complex with their target antigens at atomic resolution. This will provide critical insights into the molecular mechanisms of CAR-antigen recognition and could help design CARs with enhanced specificity and reduced off-tumor toxicity.

2. Development of Multispecific Cell CARs: Based on our structural understanding, we will design and generate multispecific CARs capable of recognizing multiple tumor antigens. This approach is expected to enhance the efficacy against heterogeneous solid tumors and prevent antigen escape, a common mechanism of resistance against CAR-T therapies.

3. In Vitro and In Vivo Evaluation of Novel CAR Cell Designs: Newly designed CARs will be expressed in T cells and their functionality will be tested in vitro and in a suitable animal model.

While the shift from structural biology to immunotherapy design represents a new direction for my research, I believe my expertise in understanding and manipulating receptor structures provides a strong foundation for this proposal. Furthermore, the study could lead to more effective and safer CAR-T therapies, significantly advancing cancer immunotherapy.
This project requires collaboration with immunologists and oncologists, which will facilitate cross-disciplinary learning and potentially open up new avenues for future research. Ultimately, by harnessing the structural knowledge of membrane receptors, we aim to take a step forward in the fight against cancer.

Applicant has a background in:
Survey design, microsociology, sociology of work

Project description, and why it is a step out:
We propose to trial new means to collect behavioural, emotional, and attitude data in experimental surveys. Willingness to participate in traditional survey types such as interviews and questionnaires, even if conducted online, is waning, and threats to data quality such as recall, attention, and fatigue persist. Experiential surveys, in which participants are prompted to respond periodically or based on trigger events, are an attempt to address some of these issues for certain information demands. To reduce intrusiveness and increase adaptivity, the use of context-aware conversational agents may allow for qualitatively new means to design and conduct surveys and experiment with such techniques in the context of remote collaboration.
While we have extensive experience in social surveys, especially in the context of office work, productive use of language technologies is entirely novel for my group. We strive to build up expertise in this area to be able to competently demonstrate and assess its potential, and to prepare ourselves and others for future studies using such technologies.

Whereas conversational agents are indeed being explored as interfaces that respondents might find more convenient to use, we intend to go beyond form factors and carry out foundational work on integrating the technology into the design of experiential surveys:

1. Prompt engineering for adaptive item design: Several steps are necessary to build competence in using generative AI to create questionnaire items in response to situational contexts, and to re-code responses so that they are usable in statistical analysis.

2. Item triggers and response quality: Through two pilot studies conducted in the Decision Science Lab at ETH, we will gather experience with participant reactions to agent interventions, and reliability and validity of reactions. Based on these results, we will specify protocols for smartphone interactions.

3. Case study in remote work: In the third year of the project, the utility of the above building blocks will be tested in a comparative study against current means of engaging participants. We are cooperating with an innovative Swiss company planning to reduce office space and to integrate several remote locations. Study design, evaluation, and ethical considerations are detailed in the proposal.

The project is heavy on technology that is developing fast, and we attempt to mitigate the risk of obtaining obsolete results by focusing on the fundamental aspects of situational context and adaptive item design. If successful, this project may lay the groundwork for a major shift in survey research.

Applicant has a background in:
Machine learning (ML), both theoretical and applied, applications to computer vision and large language models (LLMs).

Project description, and why it is a step out:
The training of very large ML models, such as LLMs in use today, is extremely compute and energy intensive. So much so, that the electricity requirements to train the largest models are comparable to the total electricity demand of a large city over a period of time. Given this context, there is a pressing need for the design of completely new machine learning architectures that do rely on traditional digital computers. One promising alternative is provided using analog electrical systems such as nonlinear electrical networks consisting of elements such as diodes, resistors, amplifiers and voltage sources. Such analog computing can be extremely energy efficient, yet the open question is whether it can be used to mimic machine learning algorithms such as deep neural networks. The aim of the proposed project is precisely to find this out.

To do so, we propose the following concrete steps:

1. We will prove a universal approximation theorem for these nonlinear electrical networks (NLENs) to show that they can approximate any continuous function. This will provide an analogous guarantee to the one that deep neural networks, based on digital computing, satisfy.

2. We will devise novel training algorithms for NLENs, which do not rely on backpropagation but rather on methods such as equilibrium propagation, which are based on the underlying physics of electrical circuits.

3. We will implement these circuits, stacked together, and simulate them with approximate ODE solvers to show how NLENs perform standard vision and text benchmarks, particu-larly in comparison to deep neural networks.

Although my group has expertise on traditional machine learning, the design and analysis of NLENs is completely different. The entire machinery has to be reinvented in this context and several standard operators such as matrix multiplication and gradient computation have to be replaced with their analog equivalents. In particular, the group has to acquire knowledge of the basic physics of electrical circuits in order to succeed in this project. Hence, this project constitutes a definitive step out. Nevertheless, the experience of the group in traditional machine learning is also necessary as several elements of ML algorithms need to be replicated in this project.
If this project is successful, it can form the basis for a collaborative grant with a group working on analog electrical circuits to realize these analog ML processors in the lab.