Combining data and theory for derivable scientific discovery with AI-Descartes

Cornelio, Cristina; Dash, Sanjeeb; Austel, Vernon; Josephson, Tyler R.; Gonçalves, J.P.M.; Clarkson, Kenneth L.; Megiddo, Nimrod; Khadir, Bachir El; Horesh, Lior

doi:10.1038/s41467-023-37236-y

Cited by 24 publications

(18 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…An early example of this is AI-Descartes, in which a symbolic regression algorithm generates equations to match experimental data, which is then combined with an automated theorem prover to establish the equations' “derivability” with respect to a scientific theory. 33 However, in this work, each theory required human expertise to be expressed in formal language, and reliance on an automated theorem prover limited the scope of theories to those expressible in first-order logic. AI tools that can autoformalize the informal scientific literature, generate novel theories, and auto-complete complex proofs could open new avenues for automating theory discovery.…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

Formalizing chemical physics using the Lean theorem prover

Bobbin,

Sharlin,

Feyzishendi

et al. 2024

Digital Discovery

Self Cite

View full text Add to dashboard Cite

show abstract

Section: Discussionmentioning

confidence: 99%

“…Artificial intelligence tools for scientific discovery have also used theorem provers in designing optical quantum experiments, 32 as well as for rediscovering and deriving scientific equations from data and background theory. 33…”

Section: Introductionmentioning

confidence: 99%

Formalizing chemical physics using the Lean theorem prover

Bobbin,

Sharlin,

Feyzishendi

et al. 2024

Digital Discovery

Self Cite

View full text Add to dashboard Cite

show abstract

“…16 The theory-infused neural network (TinNet) exhibits emergent behaviors in interpretability, whereas it is fundamentally limited in generalizability of the underlying neural network architectures, especially with out-of-distribution samples. Imposing constraints or scientific rules through other means, e.g., physics-inspired graph kernels, 15 symbolic regression with logical reasoning, 59 might provide improved accuracy while attaining interpretability.…”

Section: Conclusion and Perspectivementioning

confidence: 99%

Interpretable Machine Learning for Catalytic Materials Design toward Sustainability

Xin,

Mou,

Pillai

et al. 2023

Acc. Mater. Res.

View full text Add to dashboard Cite

Metrics & MoreArticle RecommendationsCONSPECTUS: Finding catalytic materials with optimal properties for sustainable chemical and energy transformations is one of the pressing challenges facing our society today. Traditionally, the discovery of catalysts or the philosopher's stone of alchemists relies on a trial-and-error approach with physicochemical intuition. Decades-long advances in science and engineering, particularly in quantum chemistry and computing infrastructures, popularize a paradigm of computational science for materials discovery. However, the brute-force search through a vast chemical space is hampered by its formidable cost. In recent years, machine learning (ML) has emerged as a promising approach to streamline the design of active sites by learning from data. As ML is increasingly employed to make predictions in practical settings, the demand for domain interpretability is surging. Therefore, it is of great importance to provide an in-depth review of our efforts in tackling this challenging issue in computational heterogeneous catalysis.In this Account, we present an interpretable ML framework for accelerating catalytic materials design, particularly in driving sustainable carbon, nitrogen, and oxygen cycles. By leveraging the linear adsorption-energy scaling and Brønsted−Evans−Polanyi (BEP) relationships, catalytic outcomes (i.e., activity, selectivity, and stability) of a multistep reaction can often be mapped onto one or two kinetics-informed descriptors. One type of descriptor of great importance is the adsorption energies of representative species at active site motifs that can be computed from quantum-chemical simulations. To complement such a descriptor-based design strategy, we delineate our endeavors in incorporating domain knowledge into a datadriven ML workflow. We demonstrate that the major drawbacks of black-box ML algorithms, e.g., poor explainability, can be largely circumvented by employing (1) physics-inspired feature engineering, (2) Bayesian statistical learning, and (3) theory-infused deep neural networks. The framework drastically facilitates the design of heterogeneous metal-based catalysts, some of which have been experimentally verified for an array of sustainable chemistries. We offer some remarks on the existing challenges, opportunities, and future directions of interpretable ML in predicting catalytic materials and, more importantly, on advancing catalysis theory beyond conventional wisdom. We envision that this Account will attract more researchers' attention to develop highly accurate, easily explainable, and trustworthy materials design strategies, facilitating the transition to the data science paradigm for sustainability through catalysis.

show abstract

“…Incorporating adequate prior knowledge is critical for developing generalizable deep learning models in data-deficient scientific problems 1 – 3 . The domain-specific prior knowledge of a particular task can help a model featurize generalizable patterns across its training data, leading to noteworthy successes 4 – 7 . For instance, AlphaFold2 8 utilized the co-evolutionary information to narrow down the extensive conformational space of protein folding on a macroscopic scale and the residue pair representation to reduce the structural complexity on a microscopic scale.…”

Section: Introductionmentioning

confidence: 99%

3D molecular generative framework for interaction-guided drug design

Zhung,

Kim,

Kim

2024

Nat Commun

View full text Add to dashboard Cite

Deep generative modeling has a strong potential to accelerate drug design. However, existing generative models often face challenges in generalization due to limited data, leading to less innovative designs with often unfavorable interactions for unseen target proteins. To address these issues, we propose an interaction-aware 3D molecular generative framework that enables interaction-guided drug design inside target binding pockets. By leveraging universal patterns of protein-ligand interactions as prior knowledge, our model can achieve high generalizability with limited experimental data. Its performance has been comprehensively assessed by analyzing generated ligands for unseen targets in terms of binding pose stability, affinity, geometric patterns, diversity, and novelty. Moreover, the effective design of potential mutant-selective inhibitors demonstrates the applicability of our approach to structure-based drug design.

show abstract

Combining data and theory for derivable scientific discovery with AI-Descartes

Cited by 24 publications

References 27 publications

Formalizing chemical physics using the Lean theorem prover

Formalizing chemical physics using the Lean theorem prover

Interpretable Machine Learning for Catalytic Materials Design toward Sustainability

3D molecular generative framework for interaction-guided drug design

Contact Info

Product

Resources

About