AI model disgorgement: Methods and choices.

Over the past few years, machine learning models have significantly increased in size and complexity, especially in the area of generative AI such as large language models. These models require massive amounts of data and compute capacity to train, to the extent that concerns over the training data (such as protected or private content) cannot be practically addressed by retraining the model "from scratch" with the questionable data removed or altered. Furthermore, despite significant efforts and controls dedicated to ensuring that training corpora are properly curated and composed, the sheer volume required makes it infeasible to manually inspect each datum comprising a training corpus. One potential approach to training corpus data defects is model disgorgement, by which we broadly mean the elimination or reduction of not only any improperly used data, but also the effects of improperly used data on any component of an ML model. Model disgorgement techniques can be used to address a wide range of issues, such as reducing bias or toxicity, increasing fidelity, and ensuring responsible use of intellectual property. In this paper, we survey the landscape of model disgorgement methods and introduce a taxonomy of disgorgement techniques that are applicable to modern ML systems. In particular, we investigate the various meanings of "removing the effects" of data on the trained model in a way that does not require retraining from scratch.

Confidence-ranked reconstruction of census microdata from published statistics.

A reconstruction attack on a private dataset D takes as input some publicly accessible information about the dataset and produces a list of candidate elements of D. We introduce a class of data reconstruction attacks based on randomized methods for nonconvex optimization. We empirically demonstrate that our attacks can not only reconstruct full rows of D from aggregate query statistics Q(D)∈ℝm but can do so in a way that reliably ranks reconstructed rows by their odds of appearing in the private data, providing a signature that could be used for prioritizing reconstructed rows for further actions such as identity theft or hate crime. We also design a sequence of baselines for evaluating reconstruction attacks. Our attacks significantly outperform those that are based only on access to a public distribution or population from which the private dataset D was sampled, demonstrating that they are exploiting information in the aggregate statistics Q(D) and not simply the overall structure of the distribution. In other words, the queries Q(D) are permitting reconstruction of elements of this dataset, not the distribution from which D was drawn. These findings are established both on 2010 US decennial Census data and queries and Census-derived American Community Survey datasets. Taken together, our methods and experiments illustrate the risks in releasing numerically precise aggregate statistics of a large dataset and provide further motivation for the careful application of provably private techniques such as differential privacy.

STING agonism overcomes STAT3-mediated immunosuppression and adaptive resistance to PARP inhibition in ovarian cancer.

BACKGROUND: Poly (ADP-ribose) polymerase (PARP) inhibition (PARPi) has demonstrated potent therapeutic efficacy in patients with BRCA-mutant ovarian cancer. However, acquired resistance to PARPi remains a major challenge in the clinic. METHODS: PARPi-resistant ovarian cancer mouse models were generated by long-term treatment of olaparib in syngeneic Brca1-deficient ovarian tumors. Signal transducer and activator of transcription 3 (STAT3)-mediated immunosuppression was investigated in vitro by co-culture experiments and in vivo by analysis of immune cells in the tumor microenvironment (TME) of human and mouse PARPi-resistant tumors. Whole genome transcriptome analysis was performed to assess the antitumor immunomodulatory effect of STING (stimulator of interferon genes) agonists on myeloid cells in the TME of PARPi-resistant ovarian tumors. A STING agonist was used to overcome STAT3-mediated immunosuppression and acquired PARPi resistance in syngeneic and patient-derived xenografts models of ovarian cancer. RESULTS: In this study, we uncover an adaptive resistance mechanism to PARP inhibition mediated by tumor-associated macrophages (TAMs) in the TME. Markedly increased populations of protumor macrophages are found in BRCA-deficient ovarian tumors that rendered resistance to PARPi in both murine models and patients. Mechanistically, PARP inhibition elevates the STAT3 signaling pathway in tumor cells, which in turn promotes protumor polarization of TAMs. STAT3 ablation in tumor cells mitigates polarization of protumor macrophages and increases tumor-infiltrating T cells on PARP inhibition. These findings are corroborated in patient-derived, PARPi-resistant BRCA1-mutant ovarian tumors. Importantly, STING agonists reshape the immunosuppressive TME by reprogramming myeloid cells and overcome the TME-dependent adaptive resistance to PARPi in ovarian cancer. This effect is further enhanced by addition of the programmed cell death protein-1 blockade. CONCLUSIONS: We elucidate an adaptive immunosuppression mechanism rendering resistance to PARPi in BRCA1-mutant ovarian tumors. This is mediated by enrichment of protumor TAMs propelled by PARPi-induced STAT3 activation in tumor cells. We also provide a new strategy to reshape the immunosuppressive TME with STING agonists and overcome PARPi resistance in ovarian cancer.

PARP Inhibition Elicits STING-Dependent Antitumor Immunity in Brca1-Deficient Ovarian Cancer.

PARP inhibitors have shown promising clinical activities for patients with BRCA mutations and are changing the landscape of ovarian cancer treatment. However, the therapeutic mechanisms of action for PARP inhibition in the interaction of tumors with the tumor microenvironment and the host immune system remain unclear. We find that PARP inhibition by olaparib triggers robust local and systemic antitumor immunity involving both adaptive and innate immune responses through a STING-dependent antitumor immune response in mice bearing Brca1-deficient ovarian tumors. This effect is further augmented when olaparib is combined with PD-1 blockade. Our findings thus provide a molecular mechanism underlying antitumor activity by PARP inhibition and lay a foundation to improve therapeutic outcome for cancer patients.

The Goddard and Saturn Genes Are Essential for Drosophila Male Fertility and May Have Arisen De Novo.

New genes arise through a variety of mechanisms, including the duplication of existing genes and the de novo birth of genes from noncoding DNA sequences. While there are numerous examples of duplicated genes with important functional roles, the functions of de novo genes remain largely unexplored. Many newly evolved genes are expressed in the male reproductive tract, suggesting that these evolutionary innovations may provide advantages to males experiencing sexual selection. Using testis-specific RNA interference, we screened 11 putative de novo genes in Drosophila melanogaster for effects on male fertility and identified two, goddard and saturn, that are essential for spermatogenesis and sperm function. Goddard knockdown (KD) males fail to produce mature sperm, while saturn KD males produce few sperm, and these function inefficiently once transferred to females. Consistent with a de novo origin, both genes are identifiable only in Drosophila and are predicted to encode proteins with no sequence similarity to any annotated protein. However, since high levels of divergence prevented the unambiguous identification of the noncoding sequences from which each gene arose, we consider goddard and saturn to be putative de novo genes. Within Drosophila, both genes have been lost in certain lineages, but show conserved, male-specific patterns of expression in the species in which they are found. Goddard is consistently found in single-copy and evolves under purifying selection. In contrast, saturn has diversified through gene duplication and positive selection. These data suggest that de novo genes can acquire essential roles in male reproduction.

Private algorithms for the protected in social network search.

Motivated by tensions between data privacy for individual citizens and societal priorities such as counterterrorism and the containment of infectious disease, we introduce a computational model that distinguishes between parties for whom privacy is explicitly protected, and those for whom it is not (the targeted subpopulation). The goal is the development of algorithms that can effectively identify and take action upon members of the targeted subpopulation in a way that minimally compromises the privacy of the protected, while simultaneously limiting the expense of distinguishing members of the two groups via costly mechanisms such as surveillance, background checks, or medical testing. Within this framework, we provide provably privacy-preserving algorithms for targeted search in social networks. These algorithms are natural variants of common graph search methods, and ensure privacy for the protected by the careful injection of noise in the prioritization of potential targets. We validate the utility of our algorithms with extensive computational experiments on two large-scale social network datasets.

Stress-induced changes in gene interactions in human cells.

Cells respond to variable environments by changing gene expression and gene interactions. To study how human cells response to stress, we analyzed the expression of >5000 genes in cultured B cells from nearly 100 normal individuals following endoplasmic reticulum stress and exposure to ionizing radiation. We identified thousands of genes that are induced or repressed. Then, we constructed coexpression networks and inferred interactions among genes. We used coexpression and machine learning analyses to study how genes interact with each other in response to stress. The results showed that for most genes, their interactions with each other are the same at baseline and in response to different stresses; however, a small set of genes acquired new interacting partners to engage in stress-specific responses. These genes with altered interacting partners are associated with diseases in which endoplasmic reticulum stress response or sensitivity to radiation has been implicated. Thus, our findings showed that to understand disease-specific pathways, it is important to identify not only genes that change expression levels but also those that alter interactions with other genes.

Behavioral dynamics and influence in networked coloring and consensus.

We report on human-subject experiments on the problems of coloring (a social differentiation task) and consensus (a social agreement task) in a networked setting. Both tasks can be viewed as coordination games, and despite their cognitive similarity, we find that within a parameterized family of social networks, network structure elicits opposing behavioral effects in the two problems, with increased long-distance connectivity making consensus easier for subjects and coloring harder. We investigate the influence that subjects have on their network neighbors and the collective outcome, and find that it varies considerably, beyond what can be explained by network position alone. We also find strong correlations between influence and other features of individual subject behavior. In contrast to much of the recent research in network science, which often emphasizes network topology out of the context of any specific problem and places primacy on network position, our findings highlight the potential importance of the details of tasks and individuals in social networks.

Coexpression network based on natural variation in human gene expression reveals gene interactions and functions.

Genes interact in networks to orchestrate cellular processes. Analysis of these networks provides insights into gene interactions and functions. Here, we took advantage of normal variation in human gene expression to infer gene networks, which we constructed using correlations in expression levels of more than 8.5 million gene pairs in immortalized B cells from three independent samples. The resulting networks allowed us to identify biological processes and gene functions. Among the biological pathways, we found processes such as translation and glycolysis that co-occur in the same subnetworks. We predicted the functions of poorly characterized genes, including CHCHD2 and TMEM111, and provided experimental evidence that TMEM111 is part of the endoplasmic reticulum-associated secretory pathway. We also found that IFIH1, a susceptibility gene of type 1 diabetes, interacts with YES1, which plays a role in glucose transport. Furthermore, genes that predispose to the same diseases are clustered nonrandomly in the coexpression network, suggesting that networks can provide candidate genes that influence disease susceptibility. Therefore, our analysis of gene coexpression networks offers information on the role of human genes in normal and disease processes.

Behavioral experiments on biased voting in networks.

Many distributed collective decision-making processes must balance diverse individual preferences with a desire for collective unity. We report here on an extensive session of behavioral experiments on biased voting in networks of individuals. In each of 81 experiments, 36 human subjects arranged in a virtual network were financially motivated to reach global consensus to one of two opposing choices. No payments were made unless the entire population reached a unanimous decision within 1 min, but different subjects were paid more for consensus to one choice or the other, and subjects could view only the current choices of their network neighbors, thus creating tensions between private incentives and preferences, global unity, and network structure. Along with analyses of how collective and individual performance vary with network structure and incentives generally, we find that there are well-studied network topologies in which the minority preference consistently wins globally; that the presence of "extremist" individuals, or the awareness of opposing incentives, reliably improve collective performance; and that certain behavioral characteristics of individual subjects, such as "stubbornness," are strongly correlated with earnings.

An experimental study of the coloring problem on human subject networks.

Theoretical work suggests that structural properties of naturally occurring networks are important in shaping behavior and dynamics. However, the relationships between structure and behavior are difficult to establish through empirical studies, because the networks in such studies are typically fixed. We studied networks of human subjects attempting to solve the graph or network coloring problem, which models settings in which it is desirable to distinguish one's behavior from that of one's network neighbors. Networks generated by preferential attachment made solving the coloring problem more difficult than did networks based on cyclical structures, and "small worlds" networks were easier still. We also showed that providing more information can have opposite effects on performance, depending on network structure.