Learning few-shot imitation as cultural transmission.

Cultural transmission is the domain-general social skill that allows agents to acquire and use information from each other in real-time with high fidelity and recall. It can be thought of as the process that perpetuates fit variants in cultural evolution. In humans, cultural evolution has led to the accumulation and refinement of skills, tools and knowledge across generations. We provide a method for generating cultural transmission in artificially intelligent agents, in the form of few-shot imitation. Our agents succeed at real-time imitation of a human in novel contexts without using any pre-collected human data. We identify a surprisingly simple set of ingredients sufficient for generating cultural transmission and develop an evaluation methodology for rigorously assessing it. This paves the way for cultural evolution to play an algorithmic role in the development of artificial general intelligence.

Fast, approximate kinetics of RNA folding.

In this article, we introduce the software suite Hermes, which provides fast, novel algorithms for RNA secondary structure kinetics. Using the fast Fourier transform to efficiently compute the Boltzmann probability that a secondary structure S of a given RNA sequence has base pair distance x (resp. y) from reference structure A (resp. B), Hermes computes the exact kinetics of folding from A to B in this coarse-grained model. In particular, Hermes computes the mean first passage time from the transition probability matrix by using matrix inversion, and also computes the equilibrium time from the rate matrix by using spectral decomposition. Due to the model granularity and the speed of Hermes, it is capable of determining secondary structure refolding kinetics for large RNA sequences, beyond the range of other methods. Comparative benchmarking of Hermes with other methods indicates that Hermes provides refolding kinetics of accuracy suitable for use in the computational design of RNA, an important area of synthetic biology. Source code and documentation for Hermes are available.

RNA folding pathways and kinetics using 2D energy landscapes.

RNA folding pathways play an important role in various biological processes, such as (i) the hok/sok (host-killing/suppression of killing) system in E. coli to check for sufficient plasmid copy number, (ii) the conformational switch in spliced leader (SL) RNA from Leptomonas collosoma, which controls trans splicing of a portion of the '5 exon, and (iii) riboswitches--portions of the 5' untranslated region of messenger RNA that regulate genes by allostery. Since RNA folding pathways are determined by the energy landscape, we describe a novel algorithm, FFTbor2D, which computes the 2D projection of the energy landscape for a given RNA sequence. Given two metastable secondary structures A, B for a given RNA sequence, FFTbor2D computes the Boltzmann probability p(x, y) = Z(x,y)/Z that a secondary structure has base pair distance x from A and distance y from B. Using polynomial interpolationwith the fast Fourier transform,we compute p(x, y) in O(n(5)) time and O(n(2)) space, which is an improvement over an earlier method, which runs in O(n(7)) time and O(n(4)) space. FFTbor2D has potential applications in synthetic biology, where one might wish to design bistable switches having target metastable structures A, B with favorable pathway kinetics. By inverting the transition probability matrix determined from FFTbor2D output, we show that L. collosoma spliced leader RNA has larger mean first passage time from A to B on the 2D energy landscape, than 97.145% of 20,000 sequences, each having metastable structures A, B. Source code and binaries are freely available for download at http://bioinformatics.bc.edu/clotelab/FFTbor2D. The program FFTbor2D is implemented in C++, with optional OpenMP parallelization primitives.

Computing the probability of RNA hairpin and multiloop formation.

We describe four novel algorithms, RNAhairpin, RNAmloopNum, RNAmloopOrder, and RNAmloopHP, which compute the Boltzmann partition function for global structural constraints-respectively for the number of hairpins, the number of multiloops, maximum order (or depth) of multiloops, and the simultaneous number of hairpins and multiloops. Given an RNA sequence of length n and a user-specified integer 0 ≤ K ≤ n, RNAhairpin (resp. RNAmloopNum and RNAmloopOrder) computes the partition functions Z(k) for each 0 ≤ k ≤ K in time O(K(2)n(3)) and space O(Kn(2)), while RNAmloopHP computes the partition functions Z(m, h) for 0 ≤ mm ≤ M multiloops and 0 ≤ h ≤ H hairpins, with run time O(M(2)H(2)n(3)) and space O(MHn(2)). In addition, programs such as RNAhairpin (resp. RNAmloopHP) sample from the low-energy ensemble of structures having h hairpins (resp. m multiloops and h hairpins), for given h, m. Moreover, by using the fast Fourier transform (FFT), RNAhairpin and RNAmloopNum have been improved to run in time O(n(4)) and space O(n(2)), although this improvement is not possible for RNAmloopOrder. We present two applications of the novel algorithms. First, we show that for many Rfam families of RNA, structures sampled from RNAmloopHP are more accurate than the minimum free-energy structure; for instance, sensitivity improves by almost 24% for transfer RNA, while for certain ribozyme families, there is an improvement of around 5%. Second, we show that the probabilities p(k)=Z(k)/Z of forming k hairpins (resp. multiloops) provide discriminating novel features for a support vector machine or relevance vector machine binary classifier for Rfam families of RNA. Our data suggests that multiloop order does not provide any significant discriminatory power over that of hairpin and multiloop number, and since these probabilities can be efficiently computed using the FFT, hairpin and multiloop formation probabilities could be added to other features in existent noncoding RNA gene finders. Our programs, written in C/C++, are publicly available online at: http://bioinformatics.bc.edu/clotelab/RNAparametric .

Using the fast fourier transform to accelerate the computational search for RNA conformational switches.

Using complex roots of unity and the Fast Fourier Transform, we design a new thermodynamics-based algorithm, FFTbor, that computes the Boltzmann probability that secondary structures differ by [Formula: see text] base pairs from an arbitrary initial structure of a given RNA sequence. The algorithm, which runs in quartic time O(n(4)) and quadratic space O(n(2)), is used to determine the correlation between kinetic folding speed and the ruggedness of the energy landscape, and to predict the location of riboswitch expression platform candidates. A web server is available at http://bioinformatics.bc.edu/clotelab/FFTbor/.