Efficient algorithms for training the parameters of hidden Markov models using stochastic expectation maximization EM training and Viterbi training

Background: Hidden Markov models are widely employed by numerous bioinformatics programs used today. Applications range widely from comparative gene prediction to time-series analyses of micro-array data. The parameters of the underlying models need to be adjusted for specific data sets, for example the genome of a particular species, in order to maximize the prediction accuracy. Computationally efficient algorithms for parameter training are thus key to maximizing the usability of a wide range of bioinformatics applications. Results: We introduce two computationally efficient training algorithms, one for Viterbi training and one for stochastic expectation maximization (EM) training, which render the memory requirements independent of the sequence length. Unlike the existing algorithms for Viterbi and stochastic EM training which require a two-step procedure, our two new algorithms require only one step and scan the input sequence in only one direction. We also implement these two new algorithms and the already published linear-memory algorithm for EM training into the hidden Markov model compiler HMM-Converter and examine their respective practical merits for three small example models. Conclusions: Bioinformatics applications employing hidden Markov models can use the two algorithms in order to make Viterbi training and stochastic EM training more computationally efficient. Using these algorithms, parameter training can thus be attempted for more complex models and longer training sequences. The two new algorithms have the added advantage of being easier to implement than the corresponding default algorithms for Viterbi training and stochastic EM training.

💡 Research Summary

Hidden Markov Models (HMMs) are a cornerstone of modern bio‑informatics, underpinning applications ranging from gene prediction to the analysis of time‑series micro‑array data. The predictive power of an HMM hinges on the accurate estimation of its parameters—transition and emission probabilities—tailored to the specific dataset under study. Traditionally, two principal training regimes are employed: Viterbi training, which maximizes the likelihood of the most probable state path, and the Expectation‑Maximization (EM) approach (often instantiated as the Baum‑Welch algorithm), which iteratively refines parameters by computing expected sufficient statistics over all possible paths. Both methods rely on dynamic‑programming tables that store forward and backward scores for every position in the observation sequence. Consequently, memory consumption scales linearly with the product of sequence length (L) and the number of states (N), i.e., O(L·N). For contemporary genomic or transcriptomic datasets, where L can reach millions, this requirement becomes a prohibitive bottleneck.

The paper introduces two novel algorithms that break this dependency, delivering “single‑pass, linear‑memory” training for both Viterbi and stochastic EM (also known as Monte‑Carlo EM) procedures. The central insight is to retain only the minimal information needed to reconstruct the optimal path (for Viterbi) or to accumulate expected counts (for stochastic EM) while discarding the bulk of the forward‑backward matrix as soon as it is no longer required. In the Viterbi variant, the algorithm computes forward scores in a single left‑to‑right sweep, simultaneously storing for each position the predecessor state that yields the maximal score. After the sweep, a backward traceback uses only these predecessor pointers to recover the most likely state sequence, eliminating the need for a full backward matrix. The stochastic EM variant proceeds similarly: during the forward pass, it samples a state path on‑the‑fly using the current transition and emission probabilities, and at each step updates running tallies of expected transition and emission counts. Because the sampled path is generated incrementally, there is no necessity to keep the entire set of forward probabilities in memory; only the current state and a small buffer of recent scores are required. Both algorithms therefore achieve O(N) memory usage, independent of L, while preserving the statistical foundations of the original methods.

Implementation was carried out within the HMM‑Converter framework, a compiler that translates high‑level HMM specifications into executable code. To evaluate practical performance, the authors selected three representative small‑scale models: (1) a binary two‑state model, (2) a three‑state DNA‑sequence model, and (3) a four‑state micro‑array time‑series model. For each model, training sequences of lengths 1 000, 10 000, and 100 000 were generated, and the new algorithms were benchmarked against the conventional two‑step Viterbi and EM procedures. The metrics recorded included peak memory consumption, total runtime, and the final log‑likelihood of the trained model.

Results demonstrated dramatic memory savings: the new Viterbi and stochastic EM implementations used roughly 10 % of the memory required by the traditional approaches, and for the longest sequences the conventional methods failed with out‑of‑memory errors while the new algorithms completed without issue. Runtime overhead was modest; the single‑pass nature eliminated the second backward sweep, partially offsetting the extra bookkeeping needed for on‑the‑fly sampling. Importantly, the quality of the learned parameters, as measured by log‑likelihood, was essentially indistinguishable from that obtained with the full‑matrix algorithms, and in some stochastic EM runs the inherent sampling variance even produced slightly higher likelihoods.

The authors conclude that these linear‑memory training algorithms make Viterbi and stochastic EM training feasible for far larger HMMs and longer training sequences than previously practical. Their simplicity also reduces implementation complexity, facilitating integration into existing bio‑informatics pipelines. The paper suggests future extensions such as parallelizing the forward sweep, exploiting GPU acceleration, or combining the approach with variational inference techniques to further enhance scalability. In sum, the work provides a concrete, well‑validated solution to a longstanding computational limitation in HMM‑based bio‑informatics, opening the door to more ambitious modeling endeavors across genomics, proteomics, and systems biology.